442 papers, organized into 18 research topics. Many papers span more than one topic and are listed under each of them. Within each topic, papers are listed newest first. Each title links to the paper's PDF.
Sketch-Based Design and Geometric Modeling
Interpreting freehand sketches and line drawings as 3D objects, sketch-based CAD interfaces, and geometric reasoning about engineering shapes such as sheet-metal parts and thin-walled primitives. This is the earliest thread of the work, reaching from 1990s CAD research to recent learned 3D reconstruction.
Studies projection-faithful reconstruction of a 3D wireframe from a single engineering line drawing, posed as conditional depth estimation over drawn pixels. A latent diffusion model conditioned on the sketch, with optional partial-depth input, is trained on over three million synthetic image-depth pairs from CAD wireframes. Scaling from 256 to 512 resolution roughly halves error, reaching 3.9% best-of-five normalized depth error.
Presents a fully automated system that builds 3D representations from 2D wireframe drawings using a neural network together with a genetic search algorithm. It builds on the hypothesis that angles and other geometric properties in the 2D scene support human-like 3D reconstruction, with the aim of enabling efficient and robust 3D sketch interfaces.
2009, ACM Transactions on Computer-Human Interaction
ModelCraft lets designers annotate and edit physical 3D models, such as 3D-printed or laser-cut prototypes, with a digital pen. The model's surface carries a traceable pattern, so sketches, comments and edit commands drawn on it are captured and transferred back into the digital CAD model.
Proposes that humans perceive 3D objects from a single sketch using simple 2D-3D geometrical correlations learned from experience. A correlation system trained on many object-sketch pairs learns to invert projection for unseen objects, and the same data can gauge how understandable a projection is. The authors demonstrate automatic conversion of a freehand sketch into a physical solid object.
Course notes for a SIGGRAPH session on creating 3D geometry from sketch input. They place sketch-based 3D modeling in the upper right of a complexity-ambiguity space, compounding stroke inaccuracy, the 2D-to-3D mapping ambiguity and scaling with object complexity.
Presents a freehand sketching application for CAD that reconstructs a 3D object from a single flat pen sketch. The algorithm uses stroke angular distribution and connectivity to find an orthogonal axis system and vertex depths, then reconstructs curved strokes separately assuming planarity. Objects with over 50 strokes reconstruct in interactive time using Levenberg-Marquardt optimization.
2006, ACM Symposium on User Interface Software and Technology (UIST)
Presents ModelCraft, a system that captures annotations and edits made on physical 3D models by tracking a pattern printed on the model surface, needing no tracking infrastructure or per-object calibration. Annotations and commands are merged back into the CAD model; the paper describes a SolidWorks plug-in and initial user feedback.
Presents a system that reconstructs a 3D object from a single freehand sketch in near real-time and uses it as the basis for physical simulation, with results interpreted back into the sketch. It addresses the limitation of mostly 2D sketch interfaces and is a first step toward fully interactive physics-based 3D design.
Presents a freehand sketch-based interface for CAD design and finite element analysis. Two optimization-based reconstruction algorithms handle 3D objects of straight lines and planar curves, with some objects of over 50 strokes reconstructed in interactive time. Structural properties can then be examined by finite element analysis and the object modified with the pen, supporting iterative design.
2004, National Conference on Artificial Intelligence
Presents a system that reconstructs a 3D object from a single freehand sketch in near real-time and uses it as the basis for a physical simulation, with results interpreted back into the sketch. It aims to bring physics-based analysis into early design and is a first step toward fully interactive physics-based 3D design.
2004, National Conference on Artificial Intelligence
Presents an algorithm to reconstruct a 3D object from a single freehand sketch of strokes connected at vertices. It uses the angular distribution of strokes to find one or more orthogonal 3D axis systems, then computes plausible vertex depths. It suits mostly orthogonal engineering sketches, and a Levenberg-Marquardt implementation reconstructs objects with over 100 strokes in interactive time.
Focuses attention on the role of the user interface in geometric modeling, where perceiving and manipulating 3D spatial information is difficult and good interfaces promise large benefits. The paper reviews recent related work and points to future challenges and opportunities.
Reports a system for conceptual design of sheet metal products from rough freehand sketches. Manufacturability, flat pattern, weight and cost can be estimated automatically without accurate details, so designers can explore concepts earlier. The paper describes the methodology and its algorithms, including flat-pattern determination using fuzzy AI techniques, with examples from a working implementation.
2000, Artificial intelligence for engineering design analysis and manufacturing
Argues that freehand sketching can serve as the language for conceptual design and analysis: if a CAD system can perceive a 3D model from a single sketch, established analysis tools can be brought into early design. The paper outlines the basic algorithms needed and gives several examples from an implemented system.
Proposes a thin-walled primitive for modeling inherently thin objects, with a topological invariant covering both manifold and thin-walled non-manifold objects, and corresponding Euler operators. The invariant is a necessary condition for topological validity of a thin-walled representation. It also suggests using non-integer values in the Euler-Poincare formula to merge with standard solid geometry.
Examines how the Information Age has changed the product engineering cycle and identifies three productivity-impeding forms of overload: detail, constraint and versatility. It suggests remedies: bringing computer tools in earlier, more natural in-context interfaces, and making computers active aids that suggest alternatives. Applications to design, manufacturing, simulation, quality assurance and maintenance are discussed.
Discusses converting variational geometry constraints into parametric geometry by creating auxiliary constraints and finding a sequential construction. New constraint chains create new paths in the constraint graph, which can break up strongly connected components and help conventional constraint solvers.
Analyzes the topology of schematic sheet metal parts (zero thickness, zero bend radii), which are usually non-manifold, and derives a general topological invariant relating facets, components, bends, free edges, welds, vertices, holes and volumes. Corresponding Euler operators are derived. The invariant lets one reason about manufacturing aspects from a single qualitative model, useful in early conceptual design.
1997, IFIP advances in information and communication technology
Proposes a freehand-sketching user interface for 3D conceptual design. Because a 2D sketch lacks depth and is imprecise, the algorithm first identifies and smooths lines, elliptic arcs and corners, links them into an edge-vertex graph, and then reconstructs the 3D object from geometric regularities, sketch topology and statistical distribution of entities. A working implementation is shown.
Presents a system for solving sets of 2D geometric constraints in product design that aims to combine the flexibility of variational geometry with the stability of parametric design. It decomposes the problem into sequential construction steps and, where none exists, generates auxiliary constructions by relocating constraints. The approach is implemented in a system for designing sheet metal parts.
Proposes classifying pen strokes in an on-line sketching system by least-squares fitting to a conic section equation, whose form naturally separates lines, elliptic arcs and corners; the hyperbola form is used for corner detection. The method is fast, runs in O(n) time in a single pass, and tolerates noisy input. An adaptive method for clustering disconnected endpoints is also given.
1996, IEEE Transactions on Pattern Analysis and Machine Intelligence
Addresses face identification in a 2D line drawing of a wireframe object, a key step in reconstructing the 3D object from a single projection. The paper lays out principles for finding edge circuits that correspond to real faces, including non-manifold objects, and presents an algorithm with proposed face-edge-vertex relationships, illustrated by a working implementation.
Describes an optimization-based algorithm that reconstructs a 3D wireframe model from a single inaccurate freehand sketch, taken as a 2D edge-vertex graph. Compared with line-labeling methods it tolerates imprecise vertex positions and missing entities, and it handles manifold and non-manifold objects with flat and cylindrical faces. Examples from an implementation are given.
Using evolutionary algorithms and generative representations to automatically design physical artifacts: mechanisms, structures, tensegrities, lattices and other engineered objects, including interactive and crowd-sourced evolution of 3D-printable shapes.
Presents JamTet, a physics-based sphere-packing framework for tetrahedral meshing and morphing in differentiable simulation. It combines a GPU-parallel mesher, Lagrangian morphing that keeps interior-node identities and avoids inversion, and a differentiable JAX simulator. In soft-robot swimming design experiments, interior-node gradients improve fitness over surface-only variants, and voxelized versions get 32-63% lower fitness.
Introduces ArticFlow, a two-stage flow-matching framework that learns an action-controlled velocity field from noise to target point sets, using a latent flow for a shape prior and a point flow conditioned on action. On MuJoCo Menagerie it works as both generative model and neural simulator, with higher kinematic accuracy and shape quality than object-specific and action-conditioned baselines.
Formulates planar linkage kinematic synthesis as cross-domain image generation, with an RGB dataset ranging from crank-rocker and crank-slider mechanisms to eight-bar linkages such as Jansen's. A shared-latent VAE synthesizes unseen motion curves, and speed encoded as color gradients allows conditioning on velocity profiles. Preliminary results are shown on three datasets of increasing complexity.
Presents a generative approach to optimizing 3D lattices of thousands of struts for mass and deflection. It iteratively erodes struts judged unnecessary by stress analysis and randomly regenerates new struts across the structure. Strength-to-weight ratios are on par with state-of-the-art topology optimization but qualitatively different; a GPU dynamics simulator also supports dynamic load cases.
Studies whether convolutional networks can predict static and dynamic properties of cantilever beams from raw cross-section images. Models predicted volume, maximum deflection and eigenfrequencies with 4.54% and 1.43% mean average percentage error against finite element analysis, and run over 1000 times faster, suggesting use as surrogate models in design optimization.
Demonstrates a 1D printing system, inspired by the ribosome, that folds a single string of source material into specialized robotic automata. A computational optimization tool derives designs from high-level goals. Three distinct robots for different locomotion tasks were made from the same material, and recycled material was reused to produce new designs.
2014, Genetic and Evolutionary Computation Conference (GECCO)
Proposes using the CPPN-NEAT evolutionary algorithm to tune the natural frequencies of structures through the material assignment of each voxel, keeping the object's shape and topology fixed. The goal is multi-material structures that resonate at multiple desired frequencies without external dampers, for robustness or energy harvesting.
Investigates fitness evaluation approaches for evolving nonlinear analog circuits when reverse-engineering an arbitrary black-box circuit. An incremental algorithm outperformed naive approaches, and the evolved robust circuits closely matched the black box's time-domain output for any time-domain input.
Outlines opportunities in soft robotics, including amorphous nonrigid dynamics, a new design space free of manufacturing constraints, natural mimicry and safer interaction with humans. Challenges include limited simulation and design tools, scarce soft actuation and difficult manufacturing. Proposed remedies include evolutionary design tools and porous and jamming materials.
Proposes using eye tracking instead of mouse clicks to drive interactive evolution of 3D-printable objects, aiming to reduce user fatigue and improve evolutionary success. It lays out a systematic method for testing this hypothesis, gives preliminary supporting results and outlines future larger studies.
2013, Genetic and Evolutionary Computation Conference (GECCO)
Proposes a design and manufacturing scheme inspired by ribosomal assembly, in which a custom printer folds a long ribbon with embedded control elements such as motors into a task-capable morphology. Folding patterns are encoded with CPPNs. Tests with different ribbon lengths show a single strip can fold into varied walking morphologies.
2013, Genetic and Evolutionary Computation Conference (GECCO)
Tests whether Karl Sims-style evolved morphologies are held back by rigid elements and overly regular encodings, by evolving multi-material soft robots with a CPPN generative encoding selected for locomotion speed. CPPNs evolved faster and more natural-looking robots than a direct encoding, and performance rose as more materials were added.
2013, IEEE Congress on Evolutionary Computation (CEC)
Describes how to let users evolve any uploaded 3D shape with CPPN-based interactive evolution by injecting the distance to the object's surface as a CPPN input. The seeded CPPN reproduces the original shape at arbitrary resolution while allowing it to be morphed in complex ways, and could extend to seeding neural wiring diagrams.
Introduces EndlessForms.com, a website for interactively evolving 3D shapes online using a CPPN generative encoding inspired by developmental biology, with WebGL display and an option to 3D print evolved designs. In its first year, nearly 3 million objects were evaluated over 190,000 generations by over 40,000 visitors from 150 countries.
Presents automated design and manufacture of static and locomoting objects whose function comes from the 3D distribution of materials, using evolutionary algorithms. Cantilever beams evolved to deflect in prescribed profiles were fabricated and matched within 0.5 to 7.6%. A freeform soft robot using volumetrically expanding actuator material was designed for forward locomotion, fabricated and confirmed within 15% error.
Presents a process that evolves target structures from functional requirements by minimizing strain energy under prescribed loads, and plans error-free assembly from many active components. Assembly planning samples paths by starting from the completed structure and removing accessible components one at a time, so at least one path to the target is guaranteed at every stage. It targets the simulation-to-reality gap in evolved morphologies.
Introduces a way to evolve 3D objects using compositional pattern producing networks (CPPNs), a generative encoding abstracting biological development. Interactive and target-based evolution produced complex 3D objects, and the inputs and functions available to CPPNs bias the kinds of objects produced. Objects transferred well to the real world when printed on a 3D printer.
Introduces a way to evolve 3D objects using compositional pattern producing networks (CPPNs), a generative encoding abstracting biological development. Interactive and target-based evolution produced complex 3D objects, and changing the information and functions given to CPPNs biases the results. Objects transferred well to reality when printed with a 3D printer.
2011, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
Uses evolutionary circuit design to automatically build equivalent analog circuits for neuronal ion channels, emulating the creativity behind the Hodgkin-Huxley model. From electrophysiological observations alone, the system evolved circuits that accurately simulate sodium and potassium channel behavior. It may serve as a reverse-engineering tool and as an aid in designing neuromorphic circuits.
A later listing of the Nature paper "Automatic design and manufacture of robotic lifeforms" (Lipson and Pollack, 2000): robot bodies and controllers evolved together in simulation and then fabricated automatically by 3D printing, producing physical machines that locomote.
Uses evolutionary circuit design to automatically construct equivalent analog circuits for neuronal ion channels, emulating the creative leap behind the Hodgkin-Huxley model. From electrophysiological observations alone, the system evolved circuits of basic electronic components that accurately simulate sodium and potassium channel behavior. It could serve as a reverse-engineering tool or assist in designing neuromorphic circuits.
Introduces an evolutionary algorithm with a generative representation for discovering large irregular tensegrity structures. It is shown to be effective and scalable relative to earlier methods, and it can find underlying structural patterns. The method produced the largest and most complex irregular tensegrities reported, with potential uses from space antennas to soft robotics.
Uses genetic programming to design analog circuits that stay robust under partial short-circuit and partial disconnect faults without any single point of failure, and also circuits that are highly sensitive to tampering. Robust circuits needed a similar number of components to standard ones. Surrogate models and coevolution sped up evolution, and generated circuits were physically built and tested.
Presents an evolutionary design algorithm for 3D multi-material structures that exploits the freedom of internal material distribution offered by multi-material printing. Material distribution is optimized under functional constraints such as maximizing stiffness per weight, and the method can also optimize shape. It is demonstrated on freeform shapes, non-uniform 3D structures and 3D compliant actuators.
2009, Genetic and Evolutionary Computation Conference (GECCO)
Presents an evolutionary design process for 3D multi-material structures, as an alternative to conventional topology optimization methods like homogenization that struggle in the large design space opened by multi-material printing. The algorithm designs substructures tailored to custom functions and is demonstrated on 3D non-uniform beams and 3D compliant actuators.
2008, Artificial intelligence for engineering design analysis and manufacturing
Applies genetic programming to synthesizing compound two-dimensional kinematic mechanisms and benchmarks it against a classical 19th-century kinematic challenge. The paper discusses representation choices for mechanism design and shows a number of human-competitive inventions.
Demonstrates a high-order frequency filter made of four pairs of coupled microring resonators with independent thermo-optic tuning of quality factor and resonant wavelength. An evolutionary algorithm finds the tuning needed for each ring to restore a transmission function distorted by fabrication errors.
2007, Genetic and Evolutionary Computation Conference (GECCO)
Explores developmental systems in which the developing structure and its environment affect each other on similar timescales. Using a model inspired by termite nest construction, evolution exploits these dynamics to generate adaptive, self-repairing structure more efficiently than a purely reactive system. The authors also offer a metric distinguishing reactive from interactive developmental systems.
A NASA Tech Brief describing automated design of complex modular robots by evolution with generative representations: encodings, such as programs, that can reuse parts of themselves and so scale to greater design complexity. Such representations are presented as a way around the computational obstacles that limit evolutionary design of practical engineering systems.
2007, Genetic and Evolutionary Computation Conference (GECCO)
Evolves map L-systems that grow planar graphs corresponding to the rods and strings of tensegrity structures, in place of a directly encoded structure that scales poorly. Each tensegrity is simulated in the Open Dynamics Engine and the convex hull volume of the rod endpoints serves as fitness. The map L-system significantly outperformed direct encoding in all runs.
Shows that fabrication errors in coupled high-Q silicon micro-ring resonators, which distort the transmission function, can be compensated by choosing the degree of all-optical tuning of each ring. An optimization algorithm selects the tuning, restoring a drastically distorted transmission function.
2006, Genetic and Evolutionary Computation Conference (GECCO)
Presents an interactive multi-objective coevolutionary algorithm that evolves candidate models of the user's preferences alongside candidate probes chosen to make those models disagree. On pen-stroke drawings, preferences were extracted with pairwise questions in fewer than ten probes, and the models extrapolated to preferred drawings they were never trained on.
Simulates an evolutionary process to design a high-confinement photonic structure starting from completely random patterns. Periodic patterns emerge spontaneously, along with previously unseen subwavelength bowtie regions. The evolved structure has a Q of 300 and an ultrasmall modal volume, indicating periodicity is a key condition for controlling light distribution.
Conference version of the study simulating evolution of a high-confinement photonic structure from random patterns. Periodic patterns and subwavelength bowtie regions emerge spontaneously, with an evolved Q of 300 and ultrasmall modal volume, suggesting periodicity is a principal condition for controlling light.
2005, Genetic and Evolutionary Computation Conference (GECCO)
Investigates evolutionary algorithms for form-finding of tensegrity structures, where finding stable connectivity of rigid and tensile elements is hard and existing methods cover only subsets. Evolution can explore the space of arbitrary tensegrities and find new, non-regular forms, suggesting a basis for a general design methodology.
Overview essay posing open-ended design synthesis as the meta-problem of engineering: designing machines that can design other machines. It motivates evolutionary approaches by noting that nature designed and fabricated its products autonomously, and asks whether computers can augment or replace human invention.
Demonstrates strong confinement of light in nanometer-scale waveguides and cavities, with intensity enhanced by orders of magnitude relative to the input. Unlike photonic crystals, which rely on interference in periodic media, the confinement here comes from strong field discontinuities, as in slot waveguides, and from scattering in pseudo-random structures.
Shows that genetic encoding influences which alternative optimal designs evolution finds when many optima exist. The test problem is designing a 2D shape that keeps constant height when rolled on flat ground; Reuleaux showed in the 19th century that the circle is not the only solution, and artificial evolution repeats that discovery in under an hour.
2005, Quantum Electronics and Laser Science Conference
Reports an evolutionary-algorithm-designed, compact pseudo-random arrangement of high-index dielectric rods that confines 1.55 micrometer light in a roughly 5 micrometer region through Anderson localization. The abstract frames this as direct evidence of confinement in a passive random medium, where earlier evidence was largely indirect.
Applies evolutionary algorithms to design two-dimensional photonic crystals, starting from random structures with no band gap. The best evolved crystal reached a gap-to-midgap ratio of 0.3189, which the authors report as 12.5% better than the best human design on the same index-contrast platform.
A short conference version of the photonic crystal work: an evolutionary algorithm searches for crystals with large band gaps starting from random structures without gaps. The evolved crystals were 12.5% better than the best published human design.
Describes an exploration-estimation algorithm for automated design, analysis and repair of nonlinear physical systems with minimal physical testing. The exploratory phase evolves improvements or tests in a simulator, and the estimation phase evolves the simulator itself to explain all observations so far. It is demonstrated for analysis, design and repair of electromechanical systems.
Applies genetic programming to synthesis of compound 2D kinematic mechanisms, benchmarked against the classical 19th-century challenge of drawing a straight line. The paper discusses considerations for choosing a representation for mechanism design and presents a number of human-competitive inventions.
2004, Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE
Uses evolutionary algorithms to search 2D photonic crystal unit cells, represented as bitmaps directly encoded or generated by construction trees, for maximum bandgaps at a given index contrast. Fitness rewards partial gaps and is computed with a plane-wave eigensolver. From random starts the process found novel crystals, some with bandgaps 12.5% larger than any previously reported human-designed ones.
Argues for fully automated design and construction of inexpensive special-purpose robots instead of a general-purpose humanoid, with software and hardware co-designed and built from reusable parts and rapid-prototyped materials. The central hypothesis is that robotics suffers from underestimating the complexity of control design, whereas in nature brains and bodies are selected together.
Shows that robotic systems can automatically design robots with complex morphologies and adapted controllers at low cost, using coevolution of bodies and brains in realistic physical simulation and transfer via rapid manufacturing. Reviews three generations: automatically designed LEGO structures, GOLEM truss-based electromechanical systems, and modular designs with a generative DNA-like representation.
2003, IEEE Transactions on Robotics and Automation
Presents generative representations for automated design of modular physical robots. The resulting machines remain simple compared with human-engineered ones, but are more regular, modular and hierarchical than previously evolved machines of similar function, and virtual designs have an order of magnitude more moving parts. The authors quantitatively show the generative representation searches more efficiently than a non-generative one.
2002, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Argues that automatic design systems must produce designs from reused modules to scale in complexity, and proposes a generative design grammar subject to evolution. Combining Lindenmayer systems with evolutionary algorithms gives hierarchical modularity. The system is demonstrated for the first time on physical, modular, 2D locomoting robots and their controllers.
A book chapter considering what it would take to apply non-biological evolutionary processes at the nanoscale, where large numbers of building blocks self-organize into irregular forms that meet functional requirements and adapt to changing conditions, a form of evolutionary materials. Drawing on lessons from evolving macro-scale robotic structures, it discusses scaling, modularity, an indirect method for inducing modularity, and steps toward physical implementation.
Reviews three generations of work toward automatically designing and fabricating special-purpose robots rather than a general-purpose one: buildable static structures, automatically designed and manufactured dynamic electromechanical systems, and modular robots designed through a generative encoding. The authors expect improved simulation, manufacturing and transfer to enable custom machinery for short-term tasks.
Considers how to promote modularity so evolutionary design can scale to higher complexity through reuse of building blocks. Proposes a weak, indirect mechanism based on the observation that modular designs adapt better and survive more under changing requirements, requiring no representation of partial solutions. The principle is demonstrated on an abstract design problem where modularity can be quantified.
Reviews three generations of research in which robot morphology and controller evolve together: automatically designed static structures, automatically designed and manufactured dynamic electromechanical systems, and modular robots designed through a generative DNA-like encoding. The aim is special-purpose mechanisms and controllers for specific short-term objectives, as an answer to the stasis in robot development.
Lipson and Pollack report the GOLEM project: robots whose bodies and controllers were designed together by an evolutionary process in simulation and then built automatically by 3D printing, with only motors added by hand. The resulting machines locomoted in the physical world, a step toward machines that control their own design and manufacture.
Explores evolving machines for specified tasks and replicating them into reality, closing a complete physical evolutionary design cycle. The design space uses only linear actuation elements and sigmoidal control neurons, simulated with quasi-static motion, and selected machines are built with rapid prototyping. Examples of virtual and real machines evolved for locomotion are presented.
An overview of how evolutionary computation can be applied to physical robotics, covering the evolution of controllers and robot bodies and the transfer of evolved designs from simulation to real hardware. (The PDF's text layer is unreadable, so this summary is based on the title and the figures.)
2000, First International Conference on Axiomatic Design (ICAD2000)
Proposes an architecture for a design knowledge database to support automation of engineering design. Following Axiomatic Design, each step of top-down decomposition is a database element holding modifiable parameters, physical behavior and interfaces. Information is stored in mathematical rather than textual form to reduce ambiguity and make it amenable to fully automated design.
Evolutionary Robotics and Co-Design of Body and Brain
Evolving robot morphologies and controllers together, in simulation and in hardware, from the GOLEM project's automatically designed and fabricated robots to evolved gaits, virtual creatures and evolved soft robots.
Presents Cronos, an open-source 3D simulation engine that parallelizes a nonlinear mass-spring model on the GPU for soft, rigid and hybrid robots. It reaches over 1 billion elements per second on consumer GPUs and allows mixing materials within one robot. Accuracy is validated against Euler-Bernoulli beam theory, natural frequency predictions and empirical data from a soft structure.
2017, Zenodo (CERN European Organization for Nuclear Research)
Examines why co-optimizing morphology and control is hard, noting that morphological changes tend to harm sensorimotor control. The authors introduce morphological innovation protection, which temporarily lowers selection pressure on recently changed morphologies so controllers can readapt. This avoids local optima and sustains fitness gains longer across varied initial conditions.
Tests whether morphological computation is helped by suitable geometry, material and growth by evolving geometries and growth processes of soft robots with softer or stiffer materials. For the task studied, evolved softer robots reached better performance with simpler growth processes than stiffer ones, which the authors attribute to better exploitation of morphological computation.
Asks why evolved virtual creatures have barely grown in complexity since Sims' early work. The authors propose, grounded in embodied cognition, that co-evolving morphology and control is inherently problematic. They reproduce a prior finding that morphology converges prematurely relative to controllers and discuss it as a symptom of this problem.
Simulated robots with a single communication channel evolve from non-communicating populations when reproduction requires finding a mate. Indexical communication arose in 99% of experiments, and 9% of populations evolved a symbolic scheme conveying two spatial dimensions over a one-dimensional channel. The results suggest symbolic communication can emerge under natural selection without cognitive preadaptations.
An editorial recalling the 2000 evolved, 3D-printed robot that walked out of the printer, designed through evolution in simulation and printed whole and preassembled. It poses the grand challenge for additive manufacturing of printing a complete functioning robot.
2015, Genetic and Evolutionary Computation Conference (GECCO)
Presents the first soft robot evolved to reach or squeeze through a small aperture, using CPPN-NEAT with a new multi-objective implementation that keeps speciation. More compliant, deformable robots performed better at the task, offered as a proof of concept for matching tasks to suitable morphologies.
Instead of tuning update rules, this work evolves the topology of a 3D cellular automaton while keeping one simple rule. Cells have an excited state and can contract or expand volumetrically, and an evolutionary algorithm optimizes the resulting soft 'robot' for locomotion in a simulated physical world. A variety of topologies produce effective locomotion.
Introduces electrophysiological soft robots, inspired by cardiac electrophysiology, whose behavior arises from electrical signals flowing through the tissue cells of the soft body, merging control and actuation. The paper describes how they are evolved, shows lifelike behaviors, and explores design choices such as sensing, actuation and central pattern generator placement.
Reviews selected projects that evolve both neural and morphological subsystems of robots to study how bodies and brains work together to produce adaptive behavior, within the embodied cognition paradigm. The focus is on synergies and tradeoffs between neural architecture, morphology, action and adaptive behavior, with robustness and resilience as themes.
Tests whether HyperNEAT gaits evolved first in simulation and then transferred to a physical quadruped outperform a simpler hand-tuned encoding. The hypothesis was confirmed, producing the fastest gaits yet seen for this robot, beating nine algorithms from three earlier papers, and showing that generative encodings work for real-robot gait problems.
2013, Genetic and Evolutionary Computation Conference (GECCO)
Proposes a design and manufacturing scheme inspired by ribosomal assembly, in which a custom printer folds a long ribbon with embedded control elements such as motors into a task-capable morphology. Folding patterns are encoded with CPPNs. Tests with different ribbon lengths show a single strip can fold into varied walking morphologies.
2013, Genetic and Evolutionary Computation Conference (GECCO)
Tests whether Karl Sims-style evolved morphologies are held back by rigid elements and overly regular encodings, by evolving multi-material soft robots with a CPPN generative encoding selected for locomotion speed. CPPNs evolved faster and more natural-looking robots than a direct encoding, and performance rose as more materials were added.
Introduces Aracna, an open-source quadruped built from off-the-shelf and 3D-printed parts, as a platform for gait-learning algorithms. All motors sit in the body, driving hip and knee joints through four-bar linkages, which fixes a previous platform's heavy legs and yields non-intuitive kinematics that suit evolutionary and learning methods.
Uses the ENSO approach (evolution of network symmetry and modularity) to evolve modular neural controllers in simulation for a physical quadruped, with real-world uncertainty modeled as noise. Transferred controllers produced synchronized trot gaits on different surfaces and, unlike a hand-designed PID controller, generalized to voltage loss and leg loss.
Applies the RL PoWER policy-learning algorithm with a spline parameterization to learn gaits on the open-source 3D-printed QuadraTot robot, entirely through physical experiments. It found the fastest gait reported from physical-only learning, 16.3% faster than HyperNEAT, with gaits less taxing on the robot and more repeatable.
Presents automated design and manufacture of static and locomoting objects whose function comes from the 3D distribution of materials, using evolutionary algorithms. Cantilever beams evolved to deflect in prescribed profiles were fabricated and matched within 0.5 to 7.6%. A freeform soft robot using volumetrically expanding actuator material was designed for forward locomotion, fabricated and confirmed within 15% error.
2011, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
Compares algorithms for automatically generating quadruped gaits, tested entirely on a physical robot because accurate simulators are often unavailable. Parameterized motion models optimized by six local search methods are compared with neural networks evolved by the HyperNEAT generative encoding. All outperformed a manually designed gait, and HyperNEAT was best, producing gaits nearly nine times faster than the reference.
Proposes a robot with two controllers: a reactive innate controller and a reflective controller that observes, models and modulates the innate one without access to its internals. The reflective controller synthesizes filters to adapt behavior to new situations. Tested with recurrent neural network innate controllers, it showed significant adaptation with few physical trials.
A later listing of the Nature paper "Automatic design and manufacture of robotic lifeforms" (Lipson and Pollack, 2000): robot bodies and controllers evolved together in simulation and then fabricated automatically by 3D printing, producing physical machines that locomote.
Demonstrates evolution of robots with continuous, amorphous, multi-material bodies actuated by periodic volumetric expansion and contraction of materials. It discusses representing evolvable freeform shapes, comparing genotypes that use a level-set threshold, and simulating soft bodies with a relaxation algorithm under actuation, gravity and nonlinear friction. This opens a design space closer to biological systems.
2010, Genetic and Evolutionary Computation Conference (GECCO)
Demonstrates evolution of locomoting amorphous multi-material robots actuated by periodic volumetric expansion and contraction. It discusses how to represent evolvable freeform shapes, exploring genotypes that use a level-set threshold, and how to simulate soft bodies with a relaxation algorithm including gravity and nonlinear friction. The work opens a design space resembling biological forms.
Presents a control scheme in which independent struts of a tensegrity robot coordinate to achieve locomotion by using the dynamical coupling from pre-stress stability as an ad-hoc communication channel. Time-sensitive spiking neural networks exploit this "morphological communication". The result supports morphological computation in robot control and suggests relevance to embodied computation in biological systems such as tendon networks.
2009, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
Applies metacognition to robots: an innate controller connects sensors to motors, while a meta controller monitors and modulates it without access to its internal state or architecture. Quantitative comparison with traditional evolutionary robotics methods suggests this self-reflection is a promising alternative for improving robot resiliency and learning.
2009, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
Explores a robotic theory of mind, using coevolution to evolve controllers that retrospectively explain an observed actor's behavior and to evolve new actions that elicit more revealing behavior. The evolved models can predict, manipulate and exploit the actor. In a physically realistic simulation the method significantly outperformed a direct estimation baseline.
2009, Workshop on Performance Metrics for Intelligent Systems (PerMIS)
Studies how robots can represent another agent's self model with neural networks and evolutionary learning, as a step toward theory of mind. The framework was tested on simulated and physical robots with a new prey-predator scenario for measuring performance. Results show the approach can recover the other agent's self models.
Shows that swarm behavior of similar complexity can be obtained from swarmers with only one controllable degree of freedom, the ability to move forward at different velocities. Interaction among units gives the swarm maneuverability the individuals lack, which may allow simpler, smaller units and larger swarms.
Argues that soft robots will likely be underactuated and undercontrolled, so part of actuation and control should be embodied in the robot's structural dynamics, termed morphological computation. Two examples are presented, the Manduca sexta caterpillar and a modular tensegrity tower, and the authors explore how they apply to locomotion and control of a soft robot.
An essay for the 50 Years of AI festschrift that opens with a robot demonstration to first graders who ask what the robots can actually do. It uses the anecdote to reflect on how hard it is to impress people with machine intelligence and on robots that learn, adapt and create.
A Current Biology commentary on robot-based simulations showing that communication can emerge spontaneously among evolving agents. It notes that deceptive signalling arises in purely competitive settings, while cooperative communication emerges only under group or kin selection, and places these results in the broader history of communication in evolution.
Demonstrates a large space of robots capable of autonomous self-reproduction, built from actuated modules with electromagnets that control the assembly's morphology. The authors show 2-D and 3-D machines from 3 to 2^n modules, including automatically generated and manually designed morphologies, and two physical implementations each achieving two generations of reproduction.
A NASA Tech Brief describing automated design of complex modular robots by evolution with generative representations: encodings, such as programs, that can reuse parts of themselves and so scale to greater design complexity. Such representations are presented as a way around the computational obstacles that limit evolutionary design of practical engineering systems.
Describes a method for a physical robot to synthesize a body schema from multimodal sensor data gathered through interaction. The synthesis can be parametric (refining an approximate model) or topological (building a predictive model of its own body plan with little prior knowledge). The topological case succeeded with a quadruped robot after only nine 5-second interactions.
Introduces locomotion robots built on tensegrity structures: TR3, a three-strut triangular prism, and TR4, a four-strut quadrilateral prism. Evolutionary algorithms produced controllers yielding static and dynamic gaits in simulation, and a physical robot was built from one simulation model as a proof of concept, suggesting lightweight, strong, fault-tolerant machines.
Talk abstract outlining challenges in moving evolutionary robotics from simulation to physical machines. Topics include evolving morphology and control, hybrid coevolution of reality and simulation, handling noise and uncertainty, and morphological adaptation in hardware, illustrated with a pneumatic nine-legged machine controlled by an open-loop pattern generator.
2006, International Conference on Advanced Robotics (ICAR)
Introduces a legged robot based on a triangular tensegrity prism actuated by contracting its transverse cables. A genetic algorithm designed controllers in simulation that produced multiple forward gaits, and a physical robot built from the simulated design achieved forward locomotion, suggesting very lightweight and robust mobile robots.
Evolves artificial neural network controllers on two cascaded field programmable analog arrays, evaluating every candidate in hardware with no simulation. Starting from random settings, the method produced controllers for biped locomotion and for restoring mobility to a damaged quadruped, aimed at low-power in-situ fault recovery for space applications.
Overview essay posing open-ended design synthesis as the meta-problem of engineering: designing machines that can design other machines. It motivates evolutionary approaches by noting that nature designed and fabricated its products autonomously, and asks whether computers can augment or replace human invention.
Presents a process for evolving flapping flight control patterns for an ornithopter, together with a realistic kinematic mechanism that can generate the motion in practice. The target is a hovering design that lifts only through drag on a flat wing. The goal is to carry simulated models over to a physical, untethered flier.
2005, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Tests whether robots with highly coupled mechanical structures, where one actuator moves many locations and vice versa, allow redundancy in control. Using three- and four-strut tensegrity prisms, locomotion controllers were evolved in simulation. Multiple, substantially different control strategies produced qualitatively similar forward locomotion, indicating considerable control redundancy.
2004, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Introduces a two-stage evolutionary algorithm for remote robots: first evolve a damage hypothesis after failure, then re-evolve a compensatory neural controller. Tested on a simulated 'physical' robot, preliminary results show nearly full functionality restored after three physical evaluations, versus over 3000 if evolved entirely on the robot. It also yields a diagnostic model of the failure.
Extends the estimation-exploration algorithm for recovering robot function after unanticipated damage or environmental change. A new fitness metric lets it describe partial as well as complete failures and tell internal damage from environmental change using only sensory feedback. In most cases only four hardware evaluations are needed to restore full functionality of the 'physical' robot.
Introduces a method for evolving dynamic gaits on a physical parallel-actuated pneumatic robot, assuming only that locomotion is rhythmic. A genetic algorithm evolves open-loop controllers whose encoding shapes speed and pattern while ensuring rhythmicity. Future plans include closed-loop controllers and reducing hardware trials.
Argues that solid freeform fabrication, traditionally for passive parts, can be developed into a compact robotic system that prints complete active electromechanical devices. Successes include printed zinc-air batteries, wiring and flexure joints; actuation, sensing and control remain, and conducting polymers are proposed. Coupled with evolutionary design, this would give evolution a rich physical substrate.
Presents the estimation-exploration algorithm as a co-evolutionary approach to crossing the reality gap in evolutionary robotics. It adapts the robot simulator using the target robot's behavior, and adapts the robot's behavior using the simulator. Claimed benefits are automation, a minimum of hardware trials, compatibility with other transfer approaches, and generalizability to other domains.
Examines whether periodic motion emerges under evolutionary pressure for energy-efficient, long-range, partially passive brachiation in simulated gibbon-like creatures with a two-arm, torso morphology. Genomes encode initial position and shoulder torque functions, and Pareto-optimal individuals were selected. Energy-efficient long-range motion resulted in periodic motion.
Argues for fully automated design and construction of inexpensive special-purpose robots instead of a general-purpose humanoid, with software and hardware co-designed and built from reusable parts and rapid-prototyped materials. The central hypothesis is that robotics suffers from underestimating the complexity of control design, whereas in nature brains and bodies are selected together.
Shows that robotic systems can automatically design robots with complex morphologies and adapted controllers at low cost, using coevolution of bodies and brains in realistic physical simulation and transfer via rapid manufacturing. Reviews three generations: automatically designed LEGO structures, GOLEM truss-based electromechanical systems, and modular designs with a generative DNA-like representation.
2003, IEEE Transactions on Robotics and Automation
Presents generative representations for automated design of modular physical robots. The resulting machines remain simple compared with human-engineered ones, but are more regular, modular and hierarchical than previously evolved machines of similar function, and virtual designs have an order of magnitude more moving parts. The authors quantitatively show the generative representation searches more efficiently than a non-generative one.
2002, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Argues that automatic design systems must produce designs from reused modules to scale in complexity, and proposes a generative design grammar subject to evolution. Combining Lindenmayer systems with evolutionary algorithms gives hierarchical modularity. The system is demonstrated for the first time on physical, modular, 2D locomoting robots and their controllers.
Presents the GOLEM project, which extends evolution into the physical world by evolving electromechanical machines made of thermoplastic, linear actuators and neurons for locomotion, in simulation and then reality. Using 3D solid printing, the evolved creatures replicate into reality and reproduce the performance of their virtual ancestors.
2002, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Proposes continuously reconfigurable robotics: given a task, a robot evolves a morphology and controller, prints the assembled structure, performs the task, and is then recycled into a new form. Demonstrated for locomotion with linear actuators and sigmoidal neurons in a thermoplastic body, printed pre-assembled by rapid prototyping, with examples of working physical robots.
Reviews three generations of work toward automatically designing and fabricating special-purpose robots rather than a general-purpose one: buildable static structures, automatically designed and manufactured dynamic electromechanical systems, and modular robots designed through a generative encoding. The authors expect improved simulation, manufacturing and transfer to enable custom machinery for short-term tasks.
Reviews three generations of research in which robot morphology and controller evolve together: automatically designed static structures, automatically designed and manufactured dynamic electromechanical systems, and modular robots designed through a generative DNA-like encoding. The aim is special-purpose mechanisms and controllers for specific short-term objectives, as an answer to the stasis in robot development.
Lipson and Pollack report the GOLEM project: robots whose bodies and controllers were designed together by an evolutionary process in simulation and then built automatically by 3D printing, with only motors added by hand. The resulting machines locomoted in the physical world, a step toward machines that control their own design and manufacture.
Explores evolving machines for specified tasks and replicating them into reality, closing a complete physical evolutionary design cycle. The design space uses only linear actuation elements and sigmoidal control neurons, simulated with quasi-static motion, and selected machines are built with rapid prototyping. Examples of virtual and real machines evolved for locomotion are presented.
An overview of how evolutionary computation can be applied to physical robotics, covering the evolution of controllers and robot bodies and the transfer of evolved designs from simulation to real hardware. (The PDF's text layer is unreadable, so this summary is based on the title and the figures.)
Describes the GOLEM project, which evolves electromechanical robots built from thermoplastic, linear actuators and neurons for locomotion, first in simulation and then in reality. Using 3D solid printing, the creatures replicate into hardware and reproduce the performance of their virtual ancestors, transferring evolved morphological diversity into reality for the first time.
Self-Modeling, Self-Awareness and Resilient Machines
Robots that build and use models of their own bodies, recover from damage, predict their own dynamics, and model other agents. This work spans from early resilient machines through continuous self-modeling to visual self-models and machine theory of mind.
Frames shape prediction for tendon-driven continuum robots as action-conditioned kinematic self-modeling. Using a lightweight 3D-printed platform and RGB-D data collection, a point cloud flow matching model maps motor actuation to settled 3D geometry. It improves shape accuracy over prior approaches on simulated 2-, 3- and 5-module robots and real 2- and 3-module robots, and extends to tip payload conditioning in simulation.
Proposes that a 'self' can be isolated as the invariant, slowly changing portion of a cognitive process, and tests this on robots under constant-task versus continual learning. Continual learning produces an invariant subnetwork that is significantly more stable than in controls and is functionally important: preserving it aids adaptation and damaging it impairs performance. Results hold across three robots in locomotion and manipulation.
2025, IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings
Presents AutoURDF, an unsupervised method that builds robot description files for unseen robots from point cloud frames. A cluster-based registration model tracks 6-DoF transformations of point clusters, which are used for moving-part segmentation, body topology inference and joint parameter estimation. Outputs are compatible with simulators, and it beats prior methods in registration and topology accuracy on synthetic and real scans.
Demonstrates a task-agnostic dynamic self-model learned self-supervised from a single first-person camera, with no prior knowledge of morphology, kinematics or task. On a 12-DoF robot it supports basic locomotion from visual input, lets the robot detect anomalies such as damaged components and adapt, and generalizes across robots with different configurations.
2024, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Uses neural fields to let a robot learn a task-agnostic self-model of its kinematics as a neural-implicit query model from 2D images annotated with camera poses and configurations, without depth or geometry knowledge. It introduces an encoder-based dynamic density field architecture with curricular sampling. On a 7-DoF robot the Chamfer-L2 error is 2% of workspace size, demonstrated on motion planning.
Trains a humanoid robot face with 26 degrees of freedom to anticipate a human's upcoming facial expression and execute it simultaneously, rather than mimicking with delay. The robot learns to predict a smile about 839 milliseconds before the human smiles and uses a learned inverse kinematic facial self-model to coexpress it. The authors suggest this could improve human-robot interaction.
2024, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Proposes a meta-self-modeling approach that infers a robot's morphology from proprioceptive signals. The study introduces a 12-DoF reconfigurable legged robot and a dataset of 200k configurations. A network with a robot signature encoder and a configuration decoder predicts configurations accurately from proprioception.
Proposes a quasi-static self-modeling framework that learns a predictive model only of high-level quasi-static dynamics rather than continuous dynamics, aimed at legged locomotion. On a 12-degree-of-freedom legged robot it improves on model-free and continuous model-based approaches. Across 80 robot morphologies the benefit correlates with complexity (R2=0.94).
Shows a task-agnostic dynamic self-model learned in a self-supervised way from a single first-person camera, without prior knowledge of morphology, kinematics or task. Trained by random motor babbling on a 12-DoF robot, the model supports forward, backward and turning locomotion without further physical training, outperforms an IMU-trained model, and lets the robot detect and recover from damage.
Introduces a self-supervised framework in which robots learn to model their morphology, kinematics and motor control from brief raw video of their own motion, much like watching a reflection. The learned self-simulation supports accurate motion planning and lets the robot detect abnormalities and recover from damage, without extensive data collection or kinematic priors.
Proposes a query-driven visual self-model that answers space-occupancy queries conditioned on the robot's state, rather than only predicting forward kinematics. Such models are continuous in space, memory efficient, differentiable and kinematics-aware. In physical experiments the model is accurate to about one percent of the workspace and supports motion planning, control, and detection of and recovery from damage.
Quantifies the benefit of self-modeling, where an agent learns a predictive model of its own dynamics, against robot complexity. Comparing Dyna-style learning to a direct learning baseline, the added value of self-modeling correlates with the robot's degrees of freedom (R2=0.90). The authors suggest this may help explain the origins of self-modeling and self-awareness.
Proposes a query-driven visual self-model that answers space occupancy queries conditioned on the robot's state, rather than modeling only forward kinematics. Such models are continuous in space, memory efficient, differentiable and kinematics-aware. In physical experiments the model was accurate to about one percent of the workspace and supported planning, control and damage detection and recovery.
A News and Views commentary arguing that the further ahead a person can mentally simulate their activities, the greater their capacity for self-awareness. It presents robots learning to self-simulate as more than an engineering convenience, calling it the beginning of an era of robot autonomy. The extracted text is fragmentary.
A Science Robotics Focus piece on a robot that modeled itself without prior knowledge of physics or its shape, then used the self-model to perform tasks and detect self-damage. It argues that self-images let robots plan internally rather than rely on hard-coded simulators that are laborious to build and become outdated.
Learns a self-model of a robot's dynamics from randomly sampled actions and uses it as the environment for reinforcement learning, so policies are trained without seeing states from the real environment. The policies transfer to the real robot, training is more data-efficient, and new tasks can be learned without further data collection.
Describes Spyndra, an open-source quadruped robot with rich proprioceptive feedback for studying machine self-awareness, based on the hypothesis that self-awareness is the ability to self-simulate by predicting sensations from actions. The paper covers hardware and software, analyzes two IMU datasets, and compares a simulated gait to a real gait as a baseline.
Proposes inferring robot dynamics models as mathematical expressions via symbolic regression and using them for motion planning and control. Experiments on underpowered pendulum domains show the inferred models enable fast planning and real-time control through rapid re-planning, with significantly better results than Support Vector Regression and Gaussian Process Regression.
2011, OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information)
A talk abstract asking whether machines can discover scientific laws automatically. It outlines a research arc from self-reflecting robots to systems that formulate hypotheses, design experiments and interpret results, with examples ranging from classical physics to biology.
Proposes a robot with two controllers: a reactive innate controller and a reflective controller that observes, models and modulates the innate one without access to its internals. The reflective controller synthesizes filters to adapt behavior to new situations. Tested with recurrent neural network innate controllers, it showed significant adaptation with few physical trials.
2009, Workshop on Performance Metrics for Intelligent Systems (PerMIS)
Explores robot behavior recovery through a process akin to self-reflection. A primary innate reactive controller is observed and modeled by a secondary reflective controller, which synthesizes input/output filters to adapt the innate controller without access to its internals. Tested on a six-legged robot under sensory, motor and control damage, the reflective controller diagnoses failure from a self-model and sensorimotor data and recovers performance.
2009, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
Applies metacognition to robots: an innate controller connects sensors to motors, while a meta controller monitors and modulates it without access to its internal state or architecture. Quantitative comparison with traditional evolutionary robotics methods suggests this self-reflection is a promising alternative for improving robot resiliency and learning.
2009, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
Explores a robotic theory of mind, using coevolution to evolve controllers that retrospectively explain an observed actor's behavior and to evolve new actions that elicit more revealing behavior. The evolved models can predict, manipulate and exploit the actor. In a physically realistic simulation the method significantly outperformed a direct estimation baseline.
2009, Workshop on Performance Metrics for Intelligent Systems (PerMIS)
Studies how robots can represent another agent's self model with neural networks and evolutionary learning, as a step toward theory of mind. The framework was tested on simulated and physical robots with a new prey-predator scenario for measuring performance. Results show the approach can recover the other agent's self models.
Describes a method for a physical robot to synthesize a body schema from multimodal sensor data gathered through interaction. The synthesis can be parametric (refining an approximate model) or topological (building a predictive model of its own body plan with little prior knowledge). The topological case succeeded with a quadruped robot after only nine 5-second interactions.
Evolves artificial neural network controllers on two cascaded field programmable analog arrays, evaluating every candidate in hardware with no simulation. Starting from random settings, the method produced controllers for biped locomotion and for restoring mobility to a damaged quadruped, aimed at low-power in-situ fault recovery for space applications.
A four-legged robot continuously builds an internal model of its own body by performing actions, observing the sensory results and refining competing candidate models. It uses the self-model to generate forward locomotion and, after a leg is damaged, updates the model and finds a new gait, demonstrating resilience without a pre-programmed model of itself.
Describes a coevolutionary algorithm that synthesizes and maintains multiple candidate models of a behaving robot. Models can generate controllers that elicit behavior under uncertainty, where models agree, or probe hidden components, where they disagree. It demonstrates model synthesis from sensor data and from scratch, and integrated model synthesis with controller design, with relevance to how organisms acquire models.
2005, IEEE Transactions on Evolutionary Computation
Presents a coevolutionary algorithm that infers topology and parameters of hidden nonlinear systems with minimal experimentation, evolving candidate models and informative tests that make models disagree. Applied to grammar induction, gene network inference, evolutionary robotics and robot damage recovery, it infers or manipulates hidden systems with few tests, with more benefit as systems grow more complex or unobservable.
Describes an exploration-estimation algorithm for automated design, analysis and repair of nonlinear physical systems with minimal physical testing. The exploratory phase evolves improvements or tests in a simulator, and the estimation phase evolves the simulator itself to explain all observations so far. It is demonstrated for analysis, design and repair of electromechanical systems.
2004, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Introduces a two-stage evolutionary algorithm for remote robots: first evolve a damage hypothesis after failure, then re-evolve a compensatory neural controller. Tested on a simulated 'physical' robot, preliminary results show nearly full functionality restored after three physical evaluations, versus over 3000 if evolved entirely on the robot. It also yields a diagnostic model of the failure.
Extends the estimation-exploration algorithm for recovering robot function after unanticipated damage or environmental change. A new fitness metric lets it describe partial as well as complete failures and tell internal damage from environmental change using only sensory feedback. In most cases only four hardware evaluations are needed to restore full functionality of the 'physical' robot.
Presents the estimation-exploration algorithm as a co-evolutionary approach to crossing the reality gap in evolutionary robotics. It adapts the robot simulator using the target robot's behavior, and adapts the robot's behavior using the simulator. Claimed benefits are automation, a minimum of hardware trials, compatibility with other transfer approaches, and generalizability to other domains.
Explores whether Truss Links, a modular robot platform developed for robot metabolism, can serve as a creative medium for non-experts. A preliminary research-through-design study shows them animating text and 2D and 3D shapes, forming kinetic sculptures, and hand-building a delta pick-and-place robot. The authors discuss what design properties make a research-grade robot accessible for creative use.
Introduces a single modular robot cell with a magnetic docking interface and two actuated joints that folds from flat into 3D structures, assembling into a three-module 'trilobite' and a five-module 'seal'. A human operator steers a parent robot to dock spare modules into a second identical robot that then separates, a budding-style reproduction. Results are preliminary and teleoperated.
Argues that open-ended physical adaptation requires robots built from a small repertoire of simple modules, so they can grow or heal by consuming parts from other machines or their surroundings and shedding broken ones. Demonstrated on a truss modular robot platform, robots grow bigger, faster and more capable by consuming material from their environment and other robots.
2024, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Proposes a meta-self-modeling approach that infers a robot's morphology from proprioceptive signals. The study introduces a 12-DoF reconfigurable legged robot and a dataset of 200k configurations. A network with a robot signature encoder and a configuration decoder predicts configurations accurately from proprioception.
2024, International Conference on Reconfigurable Mechanisms and Robots (ReMAR)
Introduces the Robot Link, a truss-robot module with magnetic connectors that can self-assemble and re-form broken connections, connecting to up to nine other links. Individual links crawl to each other and demonstrate assembly into a tetrahedron, turning 1D-locomotion modules into a 3D maneuverable robot. Three topologies recover from separation after impact, supporting robots that grow and self-repair.
2024, International Conference on Reconfigurable Mechanisms and Robots (ReMAR)
Introduces a particle robot of vibrating units, each with an off-axis spinning mass, enclosed by a rigid chain-link boundary that behaves like a soft robot. Individual particles cannot move directionally, but together they yield deterministic motion from stochastic interactions. Experiments show how particle count affects movement and direction, and the robot can pass tight spaces and flow around obstacles.
A Science Robotics editorial arguing that most robots today operate in isolation under central control. Drawing on the biological transition from individuals to societies and on swarm ideas, it discusses the case for robots that act as collectives that are more robust and capable than single units.
Asks whether layered assembly can give exponential gains in speed for block-based manufacturing. It describes two layered assembly principles that integrate n modules faster than linear time, in O(n^(2/3)) and beyond, offering a route to scalable parallel assembly of large numbers of small building blocks.
Demonstrates robots made of many loosely coupled particles that only perform uniform volumetric oscillations phase-modulated by a global signal, with no individual identity. Physical robots of up to two dozen particles and simulated ones of up to 100,000 achieved locomotion, object transport and phototaxis, and locomotion persisted with 20% of particles malfunctioning.
Describes self-replicating neural networks that replicate by learning to output their own weights, trained with gradient-based or non-gradient methods, and a regeneration method that injects the network with predictions of its own parameters. Alternating regeneration and optimization worked best. A network that also classifies MNIST shows a trade-off between classification and replication.
Addresses planning of assembly for stochastically reconfigurable systems, where module availability in space and time is uncertain. An on-line algorithm samples assembly paths to the target structure, is guaranteed to find a path for finite connected objects, pursues several paths in parallel, adapts to changing conditions and predicts remaining assembly time.
Soldercubes are small, light, low-cost self-reconfiguring modular robot modules designed for volume manufacturing. Each uses a 2 g soldering connector with no moving parts, making the system among the cheapest and lightest of comparable designs. The paper details the module design, reports lattice-configuration experiments, explores non-lattice uses, and releases all design files as open-source hardware.
Describes a solid-state, lightweight connector for modular robots that forms reversible soldered connections by heating a low melting point alloy. No external manipulation is needed, energy is used only when switching, and power and signals can pass through. Connections hold 173 N and survive hundreds of cycles before failure.
Presents a robot that autonomously traverses and manipulates a 3D truss, using translational and rotational motions across joints and specially designed building blocks made of bidirectional geared rods. Traversal plans were analyzed for speed, robustness and repeatability, with paths covering eight joints and automatic assembly and disassembly of elements demonstrated.
Explores the evolution of self-replication in three dimensions in a simulated world of physically realizable molecubes that rotate, attach into conglomerations and stochastically exchange mutable controller genomes. Self-replicating conglomerations emerged across a wide range of densities without any fitness function, informing both natural replication and physical self-replicating machines.
Presents Vacuubes, robotic modules that use vacuum as an adhesive force to form and hold structures. A first prototype was used to demonstrate vacuum sealing and the actuation of a valve that propagates vacuum between two modules. The authors expect vacuum to be a viable connection principle for modular systems needing easy attachment and detachment.
Presents a process that evolves target structures from functional requirements by minimizing strain energy under prescribed loads, and plans error-free assembly from many active components. Assembly planning samples paths by starting from the completed structure and removing accessible components one at a time, so at least one path to the target is guaranteed at every stage. It targets the simulation-to-reality gap in evolved morphologies.
Explores microbricks, interlocking 500 micrometer building blocks for assembling and reconfiguring 3D structures on a lattice. A space-filling, rotation and flip-invariant geometry was designed and fabricated by micromachining silicon and by building SU-8 polymer tiles in three layers. Ten-brick structures demonstrated interlocking and compatibility between the two materials.
2011, The International Journal of Robotics Research
Addresses on-line assembly planning for stochastically reconfigurable systems where module availability is uncertain. The proposed algorithm is guaranteed to find an assembly path for finite-sized connected objects by sampling feasible paths to the target, pursues multiple paths in parallel, and adapts during assembly to changing conditions while predicting remaining assembly time.
2011, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Demonstrates stochastic fluidic assembly of cm-scale modules in a tank, where ambient fluid flow transports the modules and an active substrate with an array of valves directs where they assemble. Valving programs with pressure-sensor feedback achieve fully automated hierarchical assembly of non-planar 3D structures, offering a route to programmable matter with simpler, smaller modules.
2010, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Presents a lattice-type modular robot concept based on a cuboctahedron element, noting that few modular robots have more than six regular faces. The conceptual design and prototypes of the modules are described. A directed graph with three parameters is proposed to represent the morphology given the various ways modules can connect.
2010, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Presents a robotic module concept for stochastic self-reconfiguring systems in which turbulent flow in a chamber drives assembly and disassembly. Thermorheological properties of Pluronic are used to route flow, the first use of thermorheological valving in three dimensions. A reversible connection using a low melting point alloy soldered in fluid, plus self-alignment, are the key innovations for scalable 3D assembly.
2010, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Presents an experimental system that stochastically assembles 3D target structures from simple 15 mm components by manipulating fluid flow in a 1.3 L tank. Modules are relieved of power, control and actuation needs, relying on stochastic environmental motion for delivery. Basic assembly and repair operations are demonstrated and initial assembly statistics are discussed.
Develops fluid-dynamically driven 3D assembly for programmable matter, studying docking of a single block onto a larger structure by simulation and experiment. A block with limited degrees of freedom aligned parallel to the docking site in 97% of trials. With self-aligning geometry and pulsating flow, 54% of assemblies were fully aligned.
Describes an efficient simulator of a modular robotic system assembling in a stochastic fluid environment, calibrated with CFD simulations and physical experiments. It is used to test assembly strategies and the effect of system parameters on assembly speed and accuracy for topologically different target structures, addressing the planning challenge from unpredictable component availability.
2009, ASME/IFToMM International Conference on Reconfigurable Mechanisms and Robots
Addresses the design of passive robotically reconfigurable truss structures whose elements can be inserted and removed singly in random-access order, and progress toward a hinge robot that manipulates them and traverses trusses. With reconfiguration algorithms, the authors envision a machine metabolism in which structures are decomposed and recomposed autonomously.
Proposes domain-independent definitions of replicability and treats self-replication as a continuous property of the interaction between a system and its environment, rather than a binary trait. Self-replicability measures how the presence of a system affects the future presence of similar systems. The framework is analyzed analytically and computationally on four systems, both discrete and continuous, formal and physical.
Develops fluid-dynamically driven assembly in three dimensions as a route to programmable matter, studying docking of a single block onto a larger structure by simulation and experiment. Limiting a block's degrees of freedom let it align parallel to the docking site in 97% of trials. Self-aligning block geometry and pulsating flow brought fully aligned assembly to 54%.
Demonstrates tuning of assembly affinities in fluidic self-assembly by thermally modulating the local viscosity around components, using a thermorheological fluid with a reversible sol-gel transition. With 500 micrometer silicon elements, the assembly point could be changed dynamically and elements selectively attracted to or rejected from a larger structure. This addresses limits of static, preprogrammed affinities.
2009, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Describes an optimal planning algorithm that morphs a grounded truss of known geometry into a new geometry, as a sequence of paths for moving truss elements while the truss stays grounded and connected. The structure grows gradually from the old one. Rigid bars are joined with 18-way connectors, and the paper also presents a truss-climbing robot able to execute the plan.
2009, ASME/IFToMM International Conference on Reconfigurable Mechanisms and Robots
Addresses algorithmic design and optimization of robotically reconfigurable structures, searching for a sequence of modifications that turns a modular structure into one serving a different function, where only the function is specified. Several internal representations are compared, and the algorithm is tested with components going missing during the build. A physical structure was built and its reconfiguration confirmed manually, pointing toward metabolic structures.
Reports two advances toward programmable matter, in which components assemble in fluid with dynamically switchable affinities to form irregular structures. One is a multi-chamber microfluidic chip for improved far-field assembly; the other is a demonstration of near-field inter-tile affinity switching using a thermorheological assembly fluid.
Shows that swarm behavior of similar complexity can be obtained from swarmers with only one controllable degree of freedom, the ability to move forward at different velocities. Interaction among units gives the swarm maneuverability the individuals lack, which may allow simpler, smaller units and larger swarms.
Presents a fluidic assembly concept that uses self-assembly forces locally but directs them globally, so arbitrary target structures can be built without tailoring the substrate or component interfaces. Controlling flow in a microfluidic chamber guides components to target locations where local interactions align and bond them, with structures of two to ten components demonstrated and quantified.
Develops an analytical and numerical model of fluid forces and torques on a two-dimensional subelement in fluidic self-assembly and its resulting motion. The model is used to find subelement shapes and flow conditions that give successful assembly across the broadest range of initial conditions, quantified by two docking parameters.
2008, IEEE International Conference on Micro Electro Mechanical Systems (MEMS)
Presents electrical and mechanical interfaces for fluidically assembled planar MEMS, including passive mechanical latches that bond microcomponents and electrical layers that connect them. The authors evaluate bonding within a microfluidic channel and circuit formation after assembly, supporting a strategy bridging bottom-up self-assembly and top-down manipulation.
Describes directed hierarchical self assembly, a two-stage procedure with coarse far-field fluidic transport of 500 micrometer silicon microtiles followed by near-field assembly controlled by on-tile jets and thermally actuated gel microvalves. Preliminary work covers CFD simulation of fluid forces and a triblock copolymer gel valving system.
Demonstrates a large space of robots capable of autonomous self-reproduction, built from actuated modules with electromagnets that control the assembly's morphology. The authors show 2-D and 3-D machines from 3 to 2^n modules, including automatically generated and manually designed morphologies, and two physical implementations each achieving two generations of reproduction.
2007, CaltechAUTHORS (California Institute of Technology)
A survey article on modular self-reconfigurable robots, which change shape by rearranging the connectivity of their parts. It outlines a taxonomy of architectures, such as lattice systems, and reviews progress from proof-of-concept systems to elaborate implementations, identifying key challenges and opportunities.
A survey in the Grand Challenges of Robotics series on modular self-reconfigurable robots, which change shape by rearranging the connectivity of their parts. It outlines a taxonomy of architectures, reviews progress over two decades, and identifies key challenges and opportunities.
Presents a microfabrication approach in which silicon microtiles self-assemble in a multilayer PDMS microchamber into arbitrary, reconfigurable structures. Controlling local flow in a microchannel array, the experiments demonstrate attraction, bonding, migration and rejection of single and paired tiles, and compare tile, outlet and medium designs. The goal is programmable digital matter.
Describes a redesign of the Fluidic Stochastic Modular Robotics system for stochastic self-assembly and reconfiguration of lattice modules in fluid, intended as a prototype for sub-millimeter implementations with MEMS counterparts. New features include controlled mechanical interconnections, active selective module rejection, flow control of inactive modules, and orientation sensing.
Describes self-reproducing modular robots built from 10-cm cubes with electromagnets that selectively weaken and strengthen connections. Fed cubes at two feeding locations, a four-module robot built a replica in 2.5 minutes by lifting and assembling cubes, reconfiguring to help build its copy, and a three-module robot reproduced in just over one minute.
Proposes a variant of nonuniform cellular automata whose rules can be realized in the electromechanical "molecube" robotic module, which the authors previously showed can self-replicate. Simple self-replication emerges spontaneously and continuously on this platform, and the authors observe qualitative differences between global state mutation and reproductive mutation.
Introduces one simulated and two physical three-dimensional stochastic modular robot systems that self-assemble and self-reconfigure in a fluid, where unpowered units move by random motion and draw power only when attached. Simulation shows that self-reconfiguration can speed assembly of a given shape, and the second physical system uses technologies that could scale to the microscale.
Proposes a physical substrate of simple modular units in which both simple and complex machines can build and be built by other machines, addressing the near absence of physical self-replication demonstrations. A number of designs for self-replicating machines are proposed, both hand-crafted and evolved.
2004, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Proposes self-reconfigurable robotics based on passive, stochastic self-organization rather than planned deterministic reconfiguration. Solid-state cellular units exploit Brownian-like motion in their environment and need no local power or locomotion, avoiding barriers to large numbers and small scales. Working prototypes are demonstrated along with preliminary analytical and computational scalability models.
Proposes a universal, continuously valued measure of self-replication, defined as a property of the interaction between a system and its environment that captures the effect of the system's presence on the future presence of similar systems. The authors give analytical and computational analyses of self-replicability factors for three distinct systems with discrete and continuous behavior.
2002, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Proposes continuously reconfigurable robotics: given a task, a robot evolves a morphology and controller, prints the assembled structure, performs the task, and is then recycled into a new form. Demonstrated for locomotion with linear actuators and sigmoidal neurons in a thermoplastic body, printed pre-assembled by rapid prototyping, with examples of working physical robots.
A book chapter considering what it would take to apply non-biological evolutionary processes at the nanoscale, where large numbers of building blocks self-organize into irregular forms that meet functional requirements and adapt to changing conditions, a form of evolutionary materials. Drawing on lessons from evolving macro-scale robotic structures, it discusses scaling, modularity, an indirect method for inducing modularity, and steps toward physical implementation.
Modularity, Evolvability and Evolutionary Computation
Foundational studies of how modularity, regularity and hierarchy arise and help evolution scale, together with algorithmic advances in evolutionary computation such as coevolution, fitness prediction, age-fitness Pareto optimization and representation design.
Presents EvoAug, a pipeline that uses generative models (conditional diffusion, few-shot NeRFs) with an efficient evolutionary algorithm to learn task-specific augmentations, structured as stochastic augmentation trees that compose augmentations hierarchically. It performs strongly on fine-grained classification and few-shot learning, and discovers augmentations aligned with domain knowledge even with little data.
Argues that adaptive imagination needs a generative world model plus a selection process that keeps imagined scenarios grounded, and tests this in few-shot vision with EvoAug, which pairs generative augmentation with evolutionary search. Without selection, generative augmentation hurts performance; with it, the system outperforms classical augmentation across six fine-grained datasets, with the largest gains where data is scarcest.
2018, Genetic and Evolutionary Computation Conference (GECCO)
Introduces Autostacker, an AutoML architecture combining hierarchical stacking with an evolutionary algorithm to search parameters, requiring no domain knowledge or feature preprocessing. It evolves pipelines that can be used directly or refined by experts, and on fifteen datasets achieves state-of-the-art or competitive accuracy and time cost against other AutoML systems.
Introduces Autostacker, an AutoML architecture pairing hierarchical stacking with an evolutionary algorithm for efficient parameter search, with no need for prior domain knowledge or feature preprocessing. Evolved pipelines can serve as is or as starting points for experts, and results on fifteen datasets are state of the art or competitive.
Presents a version of Autostacker that uses parallel hill climbing for parameter search, sharply reducing AutoML time. Its hierarchical stacking explores combinations and structures of models rather than a fixed-order pipeline and needs no domain knowledge. Experiments show better accuracy and time cost than human initial trials and a popular AutoML system.
A crowd-sourced collection of first-hand anecdotes from researchers in evolutionary computation and artificial life in which evolving algorithms or organisms surprised their creators, exposed bugs, or produced unexpected adaptations. The authors argue such stories deserve study rather than being discarded, and note that no such collection had been published before.
2017, Zenodo (CERN European Organization for Nuclear Research)
Examines why co-optimizing morphology and control is hard, noting that morphological changes tend to harm sensorimotor control. The authors introduce morphological innovation protection, which temporarily lowers selection pressure on recently changed morphologies so controllers can readapt. This avoids local optima and sustains fitness gains longer across varied initial conditions.
Editorial introducing a special issue drawn from the Artificial Life 14 conference held in Manhattan. It describes the conference program and review process, and outlines the eight extended articles selected, the first four on artificial chemistry and the last four on evolution.
Asks why evolved virtual creatures have barely grown in complexity since Sims' early work. The authors propose, grounded in embodied cognition, that co-evolving morphology and control is inherently problematic. They reproduce a prior finding that morphology converges prematurely relative to controllers and discuss it as a symptom of this problem.
Proposes a hybrid algorithm for symbolic modeling of multiple-time-scale dynamical systems, combining an evolutionary algorithm with other metaheuristics and conventional nonlinear regression. With minimal domain knowledge it turns raw observations into interpretable symbolic models, and its performance and scalability are studied across dynamics types and noise levels on physical and simulated systems.
Instead of tuning update rules, this work evolves the topology of a 3D cellular automaton while keeping one simple rule. Cells have an excited state and can contract or expand volumetrically, and an evolutionary algorithm optimizes the resulting soft 'robot' for locomotion in a simulated physical world. A variety of topologies produce effective locomotion.
2013, IEEE Transactions on Evolutionary Computation
Proposes selecting the most informative individuals in teacher-learner coevolution using the surprisal of the mean, from Shannon information theory, within a framework pairing symbolic regression with a genetic algorithm for experiment design. It inferred complex expressions from under 32 observations, used 21% fewer experiments than baselines, and reached 96.1% of passive learning performance on concrete strength using 16.6% of the data.
2013, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
A short GECCO companion summary of a Proceedings of the Royal Society B paper on why modularity evolves. It notes that modularity aids evolvability but that selection for it is indirect, and reviews the leading hypothesis for the origin of modular networks.
2013, Proceedings of the Royal Society B Biological Sciences
Clune, Mouret and Lipson show that modularity in evolved networks can arise as a side effect of selection to reduce connection costs. Networks evolved with a cost on connections become more modular, perform better and adapt faster to new environments than networks evolved for performance alone.
2012, Genetic and Evolutionary Computation Conference (GECCO)
Presents a co-evolutionary framework with pose and predictor populations for inferring kinematic poses from a single depth image, with no prior training and any kinematic structure. On 256 synthetic and 52 real images, it gave consistent poses for 34 and 78 degree-of-freedom models despite self-occlusion, with at least 3.5 times better accuracy and half the computational effort of baselines.
Uses the ENSO approach (evolution of network symmetry and modularity) to evolve modular neural controllers in simulation for a physical quadruped, with real-world uncertainty modeled as noise. Transferred controllers produced synchronized trot gaits on different surfaces and, unlike a hand-designed PID controller, generalized to voltage loss and leg loss.
Explores the evolution of self-replication in three dimensions in a simulated world of physically realizable molecubes that rotate, attach into conglomerations and stochastically exchange mutable controller genomes. Self-replicating conglomerations emerged across a wide range of densities without any fitness function, informing both natural replication and physical self-replicating machines.
2012, Genetic and Evolutionary Computation Conference (GECCO)
Proposes a divide-and-conquer approach to symbolic regression of systems with multiple time scales: decompose the data into components at different time scales, then run genetic programming on each separately. Results show improved accuracy and efficiency in reverse-engineering such dynamical systems.
Introduces a way to evolve 3D objects using compositional pattern producing networks (CPPNs), a generative encoding abstracting biological development. Interactive and target-based evolution produced complex 3D objects, and the inputs and functions available to CPPNs bias the kinds of objects produced. Objects transferred well to the real world when printed on a 3D printer.
Introduces a way to evolve 3D objects using compositional pattern producing networks (CPPNs), a generative encoding abstracting biological development. Interactive and target-based evolution produced complex 3D objects, and changing the information and functions given to CPPNs biases the results. Objects transferred well to reality when printed with a 3D printer.
Presents a method for estimating pose from a single depth image for an arbitrary kinematic skeleton without prior training, using an evolutionary algorithm to find the configuration that best explains the image. Results show correct pose estimation of 39 and 78 degree-of-freedom models, even with significant self-occlusion.
2011, IEEE Congress on Evolutionary Computation (CEC)
Studies how to choose the trainer population in coevolving rank predictors, which provide a cheap approximation of fitness in evolutionary algorithms. Four trainer selection strategies are benchmarked on symbolic regression over hundreds of test problems. Updating trainers with the solution of highest rank variance performed best, cutting computational effort four to ten fold at similar convergence.
2010, Genetic and Evolutionary Computation Conference (GECCO)
Proposes treating age as an explicit objective, evolving a population on the Pareto front of genotype age and fitness to avoid premature convergence. It is inspired by the Age Layered Population Structure algorithm. On symbolic regression across varied complexity and dimensionality, it found exact solutions more often than age-layered and standard populations, with a reported three-fold performance improvement.
2010, Genetic and Evolutionary Computation Conference (GECCO)
Proposes learning fitness approximations that predict the rank of individuals rather than real-valued fitness, using a coevolved rank predictor. Benchmarked on thousands of random symbolic regression problems against fitness prediction, it showed a five-fold reduction in computational effort for similar convergence and produced less bloated solutions.
A poster applying an evolutionary computational approach to estimate the topology of tendon networks across the fingers (the juncturae tendinum). A population of candidate network models is ranked by how well their output forces match a target network under the same input forces. The authors conclude that force distribution is sensitive to topology and that the approach can estimate topologies of real biological networks.
2009, Genetic and Evolutionary Computation Conference (GECCO)
Introduces a method that automatically identifies a small alphabet of building blocks for a problem domain. It examines solutions on the complexity-optimality Pareto front of sample systems and picks out terms that occur more often than their size would warrant. On symbolic regression of physics problems it found terms related to energy and momentum, and using them as building blocks significantly improved performance on new problems.
2009, Genetic and Evolutionary Computation Conference (GECCO)
Compares six ways of seeding expert knowledge into genetic programming for symbolic regression, using about a thousand random problems and approximate models. Inserting the whole approximate solution was not the best approach; inserting randomized pieces scaled best to complex problems, was least sensitive to approximation accuracy, and produced the least bloat.
2009, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
Proposes genetic programming to find analytical solutions to iterated functions, where f(f(x)) equals a given g(x). Applied to the hard problem f(f(x)) = x^2 - 2, the approach solved it without deep mathematical insight and found a previously unknown solution, suggesting genetic programming as a tool for such problems.
Explores symbolic regression for unsupervised learning by searching for implicit relationships f(x, y) = 0 rather than explicit y = f(x). Implicit equations can describe closed surfaces and manifolds but lack an obvious error metric. Several direct and indirect techniques were studied, and a successful method based on implicit derivatives found relationships in datasets such as circles, elliptic curves, spheres, equations of motion and energy manifolds.
2008, IEEE Transactions on Evolutionary Computation
Presents an algorithm that coevolves fitness predictors tuned to the current solution population, cutting the cost and frequency of fitness evaluation while preserving progress. Predictors need not represent objective fitness, which can adapt selection pressure and diversify solutions. In symbolic regression they scale well with problem complexity, reduce bloat and find solutions more reliably.
Proposes a coevolutionary algorithm for fitness models that addresses model learning cost, level of approximation and lack of convergence to optima. In symbolic regression it significantly outperformed other algorithms and fitness modeling approaches. It is also applied to interactive evolution of pen-stroke drawings, inferring a user's preference landscape with little interaction.
2007, Genetic and Evolutionary Computation Conference (GECCO)
Compares tree and graph encodings in genetic programming using symbolic regression on hundreds of random nonlinear target functions in 1 and 8 dimensions. Graph encodings bloat less but converge more slowly and suffer more deleterious crossovers, while offering computational benefits, a possible trade-off between regression performance and effort.
2007, Genetic and Evolutionary Computation Conference (GECCO)
Explores developmental systems in which the developing structure and its environment affect each other on similar timescales. Using a model inspired by termite nest construction, evolution exploits these dynamics to generate adaptive, self-repairing structure more efficiently than a purely reactive system. The authors also offer a metric distinguishing reactive from interactive developmental systems.
2007, Genetic and Evolutionary Computation Conference (GECCO)
Proposes a genetic programming approach to learning stochastic models with asymmetric noise, by including noise sources explicitly as stochastic elements inside the model rather than assuming noise averages out. The method decomposes nonuniform output noise into uniform stochastic elements and recovered exact analytical models with stochastic elements embedded in nonlinear and polynomial hidden systems.
Discusses how modularity, regularity and hierarchy enable scalability of open-ended evolutionary processes, defining each and noting that they are especially prominent in engineering design. The paper offers observations on properties, dependencies and trade-offs among the three and proposes a formal model for examining them.
2006, Genetic and Evolutionary Computation Conference (GECCO)
Presents an interactive multi-objective coevolutionary algorithm that evolves candidate models of the user's preferences alongside candidate probes chosen to make those models disagree. On pen-stroke drawings, preferences were extracted with pairwise questions in fewer than ten probes, and the models extrapolated to preferred drawings they were never trained on.
Studies coevolution of system models and system tests for identifying physical systems with minimal experimentation. Unmanaged test difficulty can grow faster than the models can explain, causing the two populations to disengage. Using variance of model outputs as a predictor of test difficulty keeps them engaged, demonstrated on nonlinear dynamical systems with nonlinear and linear approximation models.
2006, International Journal for Numerical Methods in Engineering
Addresses structural damage identification from sparse test data, where the inverse problem is ill-posed. A coevolutionary algorithm alternates an estimation phase, searching for damage scenarios that explain tests already done, and an exploration phase, searching for new tests that add the most information. Feasibility is shown on numerical examples.
2005, Genetic and Evolutionary Computation Conference (GECCO)
Introduces 'managed challenge' to counter disengagement in coevolutionary algorithms, here the estimation-exploration algorithm, where one population outperforms the other and selection pressure is lost. Data from futile interactions are stored in a test bank and reintroduced on re-engagement. This yields more accurate models and less training data for parametric and symbolic identification of nonlinear systems.
Applies the estimation-exploration algorithm to active learning for grammatical inference of deterministic finite automata, with the algorithm proposing its own training data. It outperforms one version of evidence-driven state merging on random DFAs, which only works well for particular balances of positive labelings. The authors propose a more general way of generating DFAs for benchmarking.
Introduces an estimation of distribution algorithm that coevolves fitness maximizers with fitness predictors, lightweight objects that approximate true fitness at reduced computation, to speed evolutionary search when evaluations are costly. Applied to symbolic regression, it improves speed and convergence while reducing evaluations, and the advantage grows with model complexity.
2005, IEEE Transactions on Evolutionary Computation
Presents a coevolutionary algorithm that infers topology and parameters of hidden nonlinear systems with minimal experimentation, evolving candidate models and informative tests that make models disagree. Applied to grammar induction, gene network inference, evolutionary robotics and robot damage recovery, it infers or manipulates hidden systems with few tests, with more benefit as systems grow more complex or unobservable.
Shows that genetic encoding influences which alternative optimal designs evolution finds when many optima exist. The test problem is designing a 2D shape that keeps constant height when rolled on flat ground; Reuleaux showed in the 19th century that the circle is not the only solution, and artificial evolution repeats that discovery in under an hour.
Proposes a variant of nonuniform cellular automata whose rules can be realized in the electromechanical "molecube" robotic module, which the authors previously showed can self-replicate. Simple self-replication emerges spontaneously and continuously on this platform, and the authors observe qualitative differences between global state mutation and reproductive mutation.
Studies the link between network structure and dynamics using linear network dynamics. Modularity and hierarchical organization emerge in evolved networks with asymptotically stable linear dynamics. Numerical experiments show linear stability benefits from a hierarchy of modules, and that this architecture makes stability more robust to random perturbations in network structure.
Proposes that a problem's Hessian matrix reveals genetic linkage, and that eigenstructure analysis of the Hessian yields a transformation to a space where first-order linkage is optimal. Genetic algorithms that dynamically transform the problem space can promote building blocks more efficiently. The approach is demonstrated on a real-valued adaptation of Kauffman's NK landscapes.
2003, IEEE Transactions on Robotics and Automation
Presents generative representations for automated design of modular physical robots. The resulting machines remain simple compared with human-engineered ones, but are more regular, modular and hierarchical than previously evolved machines of similar function, and virtual designs have an order of magnitude more moving parts. The authors quantitatively show the generative representation searches more efficiently than a non-generative one.
2003, Dépôt institutionnel de l'Université libre de Bruxelles (Université Libre de Bruxelles)
Addresses the lack of mechanisms in genetic algorithms for building hierarchically organized solutions, which would need grouping or social behavior that is maladaptive under competition. The paper analyzes and models a biological multi-level selection model that allows such behavior to emerge, as a first step toward computational synthesis of complex structures.
2002, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Argues that automatic design systems must produce designs from reused modules to scale in complexity, and proposes a generative design grammar subject to evolution. Combining Lindenmayer systems with evolutionary algorithms gives hierarchical modularity. The system is demonstrated for the first time on physical, modular, 2D locomoting robots and their controllers.
Studies the dynamics of modularization in a minimal substrate, using simple simulations to suggest that modularity arises spontaneously in response to variation. The amount of modular separation is logarithmically proportional to the rate of variation, implying modular architectures should correlate with high environmental change rates. The authors suggest engineering evolution would benefit from variable fitness criteria.
A book chapter considering what it would take to apply non-biological evolutionary processes at the nanoscale, where large numbers of building blocks self-organize into irregular forms that meet functional requirements and adapt to changing conditions, a form of evolutionary materials. Drawing on lessons from evolving macro-scale robotic structures, it discusses scaling, modularity, an indirect method for inducing modularity, and steps toward physical implementation.
Considers how to promote modularity so evolutionary design can scale to higher complexity through reuse of building blocks. Proposes a weak, indirect mechanism based on the observation that modular designs adapt better and survive more under changing requirements, requiring no representation of partial solutions. The principle is demonstrated on an abstract design problem where modularity can be quantified.
Automated Scientific Discovery and Symbolic Regression
Algorithms that distill equations, laws and models directly from experimental data: symbolic regression, coevolution of models and tests for system identification, discovery of hidden state variables, and data-driven analysis of stochastic processes.
2026, IEEE International Conference on Acoustics Speech and Signal Processing
Proposes LyTimeT, a two-phase framework for extracting dynamical state variables from video despite visual distractors. A TimeSformer autoencoder learns robust latent states, then physically meaningful dimensions are selected by correlation analysis and dynamics refined with a Lyapunov stability regularizer. On five synthetic and four real systems it gives estimates closest to ground truth and the lowest error among baselines.
A large multi-institution team, including Lipson, proposes applying the common task framework (shared datasets, tasks and evaluation metrics, as used to drive progress in machine learning) to accelerate AI-driven scientific discovery in dynamical systems and other areas of science.
Describes the mission of the NSF AI Institute in Dynamic Systems: developing machine learning tools that find physically interpretable, physics-constrained data-driven models for controlling complex systems, via optimal sensor selection and placement. A common task framework evaluates algorithms across engineering tasks and supports open-source challenge datasets.
Proposes a general principle for distilling hidden state variables that are more aligned with human intuition, without relying on prior physical knowledge. Since AI-found variables are often hard to interpret, the authors test on experimental and simulated systems, where the variables generated closely resemble those chosen independently by human scientists.
Addresses the problem of identifying the state variables needed before physical laws can be formulated, by having an AI extract them from raw experimental data such as video. The framework finds a compact set of variables describing a dynamical system without prior physical knowledge. The text motivates the work with the historical role of defining variables like mass or temperature.
Proposes a principle for determining how many state variables a system has and what they might be, directly from high-dimensional video. Applied to recordings of systems from elastic double pendulums to fire flames, the algorithm finds the intrinsic dimension of the dynamics and candidate sets of state variables without prior knowledge of the physics.
Proposes a hybrid algorithm for symbolic modeling of multiple-time-scale dynamical systems, combining an evolutionary algorithm with other metaheuristics and conventional nonlinear regression. With minimal domain knowledge it turns raw observations into interpretable symbolic models, and its performance and scalability are studied across dynamics types and noise levels on physical and simulated systems.
Presents a new, faster-converging approach to estimating the entropy rate of symbol sources, tested from English text to chaotic dynamical systems. Using a connection to probabilistic automata, the authors establish a convergence rate of O(log|s|/cube root of |s|), which yields explicit uncertainty estimates and required data lengths.
Proposes data smashing, a feature-free way to measure similarity between arbitrary data streams. Using an algebraic structure on symbolic models, each stream has an anti-stream, and similarity is how fully one annihilates the other's statistical structure into noise. Applied to epileptic brainwaves, cardiac rhythms and astronomical photometry, it matched or exceeded expert-tuned algorithms.
Journal version of the data smashing work: a universal principle for quantifying similarity between data sources without features, training or prior knowledge. Stochastic symbolic generators can be added and inverted, so each stream has an anti-stream. Applications in seizure brainwaves, cardiac anomalies and astronomical classification meet or exceed domain-tuned methods.
Presents an initial algorithm to infer direction-specific causal dependencies among sequence positions in retroviral genome evolution, without prior knowledge of the viral genomes. It distills a network of causality flows and identifies key regions of immunological vulnerability that could inform vaccine design for highly mutable viruses such as HIV.
Investigates fitness evaluation approaches for evolving nonlinear analog circuits when reverse-engineering an arbitrary black-box circuit. An incremental algorithm outperformed naive approaches, and the evolved robust circuits closely matched the black box's time-domain output for any time-domain input.
2013, Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences
Chattopadhyay and Lipson present an algorithm for learning probabilistic finite automata from quantized time series, giving compact generative models of stochastic processes and a basis for comparing and classifying data streams. Published in the Philosophical Transactions of the Royal Society A.
2013, Proceedings of the National Academy of Sciences
Addresses the inverse of Gillespie stochastic simulation: reconstructing the underlying reactions from sparse, intermittent aggregate observations. The authors find that relative population updates form a convex polytope whose vertices indicate the dominant reactions, and reconstruct models (structure and propensities) from simulated and experimental systems, sometimes giving mechanistic insight.
2013, IEEE Transactions on Evolutionary Computation
Proposes selecting the most informative individuals in teacher-learner coevolution using the surprisal of the mean, from Shannon information theory, within a framework pairing symbolic regression with a genetic algorithm for experiment design. It inferred complex expressions from under 32 observations, used 21% fewer experiments than baselines, and reached 96.1% of passive learning performance on concrete strength using 16.6% of the data.
Applies genetic-programming symbolic regression to infer analytical expressions for tendon excursions and moment arms from sparse data, replacing conventional polynomial regression. Tested on a tendon-driven robotic system and synthetic musculoskeletal data, it needed less training data, extrapolated better, tolerated noise and used fewer parameters than polynomials.
Infers the structure and parameters of the finger tendon network through sparse force experiments, co-evolving candidate models with informative future force inputs via the estimation-exploration algorithm. Validated on two known latex networks and a cadaveric middle finger, with training error under 4% and cross-validation error under 7.2% for the cadaver specimen.
Uses symbolic regression to infer both the form and parameters of analytical functions for tendon excursions in the index finger, instead of assuming polynomials. Data came from applying constant tension to seven tendons of a cadaveric hand and moving the finger through its range of motion.
Investigates using genetic programming to construct ODE models of dynamical systems from time series when some variables are unmeasured. Synthetic and physical examples show the challenges and the circumstances under which both the form and parameters of models with hidden variables can be reverse-engineered.
Presents multi-modal symbolic regression (MMSR) for learning symbolic models of hybrid dynamical systems from unlabeled time series. It combines clustered symbolic regression, which identifies distinct behaviors and their equations, with transition modeling that infers symbolic inequalities as boundaries. It beat other multi-modal approaches on four synthetic sets and recovered classical transistor modes from measurements.
2012, Genetic and Evolutionary Computation Conference (GECCO)
Proposes a divide-and-conquer approach to symbolic regression of systems with multiple time scales: decompose the data into components at different time scales, then run genetic programming on each separately. Results show improved accuracy and efficiency in reverse-engineering such dynamical systems.
Presents a data-driven process for estimating nonlinear material models for multi-material 3D printed materials. Rather than classical engineering stress-strain assuming constant cross-section, it runs virtual tensile tests by finite element analysis on the full 3D specimen geometry and fits material model parameters by minimizing the discrepancy with experimental results. Preliminary results are reported.
2011, Genetic and Evolutionary Computation Conference (GECCO)
Introduces an algorithm to infer exact stochastic reaction models from sparse measurements of systems of discrete quantities, such as molecules. Candidate models are simulated to maximize data likelihood, and when likelihood is too small the distance of data to the model's estimated distribution is used. It infers models reliably for both short and long measurement gaps and outperformed likelihood or distance alone.
Demonstrates automated refinement and inference of analytical models of metabolic networks from time-series data, using a yeast glycolytic oscillation model as the test system. The method corrects errors in approximated and overspecified models, can sometimes identify the full model from scratch, and designs experiments that make candidate predictions disagree. It handled high noise and predicted better than parametric regression and neural networks.
Proposes inferring robot dynamics models as mathematical expressions via symbolic regression and using them for motion planning and control. Experiments on underpowered pendulum domains show the inferred models enable fast planning and real-time control through rapid re-planning, with significantly better results than Support Vector Regression and Gaussian Process Regression.
2011, OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information)
A talk abstract asking whether machines can discover scientific laws automatically. It outlines a research arc from self-reflecting robots to systems that formulate hypotheses, design experiments and interpret results, with examples ranging from classical physics to biology.
2010, Genetic and Evolutionary Computation Conference (GECCO)
Proposes treating age as an explicit objective, evolving a population on the Pareto front of genotype age and fitness to avoid premature convergence. It is inspired by the Age Layered Population Structure algorithm. On symbolic regression across varied complexity and dimensionality, it found exact solutions more often than age-layered and standard populations, with a reported three-fold performance improvement.
Uses symbolic regression to determine both the form and parameters of tendon excursion versus joint angle relationships that define moment arms, instead of assuming a polynomial form. Demonstrated on synthetic data from a known human index finger model, cross-validation with realistic noise shows the method can recover the correct form and parameters of nonlinear relationships.
Proposes symbolic regression to find both the mathematical form and parameters of tendon excursion-joint angle relationships for index finger kinematics, rather than fitting parameters of an assumed form. Tested on synthetic data from a known finger model, with realistic noise, it extracts the correct form and values even in the presence of noise.
A short abstract describing automated search for dynamical invariants in experimental data as a route to conservation laws rather than only predictive models. Applied to motion-tracking data from systems from simple oscillators to chaotic double pendula, the algorithm found Hamiltonians, Lagrangians and momentum conservation laws without prior physics knowledge. Laws from simpler systems helped bootstrap more complex ones.
2010, IEEE Aerospace Conference (AERO) · no PDF yet
An IEEE Aerospace Conference paper on robotic and computational systems that discover scientific laws from data on their own, connecting self-modeling robots with automated scientific discovery.
No PDF yet: the file found earlier was a Spanish news web page (RTVE), not a paper; it has been removed.
A rebuttal to a comment by Hillar and Sommer on the Schmidt and Lipson paper on distilling free-form natural laws from data. The text lists the commenters' criticisms, including nonstandard terminology, missing theoretical justification, a fitness function claimed to be flat for general systems, and uncited prior work; the rebuttal itself is not in the extracted text.
2009, Genetic and Evolutionary Computation Conference (GECCO)
Introduces a method that automatically identifies a small alphabet of building blocks for a problem domain. It examines solutions on the complexity-optimality Pareto front of sample systems and picks out terms that occur more often than their size would warrant. On symbolic regression of physics problems it found terms related to energy and momentum, and using them as building blocks significantly improved performance on new problems.
Schmidt and Lipson show that a symbolic regression algorithm can distill free-form natural laws, such as conservation laws and equations of motion, directly from motion-tracking data of physical systems like oscillators and pendulums, without prior knowledge of physics.
2009, Genetic and Evolutionary Computation Conference (GECCO)
Compares six ways of seeding expert knowledge into genetic programming for symbolic regression, using about a thousand random problems and approximate models. Inserting the whole approximate solution was not the best approach; inserting randomized pieces scaled best to complex problems, was least sensitive to approximation accuracy, and produced the least bloat.
2009, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
Proposes genetic programming to find analytical solutions to iterated functions, where f(f(x)) equals a given g(x). Applied to the hard problem f(f(x)) = x^2 - 2, the approach solved it without deep mathematical insight and found a previously unknown solution, suggesting genetic programming as a tool for such problems.
Explores symbolic regression for unsupervised learning by searching for implicit relationships f(x, y) = 0 rather than explicit y = f(x). Implicit equations can describe closed surfaces and manifolds but lack an obvious error metric. Several direct and indirect techniques were studied, and a successful method based on implicit derivatives found relationships in datasets such as circles, elliptic curves, spheres, equations of motion and energy manifolds.
2008, IEEE Transactions on Evolutionary Computation
Presents an algorithm that coevolves fitness predictors tuned to the current solution population, cutting the cost and frequency of fitness evaluation while preserving progress. Predictors need not represent objective fitness, which can adapt selection pressure and diversify solutions. In symbolic regression they scale well with problem complexity, reduce bloat and find solutions more reliably.
Describes an algorithm that uses genetic programming, representing differential equations as acyclic graphs of primitive operations, to reverse-engineer symbolic models of dynamical systems from observations for modeling, control and exploratory analysis. The core algorithm and its application to system identification are shown on nonlinear mechanical and biological systems.
2007, Proceedings of the National Academy of Sciences
Introduces a method that automatically generates symbolic differential equations for coupled nonlinear dynamical systems directly from time series. It models each variable separately, perturbs and destabilizes the system to expose hidden behavior, and simplifies equations during modeling. It is demonstrated on four simulated and two real systems spanning mechanics, ecology and systems biology.
Argues that biomechanical modeling should explore model topology as well as parameters, using finger tendon networks as the example. It presents a simulator for elastic structures under large deformation, shows that assumed topology affects tension propagation to joints, and demonstrates an inference algorithm recovering hidden synthetic tendon network topologies, with discussion of observability and uniqueness.
Proposes a coevolutionary algorithm for fitness models that addresses model learning cost, level of approximation and lack of convergence to optima. In symbolic regression it significantly outperformed other algorithms and fitness modeling approaches. It is also applied to interactive evolution of pen-stroke drawings, inferring a user's preference landscape with little interaction.
2007, Genetic and Evolutionary Computation Conference (GECCO)
Compares tree and graph encodings in genetic programming using symbolic regression on hundreds of random nonlinear target functions in 1 and 8 dimensions. Graph encodings bloat less but converge more slowly and suffer more deleterious crossovers, while offering computational benefits, a possible trade-off between regression performance and effort.
2007, Genetic and Evolutionary Computation Conference (GECCO)
Proposes a genetic programming approach to learning stochastic models with asymmetric noise, by including noise sources explicitly as stochastic elements inside the model rather than assuming noise averages out. The method decomposes nonuniform output noise into uniform stochastic elements and recovered exact analytical models with stochastic elements embedded in nonlinear and polynomial hidden systems.
Studies coevolution of system models and system tests for identifying physical systems with minimal experimentation. Unmanaged test difficulty can grow faster than the models can explain, causing the two populations to disengage. Using variance of model outputs as a predictor of test difficulty keeps them engaged, demonstrated on nonlinear dynamical systems with nonlinear and linear approximation models.
2006, International Journal for Numerical Methods in Engineering
Addresses structural damage identification from sparse test data, where the inverse problem is ill-posed. A coevolutionary algorithm alternates an estimation phase, searching for damage scenarios that explain tests already done, and an exploration phase, searching for new tests that add the most information. Feasibility is shown on numerical examples.
2005, Genetic and Evolutionary Computation Conference (GECCO)
Introduces 'managed challenge' to counter disengagement in coevolutionary algorithms, here the estimation-exploration algorithm, where one population outperforms the other and selection pressure is lost. Data from futile interactions are stored in a test bank and reintroduced on re-engagement. This yields more accurate models and less training data for parametric and symbolic identification of nonlinear systems.
Applies the estimation-exploration algorithm to active learning for grammatical inference of deterministic finite automata, with the algorithm proposing its own training data. It outperforms one version of evidence-driven state merging on random DFAs, which only works well for particular balances of positive labelings. The authors propose a more general way of generating DFAs for benchmarking.
Describes a coevolutionary algorithm that synthesizes and maintains multiple candidate models of a behaving robot. Models can generate controllers that elicit behavior under uncertainty, where models agree, or probe hidden components, where they disagree. It demonstrates model synthesis from sensor data and from scratch, and integrated model synthesis with controller design, with relevance to how organisms acquire models.
Demonstrates inferring the topology of a hidden nonlinear network of cable-like, tension-only links from limited data. Hidden networks follow two classical tendon-network representations of a human finger, with 11 and 16 links. An evolutionary algorithm modifying a 72-link mesh used 30 data points to evolve networks reproducing the hidden functional behavior and closely matching its topology.
Introduces an estimation of distribution algorithm that coevolves fitness maximizers with fitness predictors, lightweight objects that approximate true fitness at reduced computation, to speed evolutionary search when evaluations are costly. Applied to symbolic regression, it improves speed and convergence while reducing evaluations, and the advantage grows with model complexity.
2005, IEEE Transactions on Evolutionary Computation
Presents a coevolutionary algorithm that infers topology and parameters of hidden nonlinear systems with minimal experimentation, evolving candidate models and informative tests that make models disagree. Applied to grammar induction, gene network inference, evolutionary robotics and robot damage recovery, it infers or manipulates hidden systems with few tests, with more benefit as systems grow more complex or unobservable.
Describes an exploration-estimation algorithm for automated design, analysis and repair of nonlinear physical systems with minimal physical testing. The exploratory phase evolves improvements or tests in a simulator, and the estimation phase evolves the simulator itself to explain all observations so far. It is demonstrated for analysis, design and repair of electromechanical systems.
Presents an evolutionary estimation-exploration algorithm for reverse engineering hidden networks from inputs and outputs using minimal tests. One stage evolves a system hypothesis and the other evolves the next most informative experiment. It is demonstrated for inferring gene regulatory networks without costly knockout studies and for inferring robot morphological properties without extensive physical testing.
Discusses high-order neurons in competitive learning, where geometric shapes replace point neurons through high-order tensors in homogeneous coordinates. Eigentensors of the synaptic tensors reveal coefficients of polynomial rules the network carries out, which may expose symbolic structure in data. The neurons use maximum-correlation activation with local Hebbian learning and separate close, overlapping clusters.
Applying automated modeling and inference to biology: tendon networks of the fingers, metabolic and genetic networks, ion channels, and molecular evolution.
Presents an initial algorithm to infer direction-specific causal dependencies among sequence positions in retroviral genome evolution, without prior knowledge of the viral genomes. It distills a network of causality flows and identifies key regions of immunological vulnerability that could inform vaccine design for highly mutable viruses such as HIV.
2013, Proceedings of the National Academy of Sciences
Addresses the inverse of Gillespie stochastic simulation: reconstructing the underlying reactions from sparse, intermittent aggregate observations. The authors find that relative population updates form a convex polytope whose vertices indicate the dominant reactions, and reconstruct models (structure and propensities) from simulated and experimental systems, sometimes giving mechanistic insight.
2013, Proceedings of the Royal Society B Biological Sciences
Clune, Mouret and Lipson show that modularity in evolved networks can arise as a side effect of selection to reduce connection costs. Networks evolved with a cost on connections become more modular, perform better and adapt faster to new environments than networks evolved for performance alone.
Applies genetic-programming symbolic regression to infer analytical expressions for tendon excursions and moment arms from sparse data, replacing conventional polynomial regression. Tested on a tendon-driven robotic system and synthetic musculoskeletal data, it needed less training data, extrapolated better, tolerated noise and used fewer parameters than polynomials.
Infers the structure and parameters of the finger tendon network through sparse force experiments, co-evolving candidate models with informative future force inputs via the estimation-exploration algorithm. Validated on two known latex networks and a cadaveric middle finger, with training error under 4% and cross-validation error under 7.2% for the cadaver specimen.
Uses symbolic regression to infer both the form and parameters of analytical functions for tendon excursions in the index finger, instead of assuming polynomials. Data came from applying constant tension to seven tendons of a cadaveric hand and moving the finger through its range of motion.
2011, Genetic and Evolutionary Computation Conference (GECCO)
Introduces an algorithm to infer exact stochastic reaction models from sparse measurements of systems of discrete quantities, such as molecules. Candidate models are simulated to maximize data likelihood, and when likelihood is too small the distance of data to the model's estimated distribution is used. It infers models reliably for both short and long measurement gaps and outperformed likelihood or distance alone.
Demonstrates automated refinement and inference of analytical models of metabolic networks from time-series data, using a yeast glycolytic oscillation model as the test system. The method corrects errors in approximated and overspecified models, can sometimes identify the full model from scratch, and designs experiments that make candidate predictions disagree. It handled high noise and predicted better than parametric regression and neural networks.
2011, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
Uses evolutionary circuit design to automatically build equivalent analog circuits for neuronal ion channels, emulating the creativity behind the Hodgkin-Huxley model. From electrophysiological observations alone, the system evolved circuits that accurately simulate sodium and potassium channel behavior. It may serve as a reverse-engineering tool and as an aid in designing neuromorphic circuits.
Uses evolutionary circuit design to automatically construct equivalent analog circuits for neuronal ion channels, emulating the creative leap behind the Hodgkin-Huxley model. From electrophysiological observations alone, the system evolved circuits of basic electronic components that accurately simulate sodium and potassium channel behavior. It could serve as a reverse-engineering tool or assist in designing neuromorphic circuits.
Uses symbolic regression to determine both the form and parameters of tendon excursion versus joint angle relationships that define moment arms, instead of assuming a polynomial form. Demonstrated on synthetic data from a known human index finger model, cross-validation with realistic noise shows the method can recover the correct form and parameters of nonlinear relationships.
Proposes symbolic regression to find both the mathematical form and parameters of tendon excursion-joint angle relationships for index finger kinematics, rather than fitting parameters of an assumed form. Tested on synthetic data from a known finger model, with realistic noise, it extracts the correct form and values even in the presence of noise.
Presents a finite element solver for mechanical force transmission through tendon networks wrapped on arbitrarily shaped, frictionless rigid objects such as bones, aimed at the tendinous networks controlling the fingers. Its speed makes it suitable as the simulation engine for evolutionary algorithms that infer network structure from sparse experiments.
A poster applying an evolutionary computational approach to estimate the topology of tendon networks across the fingers (the juncturae tendinum). A population of candidate network models is ranked by how well their output forces match a target network under the same input forces. The authors conclude that force distribution is sensitive to topology and that the approach can estimate topologies of real biological networks.
2007, Proceedings of the National Academy of Sciences
Introduces a method that automatically generates symbolic differential equations for coupled nonlinear dynamical systems directly from time series. It models each variable separately, perturbs and destabilizes the system to expose hidden behavior, and simplifies equations during modeling. It is demonstrated on four simulated and two real systems spanning mechanics, ecology and systems biology.
Argues that biomechanical modeling should explore model topology as well as parameters, using finger tendon networks as the example. It presents a simulator for elastic structures under large deformation, shows that assumed topology affects tension propagation to joints, and demonstrates an inference algorithm recovering hidden synthetic tendon network topologies, with discussion of observability and uniqueness.
Cadaveric experiments and simulations indicate that the tendon network of the fingers performs logic-like computation. Varying tension in the extensor digitorum versus the interosseous tendons changed how tension propagated to the finger joints nonlinearly, and modeling attributes this to the network's deformable structure, suggesting a non-neural form of information processing.
2005, Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC)
Describes a biomechanical modeling environment that predicts tension distribution in the finger extensor mechanism, the tendinous network over the phalanges, as part of work to simulate finger function and precision pinch. Predictions show tension distribution can be very sensitive to the assumed network topology, meaning the number of elements and their connectivity.
Describes a relaxation-based method for simulating 2D and 3D compound kinematic mechanisms, in which iterative propagation of node displacements reveals the kinematics. It is slower than simultaneous differential equation approaches but more accurate, more robust to singularities, and able to handle complex entangled chains and nonlinear links such as wrapping strings and contacts. Applications include robotics, design and biomechanics.
Demonstrates inferring the topology of a hidden nonlinear network of cable-like, tension-only links from limited data. Hidden networks follow two classical tendon-network representations of a human finger, with 11 and 16 links. An evolutionary algorithm modifying a 72-link mesh used 30 data points to evolve networks reproducing the hidden functional behavior and closely matching its topology.
Presents an evolutionary estimation-exploration algorithm for reverse engineering hidden networks from inputs and outputs using minimal tests. One stage evolves a system hypothesis and the other evolves the next most informative experiment. It is demonstrated for inferring gene regulatory networks without costly knockout studies and for inferring robot morphological properties without extensive physical testing.
3D printing processes, personal fabrication (Fab@Home), multi-material and voxel-based digital materials, file formats such as AMF, laser sintering and other manufacturing methods, plus commentary on the future of additive manufacturing.
Introduces a self-deploying formwork for concrete built from multistable tubes that expand from a compact state into a full-scale mold with integrated steel reinforcement. A steel-reinforced structure 2.36 m high is deployed within 14 seconds, and the polymer shell of the tubes triples structural strength. The concept aims at low-carbon construction with minimal on-site intervention.
Presents a physics-based simulation of the deposition process in multi-ingredient food 3D printing, taking G-code as input and using the Bifrost plug-in for Autodesk Maya. It is validated by reproducing a custom seven-ingredient dessert. Simulating designs before printing helps build intuition for stable structures, reduces waste and speeds proofing.
Presents a method for designing and fabricating 3D carbon fiber lattices by weaving carbon fiber tow through two parallel grids, reinforced with vertical carbon fiber tubes, inspired by crochet and sewing. Compression tests on three designs perform similarly to comparable lattices, and finite element analysis supports the results. Weave density can vary within one lattice, and the process could be automated.
Demonstrates inverted laser sintering for metal powders, which avoids a powder bed. A 10-layer copper part is built by depositing powder on a sapphire plate, pressing it against the part and sintering it with a 14 W 445 nm laser through the glass, repeated layer by layer. The authors discuss applications such as directly fabricating multi-material metal parts.
Proposes selective curing of carbon fiber-reinforced polymer laminates with blue and red lasers as an alternative to slow, energy-intensive conventional curing, aimed at deployable structures. A blue laser cures parts of a flat design to make hinges so it pops up, while a red laser ablates epoxy to create flexible regions. A pop-up CFRP lattice and a ruled-surface structure are developed.
Proposes embedding electronic components during selective laser sintering using an inverted laser sintering process that needs no powder bed, so voids can be created easily. An LED is placed into a cavity within an SLS part and printing resumes. The LED still works after several layers are added, pointing toward SLS producing active systems rather than passive parts.
Combines powder-based and resin-based additive manufacturing in one process, using an inverted setup in which a laser shines upward through glass to fuse powder and also cure resin. Samples containing both materials are fabricated. The authors also examine using resin to reduce porosity of powder-bed parts and to adjust surface finish.
Presents a generative approach to optimizing 3D lattices of thousands of struts for mass and deflection. It iteratively erodes struts judged unnecessary by stress analysis and randomly regenerates new struts across the structure. Strength-to-weight ratios are on par with state-of-the-art topology optimization but qualitatively different; a GPU dynamics simulator also supports dynamic load cases.
Demonstrates parallel electro-osmotic grippers for layered assembly, where each array element is controlled individually via capillary effects driven by an electric field to form droplet patterns that pick and place voxels. A 25-element array assembled 1.5 mm spheres with 95 to 98% grasping reliability, validated against analytical and computational models.
Presents PARA, a mostly 3D printed, open-source three-degree-of-freedom robotic arm that lifts 2 kg at a 940 mm reach, with about 2.6 mm precision at 250 mm/s without load. It costs about $3,400 to build, an order of magnitude below comparable market arms, and demonstrates printed parts as practical industrial tools.
Presents an optimization strategy, with experimental validation, for two-dimensional electroadhesors that grip millimeter-scale objects. Using 3D electrostatic simulation and weight-lifting tests at 600 V on 1.6 mm thick objects, the optimal pad spacing was 0.5 mm and the optimal pad-area ratio 0.75 for conductive and insulating objects. A step-by-step design method follows.
Presents an additive process in which an upward-directed laser fuses a powder monolayer onto a substrate through a clear surface, building a hanging part layer by layer. It removes the need for a large powder bed and permits different powders in one layer. A 445 nm laser made a Nylon-12 and TPU pattern, whose mechanical properties were characterized.
Asks whether layered assembly can give exponential gains in speed for block-based manufacturing. It describes two layered assembly principles that integrate n modules faster than linear time, in O(n^(2/3)) and beyond, offering a route to scalable parallel assembly of large numbers of small building blocks.
Explores laser cooking as a software-controlled way to cook printed food layers, using mirror galvanometers to direct a 445 nm blue laser and a 980 nm near-infrared laser. By varying power, speed and pattern, the authors produced a software-generated three-course meal, and present digital cooking as a new tool for chefs and makers.
Designs 2 x 2 arrays of electroadhesive grippers, with comb and spiral electrodes, to place premanufactured voxels in parallel for Layered Assembly, a multi-material additive process. Using field simulations and tests at 75 to 800 V on polymer and metal voxels, comb-shaped electrodes performed better, with about 100% reliability at 600 to 800 V.
Presents a workflow to replicate both surface and internal color texture of wood using destructive tomographic imaging and voxel printing. Wood was sliced at 27 micrometer intervals and imaged with a camera on a CNC mill, and the 230 images were dithered for a voxel printer. The printed block resembles the wood externally and internally when cut.
Illustration
Multi-Material Three-Dimensional Food Printing with Simultaneous Infrared Cooking DOI
2018, 3D Printing and Additive Manufacturing · no PDF yet
A study of multi-material 3D food printing combined with infrared heating that cooks the food as it is printed.
No PDF yet: the file found earlier was a different paper, "Characterization of CO2 laser browning of dough"; it has been removed.
Demonstrates a 1D printing system, inspired by the ribosome, that folds a single string of source material into specialized robotic automata. A computational optimization tool derives designs from high-level goals. Three distinct robots for different locomotion tasks were made from the same material, and recycled material was reused to produce new designs.
Describes producing cellular foam structures implicitly by inducing viscous thread instability while extruding material, rather than designing each cell explicitly. The process yields structures finer than the printer's native resolution, with effective elastic modulus and density tunable across two orders of magnitude. Applications cited include bioengineering, robotics and food printing.
A moderated roundtable with Jennifer Lewis, Robert Shepherd and Hod Lipson on the prospects and limits of 3D printing soft materials. Participants describe their use of printing for robots and functional materials, from hard plastics to soft and active materials, conductors, batteries, actuators, biomaterials and food.
An editorial recalling the 2000 evolved, 3D-printed robot that walked out of the printer, designed through evolution in simulation and printed whole and preassembled. It poses the grand challenge for additive manufacturing of printing a complete functioning robot.
Explores additive manufacturing for food-safe applications by producing an entirely additively manufactured dinner, including complex tableware. It notes that current AM lacks the process control and certification for widespread food contact use, with coatings mitigating concerns for prototypes and small runs, and lays out barriers to adoption.
Demonstrates additive manufacturing for the hospitality industry by producing an entirely additively manufactured dinner with complex tableware for limited audiences. It observes that current techniques lack the process control and certification needed for widespread food contact use, that coatings can help, and lists barriers for further work.
An editorial on material complexity in 3D printing: hundreds of process parameters affect the properties of printed materials and few are fully characterized. It argues that data mining of the relationship between printer settings and material properties can address this and help engineers trust printed parts.
Second tutorial in a series on the AMF (ISO/ASTM 52915) additive manufacturing file format, covering how to specify colors and color textures. It motivates the topic with the history of color in 3D printing and the importance of color in multimaterial printing.
First tutorial in a series on the AMF additive manufacturing format, walking through a complete file describing a simple two-material pyramid. It explains the limitations of STL, such as redundancy, unit ambiguity, poor scaling and no support for color or multiple materials, that motivated the ISO/ASTM 52915 standard.
2014, The International Journal of Robotics Research
Introduces Bitblox, small modular interconnecting blocks embedding simple electromechanical connectivity and functionality as printable digital materials. Different blocks, and their combination and positions, set the mechanical and electrical properties of an assembly. The paper describes the design, compares it with similar materials and demonstrates use in a working 3D printer.
Journal version of the open-source simulation engine for statics, dynamics and nonlinear deformation of heterogeneous soft bodies. It supports interspersed materials with varied stiffness, density, Poisson ratio, thermal expansion and friction, handles collisions, uses volumetric actuation, and simulates thousands of degrees of freedom at interactive rates, aimed at multimaterial design automation.
An editorial arguing that additive manufacturing is a real industrial revolution. Past revolutions each made a major cost driver, such as power or calculation, nearly vanish, and the editorial applies that pattern to 3D printing despite its market being far smaller than conventional manufacturing.
An editorial arguing that additive manufacturing is driving a revolution in design as well as manufacturing. We can now print nearly any shape we can describe, but CAD software, analysis tools and designers' imagination lag, so design tools must catch up to exploit the technology.
Uses multimaterial solid freeform fabrication to build soft tissue simulators that replicate the injection force profile of a needle stick, using blunted needles that are safer for training. Composite structures and needle pairs gave a puncture force of 2.8 N at 9 to 15 mm depth, within the reported range for human skin.
An editorial on the importance of standards in additive manufacturing, drawing an analogy to printer drivers, PostScript and PDF. It describes the shift from STL to new universal file formats and from proprietary material specifications to ASTM, ISO and NIST standards, arguing that most users need performance standards.
Inaugural editorial introducing the journal 3D Printing and Additive Manufacturing. It outlines the journal's scope, from processes, materials and design tools to medicine, education, bioprinting, food, art, policy, standards and business models, and notes features such as Chinese translation of articles.
Reviews how desktop manufacturing such as 3D printers, die cutters and 3D scanners can advance children's engineering education. It discusses links between virtual designs and physical objects to mathematics concepts, CAD tools that capture design thinking, motor skills, and sharing designs across schools, and notes that classroom best practices are not yet defined.
Measures the stress response of granular packings of 3D-printed particles with many shapes, including spheres, Platonic solids and non-convex hexapods, dolos and tetrahedral frames. Confining pressure dominates the response, but particle shape controls stress-strain details and can tune packing stiffness and yielding, revealing trends useful for designing aggregate behavior.
2013, IEEE Congress on Evolutionary Computation (CEC)
Describes how to let users evolve any uploaded 3D shape with CPPN-based interactive evolution by injecting the distance to the object's surface as a CPPN input. The seeded CPPN reproduces the original shape at arbitrary resolution while allowing it to be morphed in complex ways, and could extend to seeding neural wiring diagrams.
Describes an open-source C++ 2D and 3D simulation engine for statics, dynamics and nonlinear deformation of soft bodies made of interspersed materials with differing properties. It handles collisions and volumetric actuation, and runs objects with thousands of degrees of freedom at interactive frame rates on a single desktop thread.
Introduces EndlessForms.com, a website for interactively evolving 3D shapes online using a CPPN generative encoding inspired by developmental biology, with WebGL display and an option to 3D print evolved designs. In its first year, nearly 3 million objects were evaluated over 190,000 generations by over 40,000 visitors from 150 countries.
A magazine essay arguing that designers given unlimited 3D printing capability still produce plain objects, constrained by years of mass-produced forms and conventional CAD tools. It argues that new design tools are needed to exploit the freedom additive manufacturing provides, starting from an anecdote about students' pencil holders.
Presents a data-driven process for estimating nonlinear material models for multi-material 3D printed materials. Rather than classical engineering stress-strain assuming constant cross-section, it runs virtual tensile tests by finite element analysis on the full 3D specimen geometry and fits material model parameters by minimizing the discrepancy with experimental results. Preliminary results are reported.
Describes Fab@Home Model 3, an open hardware personal fabrication platform designed to lower skill and price barriers. It is modular in tool heads, fabrication processes and electronics, costs under $1000 and needs only simple tools to assemble. The aim is to move personal 3D printers beyond plastic FDM toward general-purpose personal manufacturing.
Presents freeLoader, a low-cost, desktop-size, open source universal testing machine for high-throughput material testing. It applies tensile and compressive loads up to 5 kN at 2 to 30 mm/min with plus or minus 1.8 N force accuracy, and costs an order of magnitude less per unit than commercial systems. Design, parts list, software and manual are freely available under a BSD license.
Explores microbricks, interlocking 500 micrometer building blocks for assembling and reconfiguring 3D structures on a lattice. A space-filling, rotation and flip-invariant geometry was designed and fabricated by micromachining silicon and by building SU-8 polymer tiles in three layers. Ten-brick structures demonstrated interlocking and compatibility between the two materials.
Lipson's essay in the National Academy of Engineering's The Bridge (Frontiers of Engineering issue) on where additive manufacturing is heading, from printing passive shapes to printing multi-material, functional and eventually active systems. The linked PDF is the full journal issue, with the essay near the start.
Describes a 3.89 g flapping-wing hovering ornithopter built with 3D-printed wings and mechanical parts. Printing allows arbitrary wing shapes and reduces wing design cycles to minutes. The insect achieved passively stable untethered hovering flight lasting 85 seconds, showing the utility of printed materials for flapping-wing experimentation.
Demonstrates in situ additive manufacturing repair of chondral and osteochondral defects in a calf femur, using geometric feedback and suitable printing materials. Alginate hydrogel repaired cartilage and a demineralized bone matrix formulation repaired bone. Mean surface errors were below 0.1 mm, with 42.8% of chondral and 83.6% of osteochondral surface points within clinically acceptable error.
A later listing of the Nature paper "Automatic design and manufacture of robotic lifeforms" (Lipson and Pollack, 2000): robot bodies and controllers evolved together in simulation and then fabricated automatically by 3D printing, producing physical machines that locomote.
Proposes the Extensible Digital Fabrication Language (XDFL) and an interpreted ToolScript language describing how geometry is translated into machine commands, since AMF and STL describe objects but fabricator control is machine-specific. An initial implementation was deployed in the Fab@Home project. A standard process control language would decouple process planning from equipment makers.
Examines producing cellular materials by freeform fabrication when microstructure features are finer than the process resolution. It studies stochastic foams made through the instability of a viscous thread and several methods for closed-cell foams, without pre-specified cell structures. Foams made from elastic materials may act as novel actuating materials.
Proposes greedy geometric feedback (GGF) control for discrete-deposition freeform fabrication, closing the loop on overall part geometry rather than process variables. The geometry is monitored during printing and a greedy algorithm decides where subsequent droplets go. In simulation and on a physical platform, GGF beat open-loop control under uncertainties in droplet shape, placement and mid-print deformation.
Develops printable food materials compatible with conventional post-processing such as baking, broiling and frying. Turkey, scallop and celery were modified with transglutaminase for slow cooking or deep frying, and multi-material turkey and celery constructs were printed. A modified cookie recipe held shape in baking, and a printed cocoa-modified inner shape within a block preserved its complex internal geometry.
Explores material properties achievable with voxel-based freeform fabrication and simulates how they can be tuned. Varying the precision, geometry and material of voxels gives continuous control over density, elastic modulus, thermal expansion, ductility and failure mode. Hierarchical voxel microstructures yield properties such as negative Poisson's ratio, so widely varying properties are possible in one desktop process.
Describes experiments printing variable-stiffness hydrogels for anatomical models and cell-seeded implants, alternating 2% alginate with a dextran-infused calcium chloride post-crosslinker. Stiffness across the construct ranged from 4 kPa to 20 kPa depending on user-specified crosslinker concentration.
Proposes modular printed bricks that integrate building systems, using a material platform of clay, gypsum cement, FabEpoxy and conductive silicone on an open-architecture freeform fabrication system. A cement wall segment with embedded electrical and fluidic conduits was fabricated. Electrical and fluidic tolerances were comparable to traditionally constructed systems.
Studies physical voxels as material building blocks for a massively parallel additive process in which 3D matter is digital. Drawing on space-filling shape geometry, numerical simulation and a technology survey, it presents self-aligning voxel geometries suited to parallel assembly. Theory and simulation predict dimensional accuracy improves as voxel count grows, and current technology could assemble billions of microscale voxels.
Presents a freeform fabricator with a six degree-of-freedom print head for additive deposition onto existing structures, useful for repair, augmentation and in situ printing. A parallel mechanism with stationary motors was developed and optimized through kinematic simulation. A working printer was built, an additive 3D structure was demonstrated, and its accuracy and reliability were analyzed.
Presents an evolutionary design algorithm for 3D multi-material structures that exploits the freedom of internal material distribution offered by multi-material printing. Material distribution is optimized under functional constraints such as maximizing stiffness per weight, and the method can also optimize shape. It is demonstrated on freeform shapes, non-uniform 3D structures and 3D compliant actuators.
Describes Fab@Home Model 2, an update to the open-source, open-architecture personal fabrication platform. A new electronics package and streamlined mechanics cut the system cost from $2500 to $1600. The authors also propose a changed business model intended to broaden adoption of freeform fabrication and spur innovation.
Demonstrates a freeform fabrication system that prints with reusable physical voxels, so multi-material objects can be recycled with minimal effort. Any number of materials can be printed together, and the voxels are reclaimed by dissolving the bonds holding the structure together. With a voxel sorting process, multiple generations of objects were fabricated from the same physical material.
Investigates hydrocolloids as printable materials for food freeform fabrication. Combinations of xanthan gum and gelatin plus flavor agents let texture and flavor be tuned independently, simulating a range of foods with few materials. A two-dimensional mouthfeel scale (stiffness versus granularity) was created, and mixtures were characterized by an expert taste panel.
2009, Genetic and Evolutionary Computation Conference (GECCO)
Presents an evolutionary design process for 3D multi-material structures, as an alternative to conventional topology optimization methods like homogenization that struggle in the large design space opened by multi-material printing. The algorithm designs substructures tailored to custom functions and is demonstrated on 3D non-uniform beams and 3D compliant actuators.
Uses only solid freeform fabrication processes to reproduce the 1844 Vail telegraph register, showing that a complete active electromechanical system can be printed. Techniques developed allow solenoid devices to be fabricated and embedded in freeform fabricated devices, which could enable electromechanical designs that are otherwise hard to manufacture.
Proposes an XML-based Additive Manufacturing File (AMF) format as a successor to STL, handling multiple and graded materials, volumetric patterns and colors. Geometry can be given as triangle meshes, functional representations or voxel bitmaps, with materials defined singly or as hierarchical combinations. The format can be converted to and from STL, and code and examples are public.
2009, ACM Transactions on Computer-Human Interaction
ModelCraft lets designers annotate and edit physical 3D models, such as 3D-printed or laser-cut prototypes, with a digital pen. The model's surface carries a traceable pattern, so sketches, comments and edit commands drawn on it are captured and transferred back into the digital CAD model.
Describes 3Dprintables.org, an open wiki-style archive of printable teaching models, built on the argument that content drives 3D printing adoption as it did for digital music. Initial models span mechanical engineering, aerospace, biochemistry, mathematics, anatomy and archaeology, and are meant to seed educators sharing models and curricula.
Develops battery materials compatible with solid freeform fabrication and uses them to print six cylindrical Zn-air cells plus a flexible two-cell battery of unusual geometry, comparing them with commercial cells. The printed cells reached 40-70 J/g, under 10% of normalized commercial performance, with contact resistance, electrolyte evaporation and catalyst quality as likely causes.
Studies how mixing of alginate hydrogel affects printing of cell-seeded tissue constructs, since local variation in material properties degrades deposition and geometric fidelity. Using a mechanical test frame with a syringe-extrusion jig, the authors found that mixing far beyond previously established limits reduced material inconsistency and improved geometric fidelity without harming cell viability.
Shows that freeform fabrication systems can use many materials and processes within a single object and can make complete active devices. The authors' multi-material system produced Zn-air batteries, wiring, flexible circuits, strain gauges, electromagnets, electroactive polymer actuators, organic transistors and relays. They also released the open-source Fab@Home Model 1, with over 100 built worldwide.
Explores using a freeform fabricator to augment everyday objects for visually impaired users, by printing a tactile alphabet on paper, plastic and metal. The authors tested non-toxic materials for dimensional tolerance, repeatability and adhesion; printing 1 by 1 cm embossed letters took 14 to 52 seconds.
Demonstrates solid freeform fabrication of metal by depositing a 17-4 PH stainless steel slurry layer by layer with the low-cost Fab@Home, then sintering. Completed parts reached densities up to 90% volume fraction and tensile strength up to 35% of the pure solid, with samples of various geometries shown.
Proposes Universal Desktop Fabrication, a framework combining volumetric digital design software with compact, automated multi-material fabrication hardware to produce almost any complete functional object. The aim is inexpensive, open, easy-to-use systems that shift manufacturing toward personal customization and point-of-use production, with current hardware and software results described.
Presents Fab@Home, a personal desktop solid freeform fabrication kit meant to get the technology to hobbyists and hackers, breaking the paradox of low demand and high cost. Designs and software are freely distributed, a first-generation kit should cost roughly $2000, and a user-group website is planned. The paper also reports experience deploying a unit for public use.
Works toward freeform fabrication of electrochemical relays built around an IPMC actuator whose electrodes open or close a load circuit, with the casting mold serving as housing. Partially printed relays, with housings made on a Fab@Home Model 1, achieved a factor-of-two change in load path resistance at up to 0.05 Hz with a 1.2 V control input.
2007, Cambridge University Engineering Department Publications Database
Reports freeform fabrication of organic electrochemical transistors, chosen over organic field effect transistors for their sub-1 V operation and ambient-condition fabrication. Functional devices made with a multi-material system on glass showed electrical behavior similar to the literature. The authors identified printable substrates and added inkjet deposition to improve thin films.
Explores digital printing of multi-material objects from many discrete self-aligning voxels, so accuracy is set by the voxels rather than the printer. To place millions of voxels quickly, the authors propose using electrostatic forces to place a whole 2D pattern at once, and demonstrate selective pick-up of 1.5 mm voxels with a custom charged print head.
Shows that a Fab@Home Model 1 personal fabricator with commercially available materials can print functional embedded electrical circuits. Demonstrations include an LED flashlight, entirely printed circuit boards in 2D and 3D, and a child's toy with embedded circuitry, aimed at making such capability available to individual creators.
Examines food as a raw material for freeform fabrication, noting uses in custom confections, as an accessible medium for teaching rapid prototyping, and as sacrificial or support material. Using Fab@Home, the authors produced multi-material edible 3D objects from cake frosting, chocolate, processed cheese and peanut butter.
Combines injection-molding style tissue engineering with solid freeform fabrication to print pre-seeded alginate hydrogel implants of arbitrary geometry without custom tooling. The process was sterile with 94 plus or minus 5% cell viability, matrix production comparable to other fabrication methods, and z-direction RMS surface roughness of 0.16 plus or minus 0.02 mm.
Argues that solid freeform fabrication, traditionally used for passive single-material parts, can be extended to a compact system with a small set of materials that produces complete active electromechanical devices. The authors discuss the long-term vision of compact, flexible factories supporting space exploration by reducing the need to launch finished equipment.
2006, ACM Symposium on User Interface Software and Technology (UIST)
Presents ModelCraft, a system that captures annotations and edits made on physical 3D models by tracking a pattern printed on the model surface, needing no tracking infrastructure or per-object calibration. Annotations and commands are merged back into the CAD model; the paper describes a SolidWorks plug-in and initial user feedback.
Focuses on freeform fabrication of complete zinc-air batteries with tailored geometry and output. A gelling agent keeps the zinc anode and catalyst layers from separating, allowing nozzles down to 1.3 mm, and polyvinyl alcohol replaces paper as the separator. Batteries reached 60 mAh/g of zinc, 7.25 mW average power and 63 h service on a 100 ohm load, about an order of magnitude below commercial cells.
IEEE Spectrum article on fabbers, machines that make useful items on demand from digital designs and could, like personal computers, become available to ordinary people. It draws on Neil Gershenfeld's vision of personal fabrication and notes that researchers at several universities are investigating technologies toward general-purpose fabricators.
Reports using rapid prototyping to document, preserve, reproduce and share historic machines in three dimensions. Several preassembled, functional mechanisms from the Cornell Collection of Reuleaux Kinematic Models were reproduced and placed in an online museum of mechanism, where visitors can download and 3D-print working replicas. The authors expect such physical preservation to become common.
Develops processes, materials and tooling to directly 3D print living, pre-seeded, patient-specific implants with spatially heterogeneous composition, by layer-wise deposition of cell-seeded alginate hydrogel. A robotic deposition system and CAD software were built. Printed gels showed 94 plus or minus 5 percent viability (n=15) and sterility, and a meniscus-shaped gel was printed from a CT scan.
Works toward producing complete electromechanical systems by freeform fabrication, building a library of compatible elements. Conducting polymers can supply actuation, sensing and control electronics. The authors manually produced air-operable actuators compatible with their power source and mechanical library, present a survey of candidate actuator materials, test two types, and demonstrate complete freeform-producible actuator devices.
Reports a fabrication platform and extensions to deposition processes for freeform fabrication of 3D functional assemblies with embedded conductive wiring and power sources. Structures use thermoplastics and elastomers, wiring uses low-melting alloys and conductive pastes, and zinc-air batteries are deposited with separators and electrodes. Process details and several printed assemblies are demonstrated.
Argues that solid freeform fabrication, traditionally for passive parts, can be developed into a compact robotic system that prints complete active electromechanical devices. Successes include printed zinc-air batteries, wiring and flexure joints; actuation, sensing and control remain, and conducting polymers are proposed. Coupled with evolutionary design, this would give evolution a rich physical substrate.
Reports a fabrication platform and deposition-process extensions for freeform fabrication of 3D functional assemblies with embedded wiring and power sources. Thermoplastics and elastomers form structure, low-melting alloys and conductive pastes form wiring, and zinc-air batteries are printed from zinc, electrolyte and catalysts. Process details and several printed assemblies are shown.
Presents the GOLEM project, which extends evolution into the physical world by evolving electromechanical machines made of thermoplastic, linear actuators and neurons for locomotion, in simulation and then reality. Using 3D solid printing, the evolved creatures replicate into reality and reproduce the performance of their virtual ancestors.
2002, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Proposes continuously reconfigurable robotics: given a task, a robot evolves a morphology and controller, prints the assembled structure, performs the task, and is then recycled into a new form. Demonstrated for locomotion with linear actuators and sigmoidal neurons in a thermoplastic body, printed pre-assembled by rapid prototyping, with examples of working physical robots.
Lipson and Pollack report the GOLEM project: robots whose bodies and controllers were designed together by an evolutionary process in simulation and then built automatically by 3D printing, with only motors added by hand. The resulting machines locomoted in the physical world, a step toward machines that control their own design and manufacture.
Describes the GOLEM project, which evolves electromechanical robots built from thermoplastic, linear actuators and neurons for locomotion, first in simulation and then in reality. Using 3D solid printing, the creatures replicate into hardware and reproduce the performance of their virtual ancestors, transferring evolved morphological diversity into reality for the first time.
Freeform fabrication of complete functional devices rather than passive parts: printed batteries, actuators, circuits, transistors and relays, embedded components, and programmable or reconfigurable materials.
Explores a rewritable 1D metamaterial of non-circular multistable corrugated tubes that can be bent into specified 3D shapes, inspired by how amino-acid sequences define protein shapes. Cross-section shape can be rewritten multiple times, allowing repair and reprogramming of multistability, and a numerical model describes the bending energy landscape. Suggested uses include robotics, deployable structures and catheters.
Proposes embedding electronic components during selective laser sintering using an inverted laser sintering process that needs no powder bed, so voids can be created easily. An LED is placed into a cavity within an SLS part and printing resumes. The LED still works after several layers are added, pointing toward SLS producing active systems rather than passive parts.
Demonstrates parallel electro-osmotic grippers for layered assembly, where each array element is controlled individually via capillary effects driven by an electric field to form droplet patterns that pick and place voxels. A 25-element array assembled 1.5 mm spheres with 95 to 98% grasping reliability, validated against analytical and computational models.
Examines electrically conductive fabrics as heaters for silicone/ethanol soft actuators, where ethanol undergoes a thermal phase change. A stretchable kirigami-based fabric design was developed that cycles the actuator to 80 C, withstands up to 195 C and 30 W, and lets the actuator exceed 30% linear strain.
2018, IEEE International Conference on Soft Robotics (RoboSoft)
Improves a thermally responsive silicone actuator, in which embedded solvent droplets undergo a liquid-gas transition, by replacing the nickel-chromium heater wire with a silicone-based conductive composite. The composite heater performed as well as the wire with better material compatibility. A 20 g actuator lifted its 4.3 kg DC power source.
Silicone-ethanol soft composite actuators degrade as ethanol escapes during prolonged actuation. The authors restore them by immersing the material in ethanol until it is saturated again; repeated rejuvenation of a soft robot built from the actuator retained up to 100% of its functionality, suggesting a way to extend the life of soft artificial muscles.
Presents a single self-contained soft composite combining an elastic polymer matrix with the large volume change of a liquid-vapor transition. It reaches strain up to 900% and stress up to 1.3 MPa at a density of 0.84 g/cm3, and is cheap and simple to make. The authors propose it for electrically driven, entirely soft robots.
Introduces electro-active hydraulic solids, a conductive elastomer filled with many microscale phase-change capsules that act like wax actuators working in parallel. A 14 mm by 50 mm cylinder is reported to produce 4,500 N of force with over 3% strain. The authors give a simple fabrication method and show compatibility with additive, subtractive and forming processes.
A moderated roundtable with Jennifer Lewis, Robert Shepherd and Hod Lipson on the prospects and limits of 3D printing soft materials. Participants describe their use of printing for robots and functional materials, from hard plastics to soft and active materials, conductors, batteries, actuators, biomaterials and food.
Develops a thermally activated actuator material made by mixing liquid silicone with liquid paraffin wax and cooling it into a suspension of wax particles. The robocast material expands by up to 6% in volume when heated above the wax melting temperature, offering a printable soft actuator.
2014, The International Journal of Robotics Research
Introduces Bitblox, small modular interconnecting blocks embedding simple electromechanical connectivity and functionality as printable digital materials. Different blocks, and their combination and positions, set the mechanical and electrical properties of an assembly. The paper describes the design, compares it with similar materials and demonstrates use in a working 3D printer.
Proposes modular printed bricks that integrate building systems, using a material platform of clay, gypsum cement, FabEpoxy and conductive silicone on an open-architecture freeform fabrication system. A cement wall segment with embedded electrical and fluidic conduits was fabricated. Electrical and fluidic tolerances were comparable to traditionally constructed systems.
Uses only solid freeform fabrication processes to reproduce the 1844 Vail telegraph register, showing that a complete active electromechanical system can be printed. Techniques developed allow solenoid devices to be fabricated and embedded in freeform fabricated devices, which could enable electromechanical designs that are otherwise hard to manufacture.
Reports two advances toward programmable matter, in which components assemble in fluid with dynamically switchable affinities to form irregular structures. One is a multi-chamber microfluidic chip for improved far-field assembly; the other is a demonstration of near-field inter-tile affinity switching using a thermorheological assembly fluid.
Develops battery materials compatible with solid freeform fabrication and uses them to print six cylindrical Zn-air cells plus a flexible two-cell battery of unusual geometry, comparing them with commercial cells. The printed cells reached 40-70 J/g, under 10% of normalized commercial performance, with contact resistance, electrolyte evaporation and catalyst quality as likely causes.
2008, IEEE International Conference on Micro Electro Mechanical Systems (MEMS)
Presents electrical and mechanical interfaces for fluidically assembled planar MEMS, including passive mechanical latches that bond microcomponents and electrical layers that connect them. The authors evaluate bonding within a microfluidic channel and circuit formation after assembly, supporting a strategy bridging bottom-up self-assembly and top-down manipulation.
Shows that freeform fabrication systems can use many materials and processes within a single object and can make complete active devices. The authors' multi-material system produced Zn-air batteries, wiring, flexible circuits, strain gauges, electromagnets, electroactive polymer actuators, organic transistors and relays. They also released the open-source Fab@Home Model 1, with over 100 built worldwide.
Describes directed hierarchical self assembly, a two-stage procedure with coarse far-field fluidic transport of 500 micrometer silicon microtiles followed by near-field assembly controlled by on-tile jets and thermally actuated gel microvalves. Preliminary work covers CFD simulation of fluid forces and a triblock copolymer gel valving system.
Works toward freeform fabrication of electrochemical relays built around an IPMC actuator whose electrodes open or close a load circuit, with the casting mold serving as housing. Partially printed relays, with housings made on a Fab@Home Model 1, achieved a factor-of-two change in load path resistance at up to 0.05 Hz with a 1.2 V control input.
2007, Cambridge University Engineering Department Publications Database
Reports freeform fabrication of organic electrochemical transistors, chosen over organic field effect transistors for their sub-1 V operation and ambient-condition fabrication. Functional devices made with a multi-material system on glass showed electrical behavior similar to the literature. The authors identified printable substrates and added inkjet deposition to improve thin films.
Shows that a Fab@Home Model 1 personal fabricator with commercially available materials can print functional embedded electrical circuits. Demonstrations include an LED flashlight, entirely printed circuit boards in 2D and 3D, and a child's toy with embedded circuitry, aimed at making such capability available to individual creators.
Presents a microfabrication approach in which silicon microtiles self-assemble in a multilayer PDMS microchamber into arbitrary, reconfigurable structures. Controlling local flow in a microchannel array, the experiments demonstrate attraction, bonding, migration and rejection of single and paired tiles, and compare tile, outlet and medium designs. The goal is programmable digital matter.
Argues that solid freeform fabrication, traditionally used for passive single-material parts, can be extended to a compact system with a small set of materials that produces complete active electromechanical devices. The authors discuss the long-term vision of compact, flexible factories supporting space exploration by reducing the need to launch finished equipment.
Develops materials and processes to freeform fabricate complete ionomeric polymer-metal composite actuators and their fabrication substrate with a multi-material system. The printed actuators operated in air for over 40 minutes and 256 bidirectional cycles. Output stress per input power was two orders of magnitude below the best conventional devices; explanations and improvements are discussed.
Focuses on freeform fabrication of complete zinc-air batteries with tailored geometry and output. A gelling agent keeps the zinc anode and catalyst layers from separating, allowing nozzles down to 1.3 mm, and polyvinyl alcohol replaces paper as the separator. Batteries reached 60 mAh/g of zinc, 7.25 mW average power and 63 h service on a 100 ohm load, about an order of magnitude below commercial cells.
Works toward producing complete electromechanical systems by freeform fabrication, building a library of compatible elements. Conducting polymers can supply actuation, sensing and control electronics. The authors manually produced air-operable actuators compatible with their power source and mechanical library, present a survey of candidate actuator materials, test two types, and demonstrate complete freeform-producible actuator devices.
Reports a fabrication platform and extensions to deposition processes for freeform fabrication of 3D functional assemblies with embedded conductive wiring and power sources. Structures use thermoplastics and elastomers, wiring uses low-melting alloys and conductive pastes, and zinc-air batteries are deposited with separators and electrodes. Process details and several printed assemblies are demonstrated.
Argues that solid freeform fabrication, traditionally for passive parts, can be developed into a compact robotic system that prints complete active electromechanical devices. Successes include printed zinc-air batteries, wiring and flexure joints; actuation, sensing and control remain, and conducting polymers are proposed. Coupled with evolutionary design, this would give evolution a rich physical substrate.
Reports a fabrication platform and deposition-process extensions for freeform fabrication of 3D functional assemblies with embedded wiring and power sources. Thermoplastics and elastomers form structure, low-melting alloys and conductive pastes form wiring, and zinc-air batteries are printed from zinc, electrolyte and catalysts. Process details and several printed assemblies are shown.
Printing living cells, hydrogels and anatomically shaped tissue constructs, and other medical uses of fabrication and AI such as implants, tissue simulators, cardiology and portable ultrasound.
Makes the case for a portable open-source 3D ultrasound system with AI, aimed at overcoming cost, training and access barriers, particularly for maternal care in developing regions. It weighs arguments for and against an open-source approach, drawing on earlier open-source hardware projects.
A review of how 3D printed patient-specific models, built from advanced imaging such as cardiac MRI or CT, are used for preprocedural planning in patients with heart failure of varied causes. It surveys reported applications and their potential to improve patient care.
Uses multimaterial solid freeform fabrication to build soft tissue simulators that replicate the injection force profile of a needle stick, using blunted needles that are safer for training. Composite structures and needle pairs gave a puncture force of 2.8 N at 9 to 15 mm depth, within the reported range for human skin.
Presents a simultaneous 3D-printing and photocrosslinking technique for rapidly making anatomically accurate, mechanically heterogeneous aortic valve scaffolds from PEG-DA hydrogels with alginate. Blends spanned a more than 10-fold elastic modulus range, valves printed in 14 to 45 minutes, and seeded porcine valve interstitial cells stayed near 100% viable over 21 days.
Demonstrates in situ additive manufacturing repair of chondral and osteochondral defects in a calf femur, using geometric feedback and suitable printing materials. Alginate hydrogel repaired cartilage and a demineralized bone matrix formulation repaired bone. Mean surface errors were below 0.1 mm, with 42.8% of chondral and 83.6% of osteochondral surface points within clinically acceptable error.
Presents a technique for characterizing homogeneity of hydrogel printing inks by measuring loads during deposition and their temporal variation ("mechanical noise"). Mixing alginate and crosslinker for 128 cycles cut mechanical noise by 82% and improved geometric fidelity. Viability was not harmed and rose 34% at a 45 minute curing time, and modulus also increased with more mixing.
Describes experiments printing variable-stiffness hydrogels for anatomical models and cell-seeded implants, alternating 2% alginate with a dextran-infused calcium chloride post-crosslinker. Stiffness across the construct ranged from 4 kPa to 20 kPa depending on user-specified crosslinker concentration.
Describes an optical, laser-triangulation method for quantifying how faithfully anatomically shaped meniscal constructs reproduce native tissue geometry. Ovine menisci were imaged by MRI and microCT, and molds and printed alginate constructs were compared with silastic molds cast directly from tissue. Silastic molds were within 10% error for all seven key measurements, while printed constructs matched fewer measurements.
Studies how mixing of alginate hydrogel affects printing of cell-seeded tissue constructs, since local variation in material properties degrades deposition and geometric fidelity. Using a mechanical test frame with a syringe-extrusion jig, the authors found that mixing far beyond previously established limits reduced material inconsistency and improved geometric fidelity without harming cell viability.
Combines injection-molding style tissue engineering with solid freeform fabrication to print pre-seeded alginate hydrogel implants of arbitrary geometry without custom tooling. The process was sterile with 94 plus or minus 5% cell viability, matrix production comparable to other fabrication methods, and z-direction RMS surface roughness of 0.16 plus or minus 0.02 mm.
Develops processes, materials and tooling to directly 3D print living, pre-seeded, patient-specific implants with spatially heterogeneous composition, by layer-wise deposition of cell-seeded alginate hydrogel. A robotic deposition system and CAD software were built. Printed gels showed 94 plus or minus 5 percent viability (n=15) and sterility, and a meniscus-shaped gel was printed from a CT scan.
Digital Food: Food Printing and Software-Controlled Cooking
3D-printed food and laser-based cooking, from multi-material food printing to precision browning, broiling and texturizing with lasers, and the case for software-controlled cooking.
Investigates blue (445 nm), near-infrared (980 nm) and mid-infrared (10.6 um) lasers for texturizing 3D-printed foods in situ, addressing the difficulty of reproducing crunchy or firm textures. Modulating laser exposure frequency across printed layers controls elasticity and chewiness, with more frequent exposure giving firmer textures, as validated by compression testing.
Presents a physics-based simulation of the deposition process in multi-ingredient food 3D printing, taking G-code as input and using the Bifrost plug-in for Autodesk Maya. It is validated by reproducing a custom seven-ingredient dessert. Simulating designs before printing helps build intuition for stable structures, reduces waste and speeds proofing.
A Viewpoint conversation between Lipson and Salah Sukkarieh on how robotics could shape food production and preparation. It covers sensor-rich agricultural robots delivering precision farming and 3D food printers making customized dishes, with the aim of producing more, healthier and more sustainable food using fewer resources and a lower carbon footprint.
A perspective arguing that laser cooking and 3D food printing may complement or replace analog cooking methods such as grills, stoves and microwaves. Software control of ingredient combination and heating could give finer control over nutrient content, enable customized meals, and reduce human handling and foodborne-illness risk. The authors suggest a 3D printer may become a staple home and industrial cooking device.
Argues that multi-ingredient food printing combined with laser cooking could offer convenient, nutritious, low-cost meals, a combination not previously demonstrated. It presents digitally crafted meals as an example of joining software with food and frames software and robotics as the next step in individualized meal production.
Explores precision laser cooking of printed chicken by combining blue (445 nm), near-infrared (980 nm) and mid-infrared (10.6 micrometer) lasers. IR light browns more efficiently than blue, NIR can brown and cook through packaging, laser-cooked food lost about 50% less in cooking than oven-broiled food, and cooking resolution is estimated at about 1 mm.
Explores laser cooking as a software-controlled way to cook printed food layers, using mirror galvanometers to direct a 445 nm blue laser and a 980 nm near-infrared laser. By varying power, speed and pattern, the authors produced a software-generated three-course meal, and present digital cooking as a new tool for chefs and makers.
Studies selective laser broiling of Atlantic salmon with a galvanometer-steered blue laser using trochoidal scan patterns, varying pattern geometry, heat flux and power. Blue laser thermal penetration approached 2 mm, and food-safe temperatures above 62.8 C with desirable color and texture change were reached using high speeds and repeated exposure.
Characterizes browning of flour-and-water dough with a 10.6 micrometer CO2 laser, varying power, beam diameter and exposure time. At 0.32 MW/m2 with a 5.7 mm beam and 180 s exposure, satisfactory browning was seen with a maximum thermal penetration of 0.77 mm. The results suggest suitability for thin goods and food layered manufacture.
Studies cooking dough with a 445 nm blue diode laser, adjusting dough water content and exposure pattern. A 1 mm thick sample was cooked with a ring pattern at 2 W. Blue-laser heat penetration is significantly higher than with an infrared laser, and combining blue and infrared lasers gave the best conditions for food layered manufacture.
Illustration
Multi-Material Three-Dimensional Food Printing with Simultaneous Infrared Cooking DOI
2018, 3D Printing and Additive Manufacturing · no PDF yet
A study of multi-material 3D food printing combined with infrared heating that cooks the food as it is printed.
No PDF yet: the file found earlier was a different paper, "Characterization of CO2 laser browning of dough"; it has been removed.
Proposes a supervised deep generative network that predicts photorealistic RGB images of laser-browned dough from laser speed, energy flux and dough moisture. Validated on a CO2 laser browning pipeline, the model characterizes surface color and pattern, and is intended as a step toward CAD software for food processing.
Explores additive manufacturing for food-safe applications by producing an entirely additively manufactured dinner, including complex tableware. It notes that current AM lacks the process control and certification for widespread food contact use, with coatings mitigating concerns for prototypes and small runs, and lays out barriers to adoption.
Demonstrates additive manufacturing for the hospitality industry by producing an entirely additively manufactured dinner with complex tableware for limited audiences. It observes that current techniques lack the process control and certification needed for widespread food contact use, that coatings can help, and lists barriers for further work.
Develops printable food materials compatible with conventional post-processing such as baking, broiling and frying. Turkey, scallop and celery were modified with transglutaminase for slow cooking or deep frying, and multi-material turkey and celery constructs were printed. A modified cookie recipe held shape in baking, and a printed cocoa-modified inner shape within a block preserved its complex internal geometry.
Investigates hydrocolloids as printable materials for food freeform fabrication. Combinations of xanthan gum and gelatin plus flavor agents let texture and flavor be tuned independently, simulating a range of foods with few materials. A two-dimensional mouthfeel scale (stiffness versus granularity) was created, and mixtures were characterized by an expert taste panel.
Examines food as a raw material for freeform fabrication, noting uses in custom confections, as an accessible medium for teaching rapid prototyping, and as sacrificial or support material. Using Fab@Home, the authors produced multi-material edible 3D objects from cake frosting, chocolate, processed cheese and peanut butter.
Soft and compliant robots and the materials that drive them: jamming-based universal grippers, soft actuators, elastic-instability swimmers and crawlers, electroadhesive grippers and simulation of soft bodies.
Frames shape prediction for tendon-driven continuum robots as action-conditioned kinematic self-modeling. Using a lightweight 3D-printed platform and RGB-D data collection, a point cloud flow matching model maps motor actuation to settled 3D geometry. It improves shape accuracy over prior approaches on simulated 2-, 3- and 5-module robots and real 2- and 3-module robots, and extends to tip payload conditioning in simulation.
Analyzes the static and dynamic properties of the hair clip mechanism, a kinked ribbon with ends pinned like a snap hair clip, validates the theory experimentally, and applies it in two carangiform fish robots. A tethered pneumatic fish reaches 1.40 body lengths per second, about twice a conventional counterpart, and an untethered motor-driven one reaches 2.03 BL/s (42.6 cm/s) with 3.89% efficiency at 3 Hz.
Presents JamTet, a physics-based sphere-packing framework for tetrahedral meshing and morphing in differentiable simulation. It combines a GPU-parallel mesher, Lagrangian morphing that keeps interior-node identities and avoids inversion, and a differentiable JAX simulator. In soft-robot swimming design experiments, interior-node gradients improve fitness over surface-only variants, and voxelized versions get 32-63% lower fitness.
Introduces a self-deploying formwork for concrete built from multistable tubes that expand from a compact state into a full-scale mold with integrated steel reinforcement. A steel-reinforced structure 2.36 m high is deployed within 14 seconds, and the polymer shell of the tubes triples structural strength. The concept aims at low-carbon construction with minimal on-site intervention.
Explores a rewritable 1D metamaterial of non-circular multistable corrugated tubes that can be bent into specified 3D shapes, inspired by how amino-acid sequences define protein shapes. Cross-section shape can be rewritten multiple times, allowing repair and reprogramming of multistability, and a numerical model describes the bending energy landscape. Suggested uses include robotics, deployable structures and catheters.
2024, International Conference on Reconfigurable Mechanisms and Robots (ReMAR)
Introduces a particle robot of vibrating units, each with an off-axis spinning mass, enclosed by a rigid chain-link boundary that behaves like a soft robot. Individual particles cannot move directionally, but together they yield deterministic motion from stochastic interactions. Experiments show how particle count affects movement and direction, and the robot can pass tight spaces and flow around obstacles.
Studies the swimming pattern produced when a hair clip mechanism, an in-plane prestressed bistable structure that snaps through, serves as a fish body. With the same energy consumption, the HCM fish cruises about twice as fast as a conventional soft fish. In single-link simulation it gives 16.7 N/m average propulsion, 2-3 times the reference and sine patterns, at efficiency close to the sine pattern.
Describes CarbonFish, a bistable, underactuated fish robot built from a carbon-fiber-reinforced plastic ribbon using the hair clip mechanism design paradigm. Preliminary tests of the single-actuated prototype show an undulation frequency approaching 10 Hz, suggesting potential for high-speed, biomimetic swimming.
Develops the theory and a design algorithm for a hair clip mechanism propulsion system for soft fish robots, built from PETG plastic, carbon fiber-reinforced plastic and steel. It derives and verifies the HCM theory and summarizes how dimensions, materials and servo specifications matter. The algorithm helps select components and operate at a desired frequency for high-speed undulatory swimming.
2023, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Describes a single-actuated untethered compliant crawler that uses a prestressed hair clip mechanism as chassis, spine and force amplifier. It gallops at 313 mm/s (1.56 body lengths per second). Experiments find speed is linearly related to actuation frequency and substrate friction, tethering slows the crawler, and asymmetric actuation produces a new galloping gait.
2023, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Presents a human-finger-inspired snapping gripper that uses elastic instability in a prestressed semi-rigid skeleton to close rapidly over a wide span. Fingers open to 86 mm and close in 46 ms, 2.7 times better in span and 10.9 times better in speed than a reference gripper. Design principles, simulation and tests on rigid, flexible and limp objects are reported.
Presents Cronos, an open-source 3D simulation engine that parallelizes a nonlinear mass-spring model on the GPU for soft, rigid and hybrid robots. It reaches over 1 billion elements per second on consumer GPUs and allows mixing materials within one robot. Accuracy is validated against Euler-Bernoulli beam theory, natural frequency predictions and empirical data from a soft structure.
Builds a compliant flapping mechanism from a prestressed bistable 'hair clip' structure made of semi-rigid plastic sheets and installs it on tethered pneumatic and untethered motor-driven robotic fish. Design rules are proposed. The pneumatic fish swims twice as fast as the reference, and the untethered fish reaches 2.03 body lengths per second, reported as a record for untethered compliant swimmers.
2021, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Presents an open-source untethered quadruped soft robot, about 80% soft by volume, with 3D printed gyroid-infill flexible body and legs driven by four servo motors. With gaits tuned in a real-time soft body simulation, the robot ran at 0.9 m/s (2.5 body lengths per second), faster than most untethered legged soft robots published to date.
Presents an optimization strategy, with experimental validation, for two-dimensional electroadhesors that grip millimeter-scale objects. Using 3D electrostatic simulation and weight-lifting tests at 600 V on 1.6 mm thick objects, the optimal pad spacing was 0.5 mm and the optimal pad-area ratio 0.75 for conductive and insulating objects. A step-by-step design method follows.
Addresses slow actuation and limited cycle life of silicone/ethanol soft actuators caused by low thermal conductivity. Adding 8 wt.% diamond nanoparticle filler raised conductivity from 0.190 to 0.212 W/mK and improved speed and cycle count without changing mechanical properties. An LSTM network predicted actuation force in multi-cyclic tests.
2020, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Presents Titan, a CUDA-based C++ library for GPU-accelerated simulation of soft-body and multi-agent robots with an asynchronous computing model, letting optimization and control run on the CPU while the GPU simulates. It reaches up to 300 million primitive updates per second and is shown on soft-body and multi-agent applications.
Examines electrically conductive fabrics as heaters for silicone/ethanol soft actuators, where ethanol undergoes a thermal phase change. A stretchable kirigami-based fabric design was developed that cycles the actuator to 80 C, withstands up to 195 C and 30 W, and lets the actuator exceed 30% linear strain.
Designs 2 x 2 arrays of electroadhesive grippers, with comb and spiral electrodes, to place premanufactured voxels in parallel for Layered Assembly, a multi-material additive process. Using field simulations and tests at 75 to 800 V on polymer and metal voxels, comb-shaped electrodes performed better, with about 100% reliability at 600 to 800 V.
2018, IEEE International Conference on Soft Robotics (RoboSoft)
Improves a thermally responsive silicone actuator, in which embedded solvent droplets undergo a liquid-gas transition, by replacing the nickel-chromium heater wire with a silicone-based conductive composite. The composite heater performed as well as the wire with better material compatibility. A 20 g actuator lifted its 4.3 kg DC power source.
Silicone-ethanol soft composite actuators degrade as ethanol escapes during prolonged actuation. The authors restore them by immersing the material in ethanol until it is saturated again; repeated rejuvenation of a soft robot built from the actuator retained up to 100% of its functionality, suggesting a way to extend the life of soft artificial muscles.
Presents a single self-contained soft composite combining an elastic polymer matrix with the large volume change of a liquid-vapor transition. It reaches strain up to 900% and stress up to 1.3 MPa at a density of 0.84 g/cm3, and is cheap and simple to make. The authors propose it for electrically driven, entirely soft robots.
Describes a robotic hand that uses phase transition of granular material in multiple fingers to achieve dexterity with minimal actuation. With only two actuators it performs six fundamental dexterous manipulations plus precision and power grasps, and is shown on real-world grasping and manipulation tasks.
Describes producing cellular foam structures implicitly by inducing viscous thread instability while extruding material, rather than designing each cell explicitly. The process yields structures finer than the printer's native resolution, with effective elastic modulus and density tunable across two orders of magnitude. Applications cited include bioengineering, robotics and food printing.
Introduces electro-active hydraulic solids, a conductive elastomer filled with many microscale phase-change capsules that act like wax actuators working in parallel. A 14 mm by 50 mm cylinder is reported to produce 4,500 N of force with over 3% strain. The authors give a simple fabrication method and show compatibility with additive, subtractive and forming processes.
Argues that shape-related tactile information is better recovered from shear strain than from normal strain. A contact mechanics analysis shows that sensing shear gives a robust edge detection mechanism. A jamming-based robot gripper used as a tactile sensor confirmed that shear processing gives accurate edge information that is invariant to pressure.
Tests whether morphological computation is helped by suitable geometry, material and growth by evolving geometries and growth processes of soft robots with softer or stiffer materials. For the task studied, evolved softer robots reached better performance with simpler growth processes than stiffer ones, which the authors attribute to better exploitation of morphological computation.
A moderated roundtable with Jennifer Lewis, Robert Shepherd and Hod Lipson on the prospects and limits of 3D printing soft materials. Participants describe their use of printing for robots and functional materials, from hard plastics to soft and active materials, conductors, batteries, actuators, biomaterials and food.
2015, Genetic and Evolutionary Computation Conference (GECCO)
Presents the first soft robot evolved to reach or squeeze through a small aperture, using CPPN-NEAT with a new multi-objective implementation that keeps speciation. More compliant, deformable robots performed better at the task, offered as a proof of concept for matching tasks to suitable morphologies.
Develops a thermally activated actuator material made by mixing liquid silicone with liquid paraffin wax and cooling it into a suspension of wax particles. The robocast material expands by up to 6% in volume when heated above the wax melting temperature, offering a printable soft actuator.
Journal version of the open-source simulation engine for statics, dynamics and nonlinear deformation of heterogeneous soft bodies. It supports interspersed materials with varied stiffness, density, Poisson ratio, thermal expansion and friction, handles collisions, uses volumetric actuation, and simulates thousands of degrees of freedom at interactive rates, aimed at multimaterial design automation.
Introduces electrophysiological soft robots, inspired by cardiac electrophysiology, whose behavior arises from electrical signals flowing through the tissue cells of the soft body, merging control and actuation. The paper describes how they are evolved, shows lifelike behaviors, and explores design choices such as sensing, actuation and central pattern generator placement.
Outlines an affordable K-12 outreach program based on building pneumatically powered soft robots, made feasible by low-cost 3D printers. The program requires no programming and is positioned as a supplement to kit and competition-based robotics, serving as a reference for students, parents and educators and as a model for similar programs.
A moderated state-of-the-field discussion among experts from rheology, evolutionary robotics, biomimetic robots, soft actuators, flexible electronics and soft robotics, convened by the Soft Robotics editor. Participants comment on their fields and on how interdisciplinary collaboration can advance soft robots; Lipson represents evolutionary robotics.
Outlines opportunities in soft robotics, including amorphous nonrigid dynamics, a new design space free of manufacturing constraints, natural mimicry and safer interaction with humans. Challenges include limited simulation and design tools, scarce soft actuation and difficult manufacturing. Proposed remedies include evolutionary design tools and porous and jamming materials.
Measures the stress response of granular packings of 3D-printed particles with many shapes, including spheres, Platonic solids and non-convex hexapods, dolos and tetrahedral frames. Confining pressure dominates the response, but particle shape controls stress-strain details and can tune packing stiffness and yielding, revealing trends useful for designing aggregate behavior.
2013, Genetic and Evolutionary Computation Conference (GECCO)
Tests whether Karl Sims-style evolved morphologies are held back by rigid elements and overly regular encodings, by evolving multi-material soft robots with a CPPN generative encoding selected for locomotion speed. CPPNs evolved faster and more natural-looking robots than a direct encoding, and performance rose as more materials were added.
Presents a passive universal gripper made of granular material in an elastic membrane that conforms to an object, hardens under vacuum to grip, and releases using positive pressure. Positive pressure improved reliability by up to 85% and error tolerance by 25%, and enabled shooting objects by fast ejection and placing multiple objects at once. Performance is compared with other grippers.
Describes an open-source C++ 2D and 3D simulation engine for statics, dynamics and nonlinear deformation of soft bodies made of interspersed materials with differing properties. It handles collisions and volumetric actuation, and runs objects with thousands of degrees of freedom at interactive frame rates on a single desktop thread.
2012, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Takes a supervised learning approach to predict successful grasps for a universal jamming gripper, whose deformation makes physical models hard. Using visual and shape features from RGBD data, the method needs no gripper model, so it also applies to a parallel plate gripper, and lets both pick up varied objects including ones outside the training set.
Presents automated design and manufacture of static and locomoting objects whose function comes from the 3D distribution of materials, using evolutionary algorithms. Cantilever beams evolved to deflect in prescribed profiles were fabricated and matched within 0.5 to 7.6%. A freeform soft robot using volumetrically expanding actuator material was designed for forward locomotion, fabricated and confirmed within 15% error.
Demonstrates evolution of robots with continuous, amorphous, multi-material bodies actuated by periodic volumetric expansion and contraction of materials. It discusses representing evolvable freeform shapes, comparing genotypes that use a level-set threshold, and simulating soft bodies with a relaxation algorithm under actuation, gravity and nonlinear friction. This opens a design space closer to biological systems.
Examines producing cellular materials by freeform fabrication when microstructure features are finer than the process resolution. It studies stochastic foams made through the instability of a viscous thread and several methods for closed-cell foams, without pre-specified cell structures. Foams made from elastic materials may act as novel actuating materials.
2010, Genetic and Evolutionary Computation Conference (GECCO)
Demonstrates evolution of locomoting amorphous multi-material robots actuated by periodic volumetric expansion and contraction. It discusses how to represent evolvable freeform shapes, exploring genotypes that use a level-set threshold, and how to simulate soft bodies with a relaxation algorithm including gravity and nonlinear friction. The work opens a design space resembling biological forms.
2010, Proceedings of the National Academy of Sciences
Presents a universal gripper in which a single mass of granular material replaces fingers: it flows around a target object when pressed on it, then hardens under vacuum to hold it without sensory feedback. Volume changes under 0.5% sufficed to grip objects reliably with forces many times their weight. The operating principle is the jamming transition of granular material.
Argues that soft robots will likely be underactuated and undercontrolled, so part of actuation and control should be embodied in the robot's structural dynamics, termed morphological computation. Two examples are presented, the Manduca sexta caterpillar and a modular tensegrity tower, and the authors explore how they apply to locomotion and control of a soft robot.
Introduces locomotion robots built on tensegrity structures: TR3, a three-strut triangular prism, and TR4, a four-strut quadrilateral prism. Evolutionary algorithms produced controllers yielding static and dynamic gaits in simulation, and a physical robot was built from one simulation model as a proof of concept, suggesting lightweight, strong, fault-tolerant machines.
Develops materials and processes to freeform fabricate complete ionomeric polymer-metal composite actuators and their fabrication substrate with a multi-material system. The printed actuators operated in air for over 40 minutes and 256 bidirectional cycles. Output stress per input power was two orders of magnitude below the best conventional devices; explanations and improvements are discussed.
Works toward producing complete electromechanical systems by freeform fabrication, building a library of compatible elements. Conducting polymers can supply actuation, sensing and control electronics. The authors manually produced air-operable actuators compatible with their power source and mechanical library, present a survey of candidate actuator materials, test two types, and demonstrate complete freeform-producible actuator devices.
Analyzes the static and dynamic properties of the hair clip mechanism, a kinked ribbon with ends pinned like a snap hair clip, validates the theory experimentally, and applies it in two carangiform fish robots. A tethered pneumatic fish reaches 1.40 body lengths per second, about twice a conventional counterpart, and an untethered motor-driven one reaches 2.03 BL/s (42.6 cm/s) with 3.89% efficiency at 3 Hz.
Introduces a single modular robot cell with a magnetic docking interface and two actuated joints that folds from flat into 3D structures, assembling into a three-module 'trilobite' and a five-module 'seal'. A human operator steers a parent robot to dock spare modules into a second identical robot that then separates, a budding-style reproduction. Results are preliminary and teleoperated.
2025, IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings
Presents AutoURDF, an unsupervised method that builds robot description files for unseen robots from point cloud frames. A cluster-based registration model tracks 6-DoF transformations of point clusters, which are used for moving-part segmentation, body topology inference and joint parameter estimation. Outputs are compatible with simulators, and it beats prior methods in registration and topology accuracy on synthetic and real scans.
2025, IEEE International Conference on Computer Vision Workshops
Presents DeepCollide, an implicit neural representation that approximates a robot's collision detection function from sampled collision data when workspace geometry is unavailable. Theory and experiments show advantages over the state of the art in time-cost scaling with training data and degrees of freedom, and in expressing complex workspace geometries. Code is released.
Argues that open-ended physical adaptation requires robots built from a small repertoire of simple modules, so they can grow or heal by consuming parts from other machines or their surroundings and shedding broken ones. Demonstrated on a truss modular robot platform, robots grow bigger, faster and more capable by consuming material from their environment and other robots.
2024, International Conference on Reconfigurable Mechanisms and Robots (ReMAR)
Introduces the Robot Link, a truss-robot module with magnetic connectors that can self-assemble and re-form broken connections, connecting to up to nine other links. Individual links crawl to each other and demonstrate assembly into a tetrahedron, turning 1D-locomotion modules into a 3D maneuverable robot. Three topologies recover from separation after impact, supporting robots that grow and self-repair.
2024, International Conference on Reconfigurable Mechanisms and Robots (ReMAR)
Introduces a particle robot of vibrating units, each with an off-axis spinning mass, enclosed by a rigid chain-link boundary that behaves like a soft robot. Individual particles cannot move directionally, but together they yield deterministic motion from stochastic interactions. Experiments show how particle count affects movement and direction, and the robot can pass tight spaces and flow around obstacles.
Describes CarbonFish, a bistable, underactuated fish robot built from a carbon-fiber-reinforced plastic ribbon using the hair clip mechanism design paradigm. Preliminary tests of the single-actuated prototype show an undulation frequency approaching 10 Hz, suggesting potential for high-speed, biomimetic swimming.
Applies a convolutional encoder-decoder to directly approximate a robot's configuration space, classifying configurations as colliding or free. For a 2-D workspace with a dual-arm robot it reaches an average 97.5% F1-score, with undetected collisions under 2.5% across obstacle translation, rotation and removal. The learned features transfer between workspaces with little or no fine-tuning.
2023, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Describes a single-actuated untethered compliant crawler that uses a prestressed hair clip mechanism as chassis, spine and force amplifier. It gallops at 313 mm/s (1.56 body lengths per second). Experiments find speed is linearly related to actuation frequency and substrate friction, tethering slows the crawler, and asymmetric actuation produces a new galloping gait.
Extends a transformer-based tidying ('knolling') robot with a self-supervised graspability estimation model. When objects are judged ungraspable, for instance in piles or tightly packed, a separating behavior runs before knolling. Combined with the earlier perception and knolling models, the system handles denser tables, with a graspability prediction accuracy of 95.7%.
Teaches a domestic robot to tidy tables by 'knolling', arranging scattered items into neat layouts, treating it as sequential prediction with a transformer that forecasts each object's placement. The method learns a generalizable notion of tidiness, produces diverse solutions for varying object sets, and can incorporate human preferences for personalized arrangements.
2023, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Presents a human-finger-inspired snapping gripper that uses elastic instability in a prestressed semi-rigid skeleton to close rapidly over a wide span. Fingers open to 86 mm and close in 46 ms, 2.7 times better in span and 10.9 times better in speed than a reference gripper. Design principles, simulation and tests on rigid, flexible and limp objects are reported.
Proposes a quasi-static self-modeling framework that learns a predictive model only of high-level quasi-static dynamics rather than continuous dynamics, aimed at legged locomotion. On a 12-degree-of-freedom legged robot it improves on model-free and continuous model-based approaches. Across 80 robot morphologies the benefit correlates with complexity (R2=0.94).
2021, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Presents an open-source untethered quadruped soft robot, about 80% soft by volume, with 3D printed gyroid-infill flexible body and legs driven by four servo motors. With gaits tuned in a real-time soft body simulation, the robot ran at 0.9 m/s (2.5 body lengths per second), faster than most untethered legged soft robots published to date.
Presents PARA, a mostly 3D printed, open-source three-degree-of-freedom robotic arm that lifts 2 kg at a 940 mm reach, with about 2.6 mm precision at 250 mm/s without load. It costs about $3,400 to build, an order of magnitude below comparable market arms, and demonstrates printed parts as practical industrial tools.
Introduces the Boombox, a container that uses sound from low-cost contact microphones to estimate the state of objects inside it, since collisions with the container cause acoustic vibrations. A convolutional network learns to reconstruct visual scenes from multimodal data, enabling state estimation without cameras when occlusion or poor lighting is a problem.
2020, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Presents Titan, a CUDA-based C++ library for GPU-accelerated simulation of soft-body and multi-agent robots with an asynchronous computing model, letting optimization and control run on the CPU while the GPU simulates. It reaches up to 300 million primitive updates per second and is shown on soft-body and multi-agent applications.
Proposes a UAV that detects, tracks and follows other small drones in GPS-denied environments using only onboard computation. A Tiny YOLO detector trained on a 58,647-image dataset, plus simple visual servoing, was validated on a physical platform that followed a target at about 1.5 m/s. Performance was limited by 77% detection accuracy in clutter and eight frames per second.
Demonstrates robots made of many loosely coupled particles that only perform uniform volumetric oscillations phase-modulated by a global signal, with no individual identity. Physical robots of up to two dozen particles and simulated ones of up to 100,000 achieved locomotion, object transport and phototaxis, and locomotion persisted with 20% of particles malfunctioning.
Describes Spyndra, an open-source quadruped robot with rich proprioceptive feedback for studying machine self-awareness, based on the hypothesis that self-awareness is the ability to self-simulate by predicting sensations from actions. The paper covers hardware and software, analyzes two IMU datasets, and compares a simulated gait to a real gait as a baseline.
Demonstrates a 1D printing system, inspired by the ribosome, that folds a single string of source material into specialized robotic automata. A computational optimization tool derives designs from high-level goals. Three distinct robots for different locomotion tasks were made from the same material, and recycled material was reused to produce new designs.
Describes a robotic hand that uses phase transition of granular material in multiple fingers to achieve dexterity with minimal actuation. With only two actuators it performs six fundamental dexterous manipulations plus precision and power grasps, and is shown on real-world grasping and manipulation tasks.
Argues that shape-related tactile information is better recovered from shear strain than from normal strain. A contact mechanics analysis shows that sensing shear gives a robust edge detection mechanism. A jamming-based robot gripper used as a tactile sensor confirmed that shear processing gives accurate edge information that is invariant to pressure.
Addresses planning of assembly for stochastically reconfigurable systems, where module availability in space and time is uncertain. An on-line algorithm samples assembly paths to the target structure, is guaranteed to find a path for finite connected objects, pursues several paths in parallel, adapts to changing conditions and predicts remaining assembly time.
Soldercubes are small, light, low-cost self-reconfiguring modular robot modules designed for volume manufacturing. Each uses a 2 g soldering connector with no moving parts, making the system among the cheapest and lightest of comparable designs. The paper details the module design, reports lattice-configuration experiments, explores non-lattice uses, and releases all design files as open-source hardware.
Describes a solid-state, lightweight connector for modular robots that forms reversible soldered connections by heating a low melting point alloy. No external manipulation is needed, energy is used only when switching, and power and signals can pass through. Connections hold 173 N and survive hundreds of cycles before failure.
Tests whether HyperNEAT gaits evolved first in simulation and then transferred to a physical quadruped outperform a simpler hand-tuned encoding. The hypothesis was confirmed, producing the fastest gaits yet seen for this robot, beating nine algorithms from three earlier papers, and showing that generative encodings work for real-robot gait problems.
Presents a robot that autonomously traverses and manipulates a 3D truss, using translational and rotational motions across joints and specially designed building blocks made of bidirectional geared rods. Traversal plans were analyzed for speed, robustness and repeatability, with paths covering eight joints and automatic assembly and disassembly of elements demonstrated.
Presents a passive universal gripper made of granular material in an elastic membrane that conforms to an object, hardens under vacuum to grip, and releases using positive pressure. Positive pressure improved reliability by up to 85% and error tolerance by 25%, and enabled shooting objects by fast ejection and placing multiple objects at once. Performance is compared with other grippers.
Introduces Aracna, an open-source quadruped built from off-the-shelf and 3D-printed parts, as a platform for gait-learning algorithms. All motors sit in the body, driving hip and knee joints through four-bar linkages, which fixes a previous platform's heavy legs and yields non-intuitive kinematics that suit evolutionary and learning methods.
Applies the RL PoWER policy-learning algorithm with a spline parameterization to learn gaits on the open-source 3D-printed QuadraTot robot, entirely through physical experiments. It found the fastest gait reported from physical-only learning, 16.3% faster than HyperNEAT, with gaits less taxing on the robot and more repeatable.
2012, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Takes a supervised learning approach to predict successful grasps for a universal jamming gripper, whose deformation makes physical models hard. Using visual and shape features from RGBD data, the method needs no gripper model, so it also applies to a parallel plate gripper, and lets both pick up varied objects including ones outside the training set.
Introduces Mini Twist, a compact twisting-string linear actuator for long-range drive, with its design, components and a theoretical model of the transmission. Experiments examine how string diameter, tension, strand count and strand type affect performance, and the analysis assesses the model's validity.
A short study of ambient wind energy harvesting with a cross-flow stalk-leaf piezoelectric structure, treated as an aerodynamic-instability driven cantilever-pendulum system. Wind tunnel experiments from 0 to 8 m/s measured output voltage and average power, and showed a bistable response with a critical speed region where the system either stays flat or bursts into large-amplitude oscillation.
2011, Genetic and Evolutionary Computation Conference Companion (GECCO Companion)
Compares algorithms for automatically generating quadruped gaits, tested entirely on a physical robot because accurate simulators are often unavailable. Parameterized motion models optimized by six local search methods are compared with neural networks evolved by the HyperNEAT generative encoding. All outperformed a manually designed gait, and HyperNEAT was best, producing gaits nearly nine times faster than the reference.
Describes a 3.89 g flapping-wing hovering ornithopter built with 3D-printed wings and mechanical parts. Printing allows arbitrary wing shapes and reduces wing design cycles to minutes. The insect achieved passively stable untethered hovering flight lasting 85 seconds, showing the utility of printed materials for flapping-wing experimentation.
2010, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Describes Tetrabot, a fully enclosed mobile robot with no moving parts exposed to the environment that locomotes by exploiting resonance. Actuators on a central sphere pull strings attached to a tetrahedral frame to excite a natural vibration mode. A simple open-loop controller produced directed hopping and sliding in a prototype, and rolling was investigated theoretically.
2010, Proceedings of the National Academy of Sciences
Presents a universal gripper in which a single mass of granular material replaces fingers: it flows around a target object when pressed on it, then hardens under vacuum to hold it without sensory feedback. Volume changes under 0.5% sufficed to grip objects reliably with forces many times their weight. The operating principle is the jamming transition of granular material.
Book chapter describing the design and fabrication of a passively stable 24-gram flapping machine that hovers at an insect-like Reynolds number of about 8,000. It uses the clap and fling effect and passive wing bending to simplify the design and improve performance, and the authors hope it aids smaller machines and understanding of insect flight.
2009, ASME/IFToMM International Conference on Reconfigurable Mechanisms and Robots
Addresses the design of passive robotically reconfigurable truss structures whose elements can be inserted and removed singly in random-access order, and progress toward a hinge robot that manipulates them and traverses trusses. With reconfiguration algorithms, the authors envision a machine metabolism in which structures are decomposed and recomposed autonomously.
Presents a freeform fabricator with a six degree-of-freedom print head for additive deposition onto existing structures, useful for repair, augmentation and in situ printing. A parallel mechanism with stationary motors was developed and optimized through kinematic simulation. A working printer was built, an additive 3D structure was demonstrated, and its accuracy and reliability were analyzed.
Presents a control scheme in which independent struts of a tensegrity robot coordinate to achieve locomotion by using the dynamical coupling from pre-stress stability as an ad-hoc communication channel. Time-sensitive spiking neural networks exploit this "morphological communication". The result supports morphological computation in robot control and suggests relevance to embodied computation in biological systems such as tendon networks.
Studies a parallelized flapping piezoelectric leaf generator that harvests ambient wind energy through wind-induced vibration. Prototypes in several configurations were built and tested, and a vertical-stalk L-shape design gave better power density than horizontal piezoelectric configurations. A power density gap to commercial wind turbines remains, though the devices are solid-state, low-cost and scalable.
Reports a robotic system that paints physical artworks with media such as acrylic on canvas, using a 6-DOF arm and a learning algorithm that determines brushstrokes to transfer a source image. A genetic algorithm used a painting simulation to maximize similarity to the target, and artists can steer results through palette, brush and algorithm parameters.
Demonstrates a passively stable, untethered 24-gram flapping-wing machine capable of hovering at an insect-like Reynolds number of about 8,000. The authors note only two earlier machines achieved untethered hovering flight, and suggest the passive-stability architecture could inform insect-sized hoverers and the aeroelastic principles of insect flight.
The IEEE Robotics & Automation Magazine article (also listed as "Demonstration of a Passively Stable, Untethered Flapping-Hovering Micro-Air Vehicle") describing a flapping-wing micro-air vehicle that achieves passively stable, untethered hovering, a capability previously seen in insects and hummingbirds and rarely reproduced in machines.
Introduces locomotion robots built on tensegrity structures: TR3, a three-strut triangular prism, and TR4, a four-strut quadrilateral prism. Evolutionary algorithms produced controllers yielding static and dynamic gaits in simulation, and a physical robot was built from one simulation model as a proof of concept, suggesting lightweight, strong, fault-tolerant machines.
2006, International Conference on Advanced Robotics (ICAR)
Introduces a legged robot based on a triangular tensegrity prism actuated by contracting its transverse cables. A genetic algorithm designed controllers in simulation that produced multiple forward gaits, and a physical robot built from the simulated design achieved forward locomotion, suggesting very lightweight and robust mobile robots.
A four-legged robot continuously builds an internal model of its own body by performing actions, observing the sensory results and refining competing candidate models. It uses the self-model to generate forward locomotion and, after a leg is damaged, updates the model and finds a new gait, demonstrating resilience without a pre-programmed model of itself.
Describes a relaxation-based method for simulating 2D and 3D compound kinematic mechanisms, in which iterative propagation of node displacements reveals the kinematics. It is slower than simultaneous differential equation approaches but more accurate, more robust to singularities, and able to handle complex entangled chains and nonlinear links such as wrapping strings and contacts. Applications include robotics, design and biomechanics.
Presents a process for evolving flapping flight control patterns for an ornithopter, together with a realistic kinematic mechanism that can generate the motion in practice. The target is a hovering design that lifts only through drag on a flat wing. The goal is to carry simulated models over to a physical, untethered flier.
2005, AIAA Guidance, Navigation, and Control Conference and Exhibit
Applies particle filters to search and exploration by autonomous rovers, adaptively controlling the resampling stage to prevent premature convergence, so the filter tunes itself for unknown environments and gains robustness to faults. Simple control strategies use the particle distribution to steer rovers toward areas of high expected payoff, and the approach is compared with earlier particle-filtering algorithms.
2005, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
Tests whether robots with highly coupled mechanical structures, where one actuator moves many locations and vice versa, allow redundancy in control. Using three- and four-strut tensegrity prisms, locomotion controllers were evolved in simulation. Multiple, substantially different control strategies produced qualitatively similar forward locomotion, indicating considerable control redundancy.
Introduces a method for evolving dynamic gaits on a physical parallel-actuated pneumatic robot, assuming only that locomotion is rhythmic. A genetic algorithm evolves open-loop controllers whose encoding shapes speed and pattern while ensuring rhythmicity. Future plans include closed-loop controllers and reducing hardware trials.
Examines whether periodic motion emerges under evolutionary pressure for energy-efficient, long-range, partially passive brachiation in simulated gibbon-like creatures with a two-arm, torso morphology. Genomes encode initial position and shoulder torque functions, and Pareto-optimal individuals were selected. Energy-efficient long-range motion resulted in periodic motion.
Presents a method for automatically tuning robot simulators with deep learning, replacing the hand-tuning that usually precedes sim-to-real transfer. The approach generates tens of thousands of randomized simulators and identifies which parameters best represent the real world. The resulting nearby simulators are then used to quickly produce a walking policy that works on the real robot.
Introduces an open-source framework for benchmarking end-to-end deep reinforcement learning on foosball, a fast game requiring fine motor control. The platform provides a high-fidelity MuJoCo physics simulation intended to support analysis of how performance transfers from simulation to a real table, and is offered as a standardized resource for the RL community.
Robots that interact socially with people: humanoid robotic faces, facial co-expression, lip and gaze generation, communication between robots, and visual behavior modeling of other agents.
Separates robot gaze into target selection by an asynchronous vision-language model, reactive fixation, and a Realistic Gaze Transformer, a Transformer-VAE trained on retargeted markerless human recordings for an 11-DoF robot. A blink predictor triggers eyelid templates. RGT-Base gets the lowest seven-joint MSE on a two-second benchmark; perceived naturalness and open-vocabulary target accuracy are not yet evaluated.
Presents a humanoid robot face with soft silicone lips driven by a 10-DoF mechanism, and a self-supervised pipeline using a VAE and a facial action transformer to infer lip trajectories directly from speech audio. It outperforms amplitude-based baselines in visual lip-audio coherence and generalizes to 10 languages unseen during training.
Trains a humanoid robot face with 26 degrees of freedom to anticipate a human's upcoming facial expression and execute it simultaneously, rather than mimicking with delay. The robot learns to predict a smile about 839 milliseconds before the human smiles and uses a learned inverse kinematic facial self-model to coexpress it. The authors suggest this could improve human-robot interaction.
Presents an animatronic robotic face with soft skin and a vision-based self-supervised framework for facial mimicry, needing no kinematic model, camera calibration or predefined expression set. Splitting learning into a generative model and an inverse model lets it train from one motor babbling dataset, and evaluations show accurate, diverse mimicry across human subjects.
Asks whether an observer can predict an actor robot's conditional behavior from visual observation alone, without access to its internal states or explicit goals, a notion called Theory of Behavior. In an example, an observer AI watches an actor seek food around obstacles and learns to visualize the actor's likely trajectory in new situations.
2021, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Proposes an end-to-end long-term visual prediction framework giving robots Visual Perspective Taking and Theory of Behavior, demonstrated on visual hide-and-seek. Unlike frame-by-frame predictors, the agent predicts directly to multiple future timestamps (25 s), extrapolating 175% beyond its training horizon.
Describes Eva, an inexpensive open-source adult-sized humanoid head for research on emotional communication between humans and machines. It reproduces facial expressions, head movements and speech using 25 muscles, including 12 facial muscles with up to 15 mm skin displacement, and costs about $900 in hardware.
Trains embodied agents from scratch to play Visual Hide and Seek, where a prey avoids a predator using only egocentric, partially observed views with occluding obstacles. Experiments suggest the agent learns to predict its own visibility, and agent weaknesses such as slower speed make learning harder but cause more useful features to be learned.
Simulated robots with a single communication channel evolve from non-communicating populations when reproduction requires finding a mate. Indexical communication arose in 99% of experiments, and 9% of populations evolved a symbolic scheme conveying two spatial dimensions over a one-dimensional channel. The results suggest symbolic communication can emerge under natural selection without cognitive preadaptations.
A Current Biology commentary on robot-based simulations showing that communication can emerge spontaneously among evolving agents. It notes that deceptive signalling arises in purely competitive settings, while cooperative communication emerges only under group or kin selection, and places these results in the broader history of communication in evolution.
Core machine learning research: understanding and visualizing deep networks, transfer learning, representation learning, label representations, data augmentation, hypernetworks, meta-learning, and evaluation of AI systems.
Presents EvoAug, a pipeline that uses generative models (conditional diffusion, few-shot NeRFs) with an efficient evolutionary algorithm to learn task-specific augmentations, structured as stochastic augmentation trees that compose augmentations hierarchically. It performs strongly on fine-grained classification and few-shot learning, and discovers augmentations aligned with domain knowledge even with little data.
Proposes that a 'self' can be isolated as the invariant, slowly changing portion of a cognitive process, and tests this on robots under constant-task versus continual learning. Continual learning produces an invariant subnetwork that is significantly more stable than in controls and is functionally important: preserving it aids adaptation and damaging it impairs performance. Results hold across three robots in locomotion and manipulation.
Separates robot gaze into target selection by an asynchronous vision-language model, reactive fixation, and a Realistic Gaze Transformer, a Transformer-VAE trained on retargeted markerless human recordings for an 11-DoF robot. A blink predictor triggers eyelid templates. RGT-Base gets the lowest seven-joint MSE on a two-second benchmark; perceived naturalness and open-vocabulary target accuracy are not yet evaluated.
Argues that adaptive imagination needs a generative world model plus a selection process that keeps imagined scenarios grounded, and tests this in few-shot vision with EvoAug, which pairs generative augmentation with evolutionary search. Without selection, generative augmentation hurts performance; with it, the system outperforms classical augmentation across six fine-grained datasets, with the largest gains where data is scarcest.
2026, IEEE International Conference on Acoustics Speech and Signal Processing
Proposes LyTimeT, a two-phase framework for extracting dynamical state variables from video despite visual distractors. A TimeSformer autoencoder learns robust latent states, then physically meaningful dimensions are selected by correlation analysis and dynamics refined with a Lyapunov stability regularizer. On five synthetic and four real systems it gives estimates closest to ground truth and the lowest error among baselines.
Studies 'mirage' behavior, where vision-language models answer image questions confidently without an image, arguing it has two modes. Mirage Probes show the behavior is linearly decodable from internal activations in two open-source VLMs, ruling out lexical confounds with a Naive Bayes baseline, and a Prior Harnessing Index separates textual-bias from spurious-image regimes, which need different mitigations.
Studies projection-faithful reconstruction of a 3D wireframe from a single engineering line drawing, posed as conditional depth estimation over drawn pixels. A latent diffusion model conditioned on the sketch, with optional partial-depth input, is trained on over three million synthetic image-depth pairs from CAD wireframes. Scaling from 256 to 512 resolution roughly halves error, reaching 3.9% best-of-five normalized depth error.
Proposes a challenge to develop building foundation models whose HVAC control policies transfer to new buildings with minimal configuration, addressing the lack of policy transfer. Participants build policies on simulated buildings and are evaluated on unseen ones that vary in size, layout, weather, materials and HVAC system, aiming to scale building optimization for climate impact.
Presents PXRD NET, a diffusion-based generative model trained on 45,229 known structures that solves nanocrystal structures from the chemical formula and a size-broadened powder diffraction pattern. It solves simulated nanocrystals as small as 10 Angstroms across 200 materials, with verifiable candidates four times out of five and about 7% average error, and handles noisy experimental data.
A large multi-institution team, including Lipson, proposes applying the common task framework (shared datasets, tasks and evaluation metrics, as used to drive progress in machine learning) to accelerate AI-driven scientific discovery in dynamical systems and other areas of science.
Introduces ArticFlow, a two-stage flow-matching framework that learns an action-controlled velocity field from noise to target point sets, using a latent flow for a shape prior and a point flow conditioned on action. On MuJoCo Menagerie it works as both generative model and neural simulator, with higher kinematic accuracy and shape quality than object-specific and action-conditioned baselines.
Tests the assumption that an individual's biometrics are statistically uncorrelated by training Bi-Encoder networks (ResNet-50 and Vision Transformer) with contrastive loss on fingerprint, iris and cross-modal matching, using 274 subjects. Iris-to-iris matching reaches 91 ROC AUC, indicating left and right irises are correlated; fingerprints reproduce prior findings, and cross-modal matching is only slightly above chance.
Formulates planar linkage kinematic synthesis as cross-domain image generation, with an RGB dataset ranging from crank-rocker and crank-slider mechanisms to eight-bar linkages such as Jansen's. A shared-latent VAE synthesizes unseen motion curves, and speed encoded as color gradients allows conditioning on velocity profiles. Preliminary results are shown on three datasets of increasing complexity.
2025, IEEE International Conference on Computer Vision Workshops
Presents DeepCollide, an implicit neural representation that approximates a robot's collision detection function from sampled collision data when workspace geometry is unavailable. Theory and experiments show advantages over the state of the art in time-cost scaling with training data and degrees of freedom, and in expressing complex workspace geometries. Code is released.
2025, IEEE International Conference on Computer Vision Workshops
Gives a quantitative benchmark of Soft Actor-Critic and DDPG agents in the Smart Buildings Control Suite simulator for Building SB1. It looks at learning efficiency, performance against a baseline schedule, and generalization across unseen seasons. Both agents beat the baseline and generalize, and extracting static schedules from policies loses little performance for up to 8-hour bins.
Proposes RL-AUX, a reinforcement learning framework that creates auxiliary tasks by assigning auxiliary labels to training examples, rewarding the agent when its choices improve the primary task, and also explores per-example auxiliary loss weights. On CIFAR-100 grouped into 20 superclasses it beats human-labeled auxiliary tasks and matches a bi-level optimization baseline, with similar results on other datasets.
Presents a humanoid robot face with soft silicone lips driven by a 10-DoF mechanism, and a self-supervised pipeline using a VAE and a facial action transformer to infer lip trajectories directly from speech audio. It outperforms amplitude-based baselines in visual lip-audio coherence and generalizes to 10 languages unseen during training.
Presents PXRD NET, a diffusion-based generative model trained on 45,229 known structures that solves nanocrystal structures from the chemical formula and a size-broadened powder diffraction pattern. It solves simulated nanocrystals as small as 10 Angstroms across 200 materials, finding verifiable candidates four times out of five with about 7% average error, and also handles noisy experimental patterns.
Proposes a general principle for distilling hidden state variables that are more aligned with human intuition, without relying on prior physical knowledge. Since AI-found variables are often hard to interpret, the authors test on experimental and simulated systems, where the variables generated closely resemble those chosen independently by human scientists.
Applies diffusion models to generating architectural floor plans, adapting the U-Net architecture and adding upscaling techniques that reduce computation while keeping high-resolution output. The authors argue that suitably tuned and conditioned diffusion models capture implicit architectural semantics and improve design efficiency and creativity.
2024, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Uses neural fields to let a robot learn a task-agnostic self-model of its kinematics as a neural-implicit query model from 2D images annotated with camera poses and configurations, without depth or geometry knowledge. It introduces an encoder-based dynamic density field architecture with curricular sampling. On a 7-DoF robot the Chamfer-L2 error is 2% of workspace size, demonstrated on motion planning.
Tackles exact parameter reconstruction of a black-box neural network with only query access. It exploits the inductive bias of random initialization plus first-order optimization and a novel query generation algorithm producing maximally informative samples. It reconstructs a network with over 1.5 million parameters and one 7 layers deep, with max parameter difference under 0.0001.
Presents the Smart Buildings Control Suite, an open-source HVAC control benchmark with real telemetry from 11 buildings over 6 years, a lightweight data-driven simulator per building, and a modular physics-informed neural network model as an alternative. It spans varied climates and sizes, is Gym-compatible, and aims to evaluate control policies that scale across buildings.
Presents an ML approach to powder crystallography that estimates unit-cell electron density with a variational coordinate-based deep network, handling ambiguities in orientation, symmetry and resolution. Using simulated powder x-ray diffraction plus partial chemical composition, it reaches up to 93.4% average structural similarity on cubic and trigonal systems.
Challenges the assumption that fingerprints from different fingers of the same person are unrelated. Using deep twin networks to extract fingerprint representations, the authors find strong similarity between fingers of one person with above 99.99% confidence, even controlling for sensor modality. Ridge orientation near the center explains much of it, while minutiae are almost nonpredictive; forensic efficiency could improve substantially.
Addresses inner-loop over-fitting in gradient-based meta-learning by sharing gradient information from earlier tasks and concurrent tasks in the batch, scaled by meta-learned parameters. On two few-shot classification datasets, gradient sharing allows larger inner-loop learning rates and speeds up meta-training by up to 134%.
Examines SATNet, a differentiable MAXSAT layer previously reported to solve visual Sudoku. Without intermediate labels linking digit images to logical symbols, SATNet gets 0% test accuracy, a failure traced to the symbol grounding problem. The authors propose an MNIST-based sanity test for differentiable symbolic solvers and discuss causes of the failure and how to avoid them.
Applies a convolutional encoder-decoder to directly approximate a robot's configuration space, classifying configurations as colliding or free. For a 2-D workspace with a dual-arm robot it reaches an average 97.5% F1-score, with undetected collisions under 2.5% across obstacle translation, rotation and removal. The learned features transfer between workspaces with little or no fine-tuning.
Extends a transformer-based tidying ('knolling') robot with a self-supervised graspability estimation model. When objects are judged ungraspable, for instance in piles or tightly packed, a separating behavior runs before knolling. Combined with the earlier perception and knolling models, the system handles denser tables, with a graspability prediction accuracy of 95.7%.
Teaches a domestic robot to tidy tables by 'knolling', arranging scattered items into neat layouts, treating it as sequential prediction with a transformer that forecasts each object's placement. The method learns a generalizable notion of tidiness, produces diverse solutions for varying object sets, and can incorporate human preferences for personalized arrangements.
Observes that classical initializations such as Glorot and He, applied directly to a hypernetwork, produce main-network weights at the wrong scale. The authors develop principled initialization techniques for hypernetworks and show they give more stable main-network weights, lower training loss and faster convergence.
Presents a variational query-based multi-branch deep network that infers crystal structure end-to-end from degraded powder x-ray diffraction data plus partial chemical composition. Structures are represented as a Cartesian mapped electron density. On simulated cubic and trigonal systems it achieves up to 93.4% average similarity to ground truth on unseen materials.
Addresses the problem of identifying the state variables needed before physical laws can be formulated, by having an AI extract them from raw experimental data such as video. The framework finds a compact set of variables describing a dynamical system without prior physical knowledge. The text motivates the work with the historical role of defining variables like mass or temperature.
Introduces Binary Labels, in which each class is represented as a large binary vector, and trains a network with multi-task learning on both one-hot and binary label tasks. Networks trained this way show higher accuracy across many datasets and architectures, from scratch and with transfer learning. They also do better with limited training data and on harder problems.
Finds that label representation strongly affects trained models, for example regressing audio labels instead of categorical probabilities gives more reliable classification. The authors hypothesize high-dimensional, high-entropy labels give a stronger error signal. Such labels match categorical accuracy but yield more robustness to adversarial attacks and work better with limited training data.
Proposes a principle for determining how many state variables a system has and what they might be, directly from high-dimensional video. Applied to recordings of systems from elastic double pendulums to fire flames, the algorithm finds the intrinsic dimension of the dynamics and candidate sets of state variables without prior knowledge of the physics.
Presents an animatronic robotic face with soft skin and a vision-based self-supervised framework for facial mimicry, needing no kinematic model, camera calibration or predefined expression set. Splitting learning into a generative model and an inverse model lets it train from one motor babbling dataset, and evaluations show accurate, diverse mimicry across human subjects.
Introduces the Boombox, a container that uses sound from low-cost contact microphones to estimate the state of objects inside it, since collisions with the container cause acoustic vibrations. A convolutional network learns to reconstruct visual scenes from multimodal data, enabling state estimation without cameras when occlusion or poor lighting is a problem.
Asks whether an observer can predict an actor robot's conditional behavior from visual observation alone, without access to its internal states or explicit goals, a notion called Theory of Behavior. In an example, an observer AI watches an actor seek food around obstacles and learns to visualize the actor's likely trajectory in new situations.
Studies whether convolutional networks can predict static and dynamic properties of cantilever beams from raw cross-section images. Models predicted volume, maximum deflection and eigenfrequencies with 4.54% and 1.43% mean average percentage error against finite element analysis, and run over 1000 times faster, suggesting use as surrogate models in design optimization.
2021, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Proposes an end-to-end long-term visual prediction framework giving robots Visual Perspective Taking and Theory of Behavior, demonstrated on visual hide-and-seek. Unlike frame-by-frame predictors, the agent predicts directly to multiple future timestamps (25 s), extrapolating 175% beyond its training horizon.
Addresses slow actuation and limited cycle life of silicone/ethanol soft actuators caused by low thermal conductivity. Adding 8 wt.% diamond nanoparticle filler raised conductivity from 0.190 to 0.212 W/mK and improved speed and cycle count without changing mechanical properties. An LSTM network predicted actuation force in multi-cyclic tests.
Trains embodied agents from scratch to play Visual Hide and Seek, where a prey avoids a predator using only egocentric, partially observed views with occluding obstacles. Experiments suggest the agent learns to predict its own visibility, and agent weaknesses such as slower speed make learning harder but cause more useful features to be learned.
Shows that neural network agents can be reduced to low-dimensional embeddings by vectorizing weights from existing networks and learning a generative model over weight space. In a Cart-Pole case study, new networks are generated from embeddings without simulator data. Interpolating between good and bad agents gives intermediate performance, and extrapolation helps up to a point.
Classifies maize versus grass and grass-like weed segments using surrounding context, by feeding a CNN a central image plus a border of its surroundings. In a case study on foxtail and yellow nutsedge, context nearly halved error from 7.1% to 4.3%, and one epoch with context outperformed 50 epochs without.
Classifies maize versus grass and grass-like weed segments using surrounding context, by feeding a CNN a central image plus a border of its surroundings. In a case study on foxtail and yellow nutsedge, context nearly halved error from 7.1% to 4.3%, and one epoch with context outperformed 50 epochs without.
Classifies maize versus grass and grass-like weed segments using surrounding context, by feeding a CNN a central image plus a border of its surroundings. In a case study on foxtail and yellow nutsedge, context nearly halved error from 7.1% to 4.3%, and one epoch with context outperformed 50 epochs without.
A deep learning model detects northern leaf blight symptoms in maize from UAV images with 95% accuracy. Transfer learning sped up model optimization, and the method detects symptoms at a very fine spatial scale, offering growers and breeders an alternative to time-consuming visual phenotyping.
Constructs a general variational family for ensemble-based Bayesian neural networks that includes dropout as a special case, and presents two members that work well with batch normalization while keeping parameter and programming overhead low. On a ResNet-18 trained on ImageNet, they improve predictive accuracy and achieve almost perfect calibration.
Describes a two-step approach to millimeter-level plant disease detection in aerial photographs, using deep learning with data labeled via crowdsourcing, since only trained experts can accurately identify symptoms. The text provided is cut off before results, so details beyond the approach are limited.
Trains a Mask R-CNN to segment individual northern leaf blight lesions in UAV images of maize. On a hold-out test set the mean intersection over union was 0.73, with average precision 0.96 at an IOU threshold of 0.50 and 0.61 across thresholds from 0.50 to 0.95. The work supports high-throughput quantitative disease phenotyping.
An archival copy of a blog post listing seven commonly held myths in machine learning research as of early 2019. They concern TensorFlow being a tensor library, representativeness of image datasets, test-set use for validation, use of every datapoint in training, the need for normalization in deep residual networks, attention versus convolution, and the robustness of saliency maps.
Learns a self-model of a robot's dynamics from randomly sampled actions and uses it as the environment for reinforcement learning, so policies are trained without seeing states from the real environment. The policies transfer to the real robot, training is more data-efficient, and new tasks can be learned without further data collection.
2018, Genetic and Evolutionary Computation Conference (GECCO)
Introduces Autostacker, an AutoML architecture combining hierarchical stacking with an evolutionary algorithm to search parameters, requiring no domain knowledge or feature preprocessing. It evolves pipelines that can be used directly or refined by experts, and on fifteen datasets achieves state-of-the-art or competitive accuracy and time cost against other AutoML systems.
Introduces Autostacker, an AutoML architecture pairing hierarchical stacking with an evolutionary algorithm for efficient parameter search, with no need for prior domain knowledge or feature preprocessing. Evolved pipelines can serve as is or as starting points for experts, and results on fifteen datasets are state of the art or competitive.
Presents a version of Autostacker that uses parallel hill climbing for parameter search, sharply reducing AutoML time. Its hierarchical stacking explores combinations and structures of models rather than a fixed-order pipeline and needs no domain knowledge. Experiments show better accuracy and time cost than human initial trials and a popular AutoML system.
Generalizes HashedNets into ArbNets, a framework for arbitrary weight-sharing, and shows that common networks can be expressed as ArbNets with different hash functions. Two new hash functions, the Dirichlet hash and the Neighborhood hash, are used to show experimentally that balanced and deterministic weight-sharing improves network performance.
Describes self-replicating neural networks that replicate by learning to output their own weights, trained with gradient-based or non-gradient methods, and a regeneration method that injects the network with predictions of its own parameters. Alternating regeneration and optimization worked best. A network that also classifies MNIST shows a trade-off between classification and replication.
Proposes a supervised deep generative network that predicts photorealistic RGB images of laser-browned dough from laser speed, energy flux and dough moisture. Validated on a CO2 laser browning pipeline, the model characterizes surface color and pattern, and is intended as a step toward CAD software for food processing.
A pipeline of convolutional neural networks identifies northern leaf blight lesions in field images of maize. Several CNNs classify small image regions, their predictions form heat maps, and a final CNN classifies the whole image, reaching 96.7% accuracy on held-out test images. The authors suggest uses in phenotyping, breeding and targeted pesticide application.
2016, ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)
Proposes a data-driven way to build multiple-choice questions from students' own past submissions. The authors develop a multinomial-logit discrete choice model, derive an objective for selecting optimally discriminative option sets, give an algorithm for a globally optimal solution, and test it with synthetic experiments and a user study, including crowd-sourcing tests from online forums.
Proposes organizing heterogeneous web tutorials, blog posts and lecture notes into a textbook-like path from a learner's prior knowledge to a target topic. Contributions are a supervised classifier for identifying explained and assumed terms in a document, and an algorithm that finds optimal paths through resources given the learner's goal and background.
Investigates convergent learning, whether separately trained neural networks learn the same representations. It introduces three alignment techniques: bipartite matching, sparse prediction with clustering, and spectral clustering. Findings include that some features are learned reliably across networks while others are not, and that units span low-dimensional subspaces.
Presents two open-source tools for interpreting deep convolutional networks. One displays live activations at every layer as a trained convnet processes an image or webcam stream. The other visualizes learned features through regularized optimization in image space, with several new regularizers that yield clearer, more recognizable images.
Proposes data smashing, a feature-free way to measure similarity between arbitrary data streams. Using an algebraic structure on symbolic models, each stream has an anti-stream, and similarity is how fully one annihilates the other's statistical structure into noise. Applied to epileptic brainwaves, cardiac rhythms and astronomical photometry, it matched or exceeded expert-tuned algorithms.
Journal version of the data smashing work: a universal principle for quantifying similarity between data sources without features, training or prior knowledge. Stochastic symbolic generators can be added and inverted, so each stream has an anti-stream. Applications in seizure brainwaves, cardiac anomalies and astronomical classification meet or exceed domain-tuned methods.
Experimentally quantifies how general versus specific the features are at each layer of a deep convolutional network. Transferability suffers from higher-layer specialization and from optimization difficulties when splitting co-adapted layers. Transfer degrades as tasks grow more distant but still beats random features, and transferred initialization can boost generalization even after fine-tuning.
2013, Philosophical Transactions of the Royal Society A Mathematical Physical and Engineering Sciences
Chattopadhyay and Lipson present an algorithm for learning probabilistic finite automata from quantized time series, giving compact generative models of stochastic processes and a basis for comparing and classifying data streams. Published in the Philosophical Transactions of the Royal Society A.
Presents a fast method to re-purpose existing word embeddings for a supervised task, taking an embedding and labeled data and producing a new embedding in the same space with better predictive performance. It improves sentiment classification over several baselines, and is most useful when the training set is small.
2012, Proceedings - IEEE International Conference on Robotics and Automation/Proceedings
Takes a supervised learning approach to predict successful grasps for a universal jamming gripper, whose deformation makes physical models hard. Using visual and shape features from RGBD data, the method needs no gripper model, so it also applies to a parallel plate gripper, and lets both pick up varied objects including ones outside the training set.
Presents multi-modal symbolic regression (MMSR) for learning symbolic models of hybrid dynamical systems from unlabeled time series. It combines clustered symbolic regression, which identifies distinct behaviors and their equations, with transition modeling that infers symbolic inequalities as boundaries. It beat other multi-modal approaches on four synthetic sets and recovered classical transistor modes from measurements.
Describes visualizations for choosing the many brittle training parameters of Restricted Boltzmann Machines, adapted from Hinton's practical guide, with Python and Octave/Matlab code snippets. It also demonstrates a successful application of an RBM to learning a representation of synthetic 3D shapes.
Presents a fully automated system that builds 3D representations from 2D wireframe drawings using a neural network together with a genetic search algorithm. It builds on the hypothesis that angles and other geometric properties in the 2D scene support human-like 3D reconstruction, with the aim of enabling efficient and robust 3D sketch interfaces.
Applies the estimation-exploration algorithm to active learning for grammatical inference of deterministic finite automata, with the algorithm proposing its own training data. It outperforms one version of evidence-driven state merging on random DFAs, which only works well for particular balances of positive labelings. The authors propose a more general way of generating DFAs for benchmarking.
Introduces a clustering method for irregularly shaped data using high-order neurons, in which the synaptic weight is replaced by a high-order tensor in homogeneous coordinates. First and second order cases correspond to a classic neuron and an ellipsoidal-metric neuron. The neurons follow a maximum-correlation principle with simple Hebbian learning, separate close and overlapping clusters, and give superior results on the Iris data.
Discusses high-order neurons in competitive learning, where geometric shapes replace point neurons through high-order tensors in homogeneous coordinates. Eigentensors of the synaptic tensors reveal coefficients of polynomial rules the network carries out, which may expose symbolic structure in data. The neurons use maximum-correlation activation with local Hebbian learning and separate close, overlapping clusters.
Presents a method for automatically tuning robot simulators with deep learning, replacing the hand-tuning that usually precedes sim-to-real transfer. The approach generates tens of thousands of randomized simulators and identifies which parameters best represent the real world. The resulting nearby simulators are then used to quickly produce a walking policy that works on the real robot.
Introduces an open-source framework for benchmarking end-to-end deep reinforcement learning on foosball, a fast game requiring fine motor control. The platform provides a high-fidelity MuJoCo physics simulation intended to support analysis of how performance transfers from simulation to a real table, and is offered as a standardized resource for the RL community.
Describes CDaug, a data augmentation method that uses ControlNet conditioned on existing images, with captions from the LLaVA2 multimodal model, to generate synthetic variations without fine-tuning a diffusion model. It relies on open-source, modular components and reports improved performance on 7 fine-grained datasets in few-shot and full-data settings across several architectures.
Applied AI: Agriculture, Buildings, Materials and More
Applying machine learning to real-world problems: crop disease and weed detection from aerial imagery, smart-building control, crystal structure determination from diffraction data, biometrics, and other application domains.
Proposes a challenge to develop building foundation models whose HVAC control policies transfer to new buildings with minimal configuration, addressing the lack of policy transfer. Participants build policies on simulated buildings and are evaluated on unseen ones that vary in size, layout, weather, materials and HVAC system, aiming to scale building optimization for climate impact.
Presents PXRD NET, a diffusion-based generative model trained on 45,229 known structures that solves nanocrystal structures from the chemical formula and a size-broadened powder diffraction pattern. It solves simulated nanocrystals as small as 10 Angstroms across 200 materials, with verifiable candidates four times out of five and about 7% average error, and handles noisy experimental data.
A correction notice for the Nature Materials paper on ab initio structure solution from nanocrystalline powder diffraction data. It states that the article's license changed from open-access Creative Commons to the standard Springer Nature license, and that the HTML and PDF versions were corrected.
Tests the assumption that an individual's biometrics are statistically uncorrelated by training Bi-Encoder networks (ResNet-50 and Vision Transformer) with contrastive loss on fingerprint, iris and cross-modal matching, using 274 subjects. Iris-to-iris matching reaches 91 ROC AUC, indicating left and right irises are correlated; fingerprints reproduce prior findings, and cross-modal matching is only slightly above chance.
2025, IEEE International Conference on Computer Vision Workshops
Gives a quantitative benchmark of Soft Actor-Critic and DDPG agents in the Smart Buildings Control Suite simulator for Building SB1. It looks at learning efficiency, performance against a baseline schedule, and generalization across unseen seasons. Both agents beat the baseline and generalize, and extracting static schedules from policies loses little performance for up to 8-hour bins.
Presents PXRD NET, a diffusion-based generative model trained on 45,229 known structures that solves nanocrystal structures from the chemical formula and a size-broadened powder diffraction pattern. It solves simulated nanocrystals as small as 10 Angstroms across 200 materials, finding verifiable candidates four times out of five with about 7% average error, and also handles noisy experimental patterns.
Applies diffusion models to generating architectural floor plans, adapting the U-Net architecture and adding upscaling techniques that reduce computation while keeping high-resolution output. The authors argue that suitably tuned and conditioned diffusion models capture implicit architectural semantics and improve design efficiency and creativity.
Presents the Smart Buildings Control Suite, an open-source HVAC control benchmark with real telemetry from 11 buildings over 6 years, a lightweight data-driven simulator per building, and a modular physics-informed neural network model as an alternative. It spans varied climates and sizes, is Gym-compatible, and aims to evaluate control policies that scale across buildings.
Presents an ML approach to powder crystallography that estimates unit-cell electron density with a variational coordinate-based deep network, handling ambiguities in orientation, symmetry and resolution. Using simulated powder x-ray diffraction plus partial chemical composition, it reaches up to 93.4% average structural similarity on cubic and trigonal systems.
Challenges the assumption that fingerprints from different fingers of the same person are unrelated. Using deep twin networks to extract fingerprint representations, the authors find strong similarity between fingers of one person with above 99.99% confidence, even controlling for sensor modality. Ridge orientation near the center explains much of it, while minutiae are almost nonpredictive; forensic efficiency could improve substantially.
Presents a variational query-based multi-branch deep network that infers crystal structure end-to-end from degraded powder x-ray diffraction data plus partial chemical composition. Structures are represented as a Cartesian mapped electron density. On simulated cubic and trigonal systems it achieves up to 93.4% average similarity to ground truth on unseen materials.
Classifies maize versus grass and grass-like weed segments using surrounding context, by feeding a CNN a central image plus a border of its surroundings. In a case study on foxtail and yellow nutsedge, context nearly halved error from 7.1% to 4.3%, and one epoch with context outperformed 50 epochs without.
Classifies maize versus grass and grass-like weed segments using surrounding context, by feeding a CNN a central image plus a border of its surroundings. In a case study on foxtail and yellow nutsedge, context nearly halved error from 7.1% to 4.3%, and one epoch with context outperformed 50 epochs without.
Classifies maize versus grass and grass-like weed segments using surrounding context, by feeding a CNN a central image plus a border of its surroundings. In a case study on foxtail and yellow nutsedge, context nearly halved error from 7.1% to 4.3%, and one epoch with context outperformed 50 epochs without.
A deep learning model detects northern leaf blight symptoms in maize from UAV images with 95% accuracy. Transfer learning sped up model optimization, and the method detects symptoms at a very fine spatial scale, offering growers and breeders an alternative to time-consuming visual phenotyping.
Proposes a UAV that detects, tracks and follows other small drones in GPS-denied environments using only onboard computation. A Tiny YOLO detector trained on a 58,647-image dataset, plus simple visual servoing, was validated on a physical platform that followed a target at about 1.5 m/s. Performance was limited by 77% detection accuracy in clutter and eight frames per second.
Describes a two-step approach to millimeter-level plant disease detection in aerial photographs, using deep learning with data labeled via crowdsourcing, since only trained experts can accurately identify symptoms. The text provided is cut off before results, so details beyond the approach are limited.
Trains a Mask R-CNN to segment individual northern leaf blight lesions in UAV images of maize. On a hold-out test set the mean intersection over union was 0.73, with average precision 0.96 at an IOU threshold of 0.50 and 0.61 across thresholds from 0.50 to 0.95. The work supports high-throughput quantitative disease phenotyping.
Describes a public dataset of maize field images annotated for northern leaf blight, captured by hand-held camera, boom-mounted camera and drone. Together the three sets contain 18,222 images with 105,705 annotated lesions, labeled by two experts, making it the largest publicly available image set annotated for a single plant disease.
A pipeline of convolutional neural networks identifies northern leaf blight lesions in field images of maize. Several CNNs classify small image regions, their predictions form heat maps, and a final CNN classifies the whole image, reaching 96.7% accuracy on held-out test images. The authors suggest uses in phenotyping, breeding and targeted pesticide application.
2014, Proceedings of the conference - Association for Computational Linguistics. Meeting
Proposes a probabilistic method for generating code-switched text to aid adult second-language vocabulary learning through reading. Given a bilingual dictionary and English text, a factor graph over lexical mentions optimizes a learnability metric. Algorithms were evaluated with Mechanical Turk subjects using an artificial language and a sentence completion recall task.
Proposes peer-mediated testing for large-scale online education, where students act as both question generators and answerers, crowdsourcing the quiz material and the ranking of participants. A user study compares the resulting rankings with those from expert-curated questions, asking whether students can accurately assess their own knowledge autonomously.
2012, Genetic and Evolutionary Computation Conference (GECCO)
Presents a co-evolutionary framework with pose and predictor populations for inferring kinematic poses from a single depth image, with no prior training and any kinematic structure. On 256 synthetic and 52 real images, it gave consistent poses for 34 and 78 degree-of-freedom models despite self-occlusion, with at least 3.5 times better accuracy and half the computational effort of baselines.
A first step toward implementing the Semantic-Script Theory of Humor computationally. The algorithm mines dual scripts from the ConceptNet semantic network that jointly maximize overlap and incongruity metrics, generating humor beyond simple puns. The metrics are evaluated through user-assessed quality of generated two-liners.
Presents a method for estimating pose from a single depth image for an arbitrary kinematic skeleton without prior training, using an evolutionary algorithm to find the configuration that best explains the image. Results show correct pose estimation of 39 and 78 degree-of-freedom models, even with significant self-occlusion.
Reports a robotic system that paints physical artworks with media such as acrylic on canvas, using a 6-DOF arm and a learning algorithm that determines brushstrokes to transfer a source image. A genetic algorithm used a painting simulation to maximize similarity to the target, and artists can steer results through palette, brush and algorithm parameters.
Essays, editorials, tutorials and educational work: engineering education and children's making, learning technologies, editorials for journals and conferences, and broad perspectives on robotics, AI and creativity.
Explores whether Truss Links, a modular robot platform developed for robot metabolism, can serve as a creative medium for non-experts. A preliminary research-through-design study shows them animating text and 2D and 3D shapes, forming kinetic sculptures, and hand-building a delta pick-and-place robot. The authors discuss what design properties make a research-grade robot accessible for creative use.
A large multi-institution team, including Lipson, proposes applying the common task framework (shared datasets, tasks and evaluation metrics, as used to drive progress in machine learning) to accelerate AI-driven scientific discovery in dynamical systems and other areas of science.
Describes the mission of the NSF AI Institute in Dynamic Systems: developing machine learning tools that find physically interpretable, physics-constrained data-driven models for controlling complex systems, via optimal sensor selection and placement. A common task framework evaluates algorithms across engineering tasks and supports open-source challenge datasets.
A Viewpoint conversation between Lipson and Salah Sukkarieh on how robotics could shape food production and preparation. It covers sensor-rich agricultural robots delivering precision farming and 3D food printers making customized dishes, with the aim of producing more, healthier and more sustainable food using fewer resources and a lower carbon footprint.
A perspective arguing that laser cooking and 3D food printing may complement or replace analog cooking methods such as grills, stoves and microwaves. Software control of ingredient combination and heating could give finer control over nutrient content, enable customized meals, and reduce human handling and foodborne-illness risk. The authors suggest a 3D printer may become a staple home and industrial cooking device.
Argues that multi-ingredient food printing combined with laser cooking could offer convenient, nutritious, low-cost meals, a combination not previously demonstrated. It presents digitally crafted meals as an example of joining software with food and frames software and robotics as the next step in individualized meal production.
A Science Robotics editorial arguing that most robots today operate in isolation under central control. Drawing on the biological transition from individuals to societies and on swarm ideas, it discusses the case for robots that act as collectives that are more robust and capable than single units.
A News and Views commentary arguing that the further ahead a person can mentally simulate their activities, the greater their capacity for self-awareness. It presents robots learning to self-simulate as more than an engineering convenience, calling it the beginning of an era of robot autonomy. The extracted text is fragmentary.
An archival copy of a blog post listing seven commonly held myths in machine learning research as of early 2019. They concern TensorFlow being a tensor library, representativeness of image datasets, test-set use for validation, use of every datapoint in training, the need for normalization in deep residual networks, attention versus convolution, and the robustness of saliency maps.
Makes the case for a portable open-source 3D ultrasound system with AI, aimed at overcoming cost, training and access barriers, particularly for maternal care in developing regions. It weighs arguments for and against an open-source approach, drawing on earlier open-source hardware projects.
A crowd-sourced collection of first-hand anecdotes from researchers in evolutionary computation and artificial life in which evolving algorithms or organisms surprised their creators, exposed bugs, or produced unexpected adaptations. The authors argue such stories deserve study rather than being discarded, and note that no such collection had been published before.
Editorial introducing a special issue drawn from the Artificial Life 14 conference held in Manhattan. It describes the conference program and review process, and outlines the eight extended articles selected, the first four on artificial chemistry and the last four on evolution.
2016, ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)
Proposes a data-driven way to build multiple-choice questions from students' own past submissions. The authors develop a multinomial-logit discrete choice model, derive an objective for selecting optimally discriminative option sets, give an algorithm for a globally optimal solution, and test it with synthetic experiments and a user study, including crowd-sourcing tests from online forums.
Proposes organizing heterogeneous web tutorials, blog posts and lecture notes into a textbook-like path from a learner's prior knowledge to a target topic. Contributions are a supervised classifier for identifying explained and assumed terms in a document, and an algorithm that finds optimal paths through resources given the learner's goal and background.
An editorial recalling the 2000 evolved, 3D-printed robot that walked out of the printer, designed through evolution in simulation and printed whole and preassembled. It poses the grand challenge for additive manufacturing of printing a complete functioning robot.
An editorial on material complexity in 3D printing: hundreds of process parameters affect the properties of printed materials and few are fully characterized. It argues that data mining of the relationship between printer settings and material properties can address this and help engineers trust printed parts.
Reviews selected projects that evolve both neural and morphological subsystems of robots to study how bodies and brains work together to produce adaptive behavior, within the embodied cognition paradigm. The focus is on synergies and tradeoffs between neural architecture, morphology, action and adaptive behavior, with robustness and resilience as themes.
2014, Proceedings of the conference - Association for Computational Linguistics. Meeting
Proposes a probabilistic method for generating code-switched text to aid adult second-language vocabulary learning through reading. Given a bilingual dictionary and English text, a factor graph over lexical mentions optimizes a learnability metric. Algorithms were evaluated with Mechanical Turk subjects using an artificial language and a sentence completion recall task.
An editorial arguing that additive manufacturing is a real industrial revolution. Past revolutions each made a major cost driver, such as power or calculation, nearly vanish, and the editorial applies that pattern to 3D printing despite its market being far smaller than conventional manufacturing.
An editorial arguing that additive manufacturing is driving a revolution in design as well as manufacturing. We can now print nearly any shape we can describe, but CAD software, analysis tools and designers' imagination lag, so design tools must catch up to exploit the technology.
Proposes peer-mediated testing for large-scale online education, where students act as both question generators and answerers, crowdsourcing the quiz material and the ranking of participants. A user study compares the resulting rankings with those from expert-curated questions, asking whether students can accurately assess their own knowledge autonomously.
An editorial on the importance of standards in additive manufacturing, drawing an analogy to printer drivers, PostScript and PDF. It describes the shift from STL to new universal file formats and from proprietary material specifications to ASTM, ISO and NIST standards, arguing that most users need performance standards.
Inaugural editorial introducing the journal 3D Printing and Additive Manufacturing. It outlines the journal's scope, from processes, materials and design tools to medicine, education, bioprinting, food, art, policy, standards and business models, and notes features such as Chinese translation of articles.
Reviews how desktop manufacturing such as 3D printers, die cutters and 3D scanners can advance children's engineering education. It discusses links between virtual designs and physical objects to mathematics concepts, CAD tools that capture design thinking, motor skills, and sharing designs across schools, and notes that classroom best practices are not yet defined.
Outlines an affordable K-12 outreach program based on building pneumatically powered soft robots, made feasible by low-cost 3D printers. The program requires no programming and is positioned as a supplement to kit and competition-based robotics, serving as a reference for students, parents and educators and as a model for similar programs.
A moderated state-of-the-field discussion among experts from rheology, evolutionary robotics, biomimetic robots, soft actuators, flexible electronics and soft robotics, convened by the Soft Robotics editor. Participants comment on their fields and on how interdisciplinary collaboration can advance soft robots; Lipson represents evolutionary robotics.
A magazine essay arguing that designers given unlimited 3D printing capability still produce plain objects, constrained by years of mass-produced forms and conventional CAD tools. It argues that new design tools are needed to exploit the freedom additive manufacturing provides, starting from an anecdote about students' pencil holders.
2011, OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information)
A talk abstract asking whether machines can discover scientific laws automatically. It outlines a research arc from self-reflecting robots to systems that formulate hypotheses, design experiments and interpret results, with examples ranging from classical physics to biology.
Lipson's essay in the National Academy of Engineering's The Bridge (Frontiers of Engineering issue) on where additive manufacturing is heading, from printing passive shapes to printing multi-material, functional and eventually active systems. The linked PDF is the full journal issue, with the essay near the start.
2010, IEEE Aerospace Conference (AERO) · no PDF yet
An IEEE Aerospace Conference paper on robotic and computational systems that discover scientific laws from data on their own, connecting self-modeling robots with automated scientific discovery.
No PDF yet: the file found earlier was a Spanish news web page (RTVE), not a paper; it has been removed.
Describes 3Dprintables.org, an open wiki-style archive of printable teaching models, built on the argument that content drives 3D printing adoption as it did for digital music. Initial models span mechanical engineering, aerospace, biochemistry, mathematics, anatomy and archaeology, and are meant to seed educators sharing models and curricula.
Explores using a freeform fabricator to augment everyday objects for visually impaired users, by printing a tactile alphabet on paper, plastic and metal. The authors tested non-toxic materials for dimensional tolerance, repeatability and adhesion; printing 1 by 1 cm embossed letters took 14 to 52 seconds.
Proposes Universal Desktop Fabrication, a framework combining volumetric digital design software with compact, automated multi-material fabrication hardware to produce almost any complete functional object. The aim is inexpensive, open, easy-to-use systems that shift manufacturing toward personal customization and point-of-use production, with current hardware and software results described.
An essay for the 50 Years of AI festschrift that opens with a robot demonstration to first graders who ask what the robots can actually do. It uses the anecdote to reflect on how hard it is to impress people with machine intelligence and on robots that learn, adapt and create.
A Current Biology commentary on robot-based simulations showing that communication can emerge spontaneously among evolving agents. It notes that deceptive signalling arises in purely competitive settings, while cooperative communication emerges only under group or kin selection, and places these results in the broader history of communication in evolution.
2007, CaltechAUTHORS (California Institute of Technology)
A survey article on modular self-reconfigurable robots, which change shape by rearranging the connectivity of their parts. It outlines a taxonomy of architectures, such as lattice systems, and reviews progress from proof-of-concept systems to elaborate implementations, identifying key challenges and opportunities.
A survey in the Grand Challenges of Robotics series on modular self-reconfigurable robots, which change shape by rearranging the connectivity of their parts. It outlines a taxonomy of architectures, reviews progress over two decades, and identifies key challenges and opportunities.
Talk abstract outlining challenges in moving evolutionary robotics from simulation to physical machines. Topics include evolving morphology and control, hybrid coevolution of reality and simulation, handling noise and uncertainty, and morphological adaptation in hardware, illustrated with a pneumatic nine-legged machine controlled by an open-loop pattern generator.
Overview essay posing open-ended design synthesis as the meta-problem of engineering: designing machines that can design other machines. It motivates evolutionary approaches by noting that nature designed and fabricated its products autonomously, and asks whether computers can augment or replace human invention.
IEEE Spectrum article on fabbers, machines that make useful items on demand from digital designs and could, like personal computers, become available to ordinary people. It draws on Neil Gershenfeld's vision of personal fabrication and notes that researchers at several universities are investigating technologies toward general-purpose fabricators.
Reports using rapid prototyping to document, preserve, reproduce and share historic machines in three dimensions. Several preassembled, functional mechanisms from the Cornell Collection of Reuleaux Kinematic Models were reproduced and placed in an online museum of mechanism, where visitors can download and 3D-print working replicas. The authors expect such physical preservation to become common.
Shows that robotic systems can automatically design robots with complex morphologies and adapted controllers at low cost, using coevolution of bodies and brains in realistic physical simulation and transfer via rapid manufacturing. Reviews three generations: automatically designed LEGO structures, GOLEM truss-based electromechanical systems, and modular designs with a generative DNA-like representation.
Focuses attention on the role of the user interface in geometric modeling, where perceiving and manipulating 3D spatial information is difficult and good interfaces promise large benefits. The paper reviews recent related work and points to future challenges and opportunities.
2000, First International Conference on Axiomatic Design (ICAD2000)
Proposes an architecture for a design knowledge database to support automation of engineering design. Following Axiomatic Design, each step of top-down decomposition is a database element holding modifiable parameters, physical behavior and interfaces. Information is stored in mathematical rather than textual form to reduce ambiguity and make it amenable to fully automated design.
Examines how the Information Age has changed the product engineering cycle and identifies three productivity-impeding forms of overload: detail, constraint and versatility. It suggests remedies: bringing computer tools in earlier, more natural in-context interfaces, and making computers active aids that suggest alternatives. Applications to design, manufacturing, simulation, quality assurance and maintenance are discussed.
![Figure from Modular Self-Reconfigurable Robot Systems [Grand Challenges of Robotics]](images/2007%20-%20Modular%20Self-Reconfigurable%20Robot%20Systems%20%5BGrand%20Challenges%20of%20Robotics%5D.jpg)
![Figure from Homemade [fabrication technology]](images/2005%20-%20Homemade%20%5Bfabrication%20technology%5D.jpg)
Expressive Faces and Social Robots
Robots that interact socially with people: humanoid robotic faces, facial co-expression, lip and gaze generation, communication between robots, and visual behavior modeling of other agents.
Eye, Robot: Learning Expressive Gaze Dynamics via a VLM-Guided Transformer-VAE DOI
Separates robot gaze into target selection by an asynchronous vision-language model, reactive fixation, and a Realistic Gaze Transformer, a Transformer-VAE trained on retargeted markerless human recordings for an 11-DoF robot. A blink predictor triggers eyelid templates. RGT-Base gets the lowest seven-joint MSE on a two-second benchmark; perceived naturalness and open-vocabulary target accuracy are not yet evaluated.
Learning realistic lip motions for humanoid face robots DOI
Presents a humanoid robot face with soft silicone lips driven by a 10-DoF mechanism, and a self-supervised pipeline using a VAE and a facial action transformer to infer lip trajectories directly from speech audio. It outperforms amplitude-based baselines in visual lip-audio coherence and generalizes to 10 languages unseen during training.
Human-robot facial coexpression DOI
Trains a humanoid robot face with 26 degrees of freedom to anticipate a human's upcoming facial expression and execute it simultaneously, rather than mimicking with delay. The robot learns to predict a smile about 839 milliseconds before the human smiles and uses a learned inverse kinematic facial self-model to coexpress it. The authors suggest this could improve human-robot interaction.
Smile Like You Mean It: Driving Animatronic Robotic Face with Learned Models DOI
Presents an animatronic robotic face with soft skin and a vision-based self-supervised framework for facial mimicry, needing no kinematic model, camera calibration or predefined expression set. Splitting learning into a generative model and an inverse model lets it train from one motor babbling dataset, and evaluations show accurate, diverse mimicry across human subjects.
Visual behavior modelling for robotic theory of mind DOI
Asks whether an observer can predict an actor robot's conditional behavior from visual observation alone, without access to its internal states or explicit goals, a notion called Theory of Behavior. In an example, an observer AI watches an actor seek food around obstacles and learns to visualize the actor's likely trajectory in new situations.
Visual Perspective Taking for Opponent Behavior Modeling DOI
Proposes an end-to-end long-term visual prediction framework giving robots Visual Perspective Taking and Theory of Behavior, demonstrated on visual hide-and-seek. Unlike frame-by-frame predictors, the agent predicts directly to multiple future timestamps (25 s), extrapolating 175% beyond its training horizon.
Facially expressive humanoid robotic face DOI
Describes Eva, an inexpensive open-source adult-sized humanoid head for research on emotional communication between humans and machines. It reproduces facial expressions, head movements and speech using 25 muscles, including 12 facial muscles with up to 15 mm skin displacement, and costs about $900 in hardware.
Visual Hide and Seek DOI
Trains embodied agents from scratch to play Visual Hide and Seek, where a prey avoids a predator using only egocentric, partially observed views with occluding obstacles. Experiments suggest the agent learns to predict its own visibility, and agent weaknesses such as slower speed make learning harder but cause more useful features to be learned.
On The Evolutionary Origin of Symbolic Communication DOI
Simulated robots with a single communication channel evolve from non-communicating populations when reproduction requires finding a mate. Indexical communication arose in 99% of experiments, and 9% of populations evolved a symbolic scheme conveying two spatial dimensions over a one-dimensional channel. The results suggest symbolic communication can emerge under natural selection without cognitive preadaptations.
Evolutionary Robotics: Emergence of Communication DOI
A Current Biology commentary on robot-based simulations showing that communication can emerge spontaneously among evolving agents. It notes that deceptive signalling arises in purely competitive settings, while cooperative communication emerges only under group or kin selection, and places these results in the broader history of communication in evolution.
Back to top