Source-linked AI summary
A soft robot that adapts to environments through shape change
Dylan S. Shah, Joshua P. Powers, Liana G. Tilton, Sam Kriegman, Josh Bongard, Rebecca Kramer-Bottiglio
TL;DR
The paper addresses robots’ inability to actively switch body shape and progress toward dynamically adapting resting shape. It demonstrates a simulation-to-reality pipeline for shape-changing locomotion and reports faster traversal than an equivalent non-morphing robot, with shape change viable and physically realizable.
Problem
Robots could not actively switch body shape, while dynamically adapting resting shape remained an emerging capability.
Method
The paper demonstrates a simulation-to-reality pipeline for designing and realizing shape-changing robots across representative tasks and environments.
Results
The robot traversed two test environments faster than an equivalent non-morphing robot, and shape change was viable and physically realizable.
Takeaways & Limitations
The work represents progress toward automated simulation-to-reality transfer and closed-loop automated design of robots.
Takeaways & Limitations
Simulation generated some designs that did not transfer to physical hardware, constraining current simulation-to-reality reliability.
Abstract
from arXiv · showhide
Many organisms, including various species of spiders and caterpillars, change their shape to switch gaits and adapt to different environments. Recent technological advances, ranging from stretchable circuits to highly deformable soft robots, have begun to make shape-changing robots a possibility. However, it is currently unclear how and when shape change should occur, and what capabilities could be gained, leading to a wide range of unsolved design and control problems. To begin addressing these questions, here we simulate, design, and build a soft robot that utilizes shape change to achieve locomotion over both a flat and inclined surface. Modeling this robot in simulation, we explore its capabilities in two environments and demonstrate the existence of environment-specific shapes and gaits that successfully transfer to the physical hardware. We found that the shape-changing robot traverses these environments better than an equivalent but non-morphing robot, in simulation and reality.
1 Introduction
Shape-changing soft robots address the unresolved challenge of adapting morphology, gait, and control across environments. This study combines simulation and physical hardware to search for environment-specific shape–behavior strategies and transfer them to a robot.
- Research gap: Prior robots mainly adapted to multiple environments through control, while their resting component dimensions remained constant.Such robots could not actively switch between forms optimized for distinct locomotion modes.
- Research gap: Shape-changing robot design and control remain manually intensive, limiting systems to shape–controller pairings that human intuition can design.Non-intuitive pairings may improve task performance in specific environments.
- Research gap: Manufacturing is more time-consuming and expensive than simulation, yet transferring viable shape–behavior pairs from simulation to physical hardware remains challenging.The paper identifies simulation-to-reality transfer as a central obstacle for shape-changing robots.
- Approach: The study tests whether shape change can improve average locomotion speed across flat and inclined environments more effectively than control adaptations.The robot actively changes shape using an internal bladder while external inflatable bladders provide locomotion.
- Approach: An iterative hill-climbing algorithm automatically generated and modified multiple robot shapes and controllers in a soft multi-material simulator.The search aimed to discover new locomotion strategies and environment-specific shape–controller pairs.
- Findings: No single shape–controller pair efficiently locomoted in both environments, but relatively small shape changes paired with policy adaptations enabled locomotion in both.The fastest strategy inflated and rolled on flat or slightly inclined surfaces, while flattening and inchworm locomotion improved performance above a critical slope angle.
- Findings: The physical robot reproduced two environmentally effective gaits, demonstrating that shape change is physically realistic and that some simulated shape–behavior strategies transfer to reality.The work therefore advances an automated sim2real pipeline for metamorphosing soft robots operating in different environments.
2 Results
The study used automated simulation searches to identify environment-specific shapes, orientations, and control policies for locomotion on flat and inclined surfaces. Shape change enabled distinct gaits and improved combined performance across both environments.
- Automated search: Automated searches optimized shape, orientation, and control policy separately or jointly for flat-ground and inclined-surface locomotion under identical computational resources.The three parameter sets were shape, orientation relative to equal-elevation contours, and control policy.
- Fixed parameters: With fixed shape and orientation, optimization found rolling on flat ground at 0.1576 BL/s but failed uphill at -0.0023 BL/s; manual inchworm motion reached 0.0246 BL/s uphill.The fixed-shape experiments used an inflated cylinder on flat ground and a deflated shape on the incline.
- Joint optimization: Joint optimization found an angled inflated shape that shuffled uphill at 0.0158 BL/s and achieved combined maximum velocity of 0.1023 BL/s, versus 0.0998 BL/s for hand-designed strategies.The hand-designed comparison combined rolling on flat ground with crawling on inclines.
- Cross-environment performance: The best evolved shape-controller pair reached 0.1888 BL/s on flat ground and 0.0158 BL/s uphill, averaging 0.1023 BL/s across environments.The corresponding fixed-shape hand-designed averages were -0.1995 BL/s for the round rolling shape and 0.0424 BL/s for the flat inchworm shape.
- Physical transfer: Physical tests transferred simulated gaits, allowing the robot to roll on flat terrain and flatten to climb moderate inclines up to approximately 14° at 0.008 BL/s.The physical robot rolled at 0.05 BL/s on flat terrain before switching to a flattened shape uphill.
- Interpretation: Shape-changing robots accessed multiple locomotion modes by switching shapes, while evolutionary policies were less finely tuned and sometimes included counterproductive or superfluous actuations.The authors suggest that broader design freedom may help automated searches avoid local minima.
3 Discussion
The study shows that shape change can improve locomotion across environmental transitions, with environment-specific shapes and gaits transferring from simulation to physical hardware. It also identifies search and sim2real limitations that constrain broader automated design.
- Shape-changing robots traversed two test environments faster than an equivalent non-morphing robot.
- The physical robot rolled when round on flat ground and flattened to use an inchworm-like gait on inclines up to ∼14 degrees.
- The pipeline searched for environment-specific shapes and control policies, then transferred effective pairs to physical hardware.
- The search struggled to discover a successful inchworm gait, likely because the policy space lacked a useful gradient toward inching.
- Some simulated behaviors were non-transferable because inaccurate parameters could produce unrealistic motions such as sliding over the ground.
- Future progress requires improved search, sensing, and iterative sim2real methods to design robots that operate across more challenging environmental changes.
4 Methods
The study simulates and fabricates a shape-changing soft robot, modeling voxel-based mechanics, inflation, friction, and actuation while optimizing shape, orientation, and control sequences for flat and inclined locomotion.
- Simulation environment: The robot is modeled in Voxelyze as cubic voxels connected by Euler-Bernoulli beams, with motion produced by external forces or voxel expansion.Adjacent voxels are represented as points connected by beams, and voxel expansion can occur along one or more dimensions.
- Simulation-to-reality transfer: Simulation parameters were iteratively adjusted so simulated and physical robots using the same controllers exhibited similar behavior on flat and inclined surfaces.The calibration included material parameters, pressure-related behavior, and friction settings.
- Simulation environment: Slope locomotion is simulated by changing gravity direction rather than rotating the floor, while shape change is represented by pressure-like outward forces.The maximum simulated pressure was 12 kPa.
- Actuation: The actuation sequence is a binary 10×T matrix: eight rows control inflatable bladders, and two control variable-friction feet.A 1 denotes bladder expansion and 0 denotes deflation; each column represents one discrete actuation step.
- Physical robot: The physical robot uses silicone bladders, broadcloth, variable-friction feet, and different silicone stiffnesses to balance force output, flexibility, and transferability.DS30 was used for higher outer-actuator bending force, while DS10 was used elsewhere to retain flexibility.
Code Availability
The study’s soft-robot simulation code is publicly available in a GitHub repository.
- Code availability: The repository contains the code necessary to reproduce the soft-robot simulations.The repository is identified as github.com/jpp46/NATURE MI2020.
Supplementary Material
Supplementary materials include a movie and figures showing the simulated robot and the manufacturing process for the physical robot.
- Supplementary material: Movie S1 documents a soft robot that adapts to environments through shape change.
- Supplementary material: Figure S1 shows the simulated robot switching between round and flat shapes using a shape-changing lattice.The figure also depicts voxel-beam connections and pressure vectors acting on interior voxels.
- Supplementary material: Figure S2 illustrates fabrication of silicone bladders, variable-friction feet, and assembly of the physical robot.The schematics include simplified cross-sections of the robot.