PaperlayerView exact revision on arXiv

Source-linked AI summary

Evolution Gym: A Large-Scale Benchmark for Evolving Soft Robots

Jagdeep Singh Bhatia, Holly Jackson, Yunsheng Tian, Jie Xu, Wojciech Matusik

arXiv:2201.09863v1cs.ROcs.LGcs.NE

TL;DR

Co-optimizing robot design and control is challenging, and comprehensive benchmarks for evaluating co-design methods are lacking. Evolution Gym provides a large-scale soft-robot benchmark and baseline algorithms; the tested methods outperform hand-designed robots on easier tasks but do not solve the hardest environments.

  • Problem

    Co-optimizing robot structure and control is challenging, and a comprehensive benchmark for evaluating co-design methods is lacking.

    Source: §1 · p. 2

  • Method

    Evolution Gym benchmarks soft-robot structure and control co-optimization across varied tasks using modular voxel designs and baseline algorithms combining design optimization with reinforcement learning.

    Source: §1 · p. 2

  • Results

    The baseline algorithms produce robots that outperform hand-designed robots on easier tasks, but none succeeds in the hardest environments.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Evolution Gym offers a testbed for comparing co-design methods and studying approaches beyond the tested baselines.

    Source: §6 · p. 10

    Existing methods do not solve the difficult tasks, and the implemented baselines use bilevel optimization with design outermost and control innermost.

    Source: §6 · p. 10

Abstract

from arXiv · show

Both the design and control of a robot play equally important roles in its task performance. However, while optimal control is well studied in the machine learning and robotics community, less attention is placed on finding the optimal robot design. This is mainly because co-optimizing design and control in robotics is characterized as a challenging problem, and more importantly, a comprehensive evaluation benchmark for co-optimization does not exist. In this paper, we propose Evolution Gym, the first large-scale benchmark for co-optimizing the design and control of soft robots. In our benchmark, each robot is composed of different types of voxels (e.g., soft, rigid, actuators), resulting in a modular and expressive robot design space. Our benchmark environments span a wide range of tasks, including locomotion on various types of terrains and manipulation. Furthermore, we develop several robot co-evolution algorithms by combining state-of-the-art design optimization methods and deep reinforcement learning techniques. Evaluating the algorithms on our benchmark platform, we observe robots exhibiting increasingly complex behaviors as evolution progresses, with the best evolved designs solving many of our proposed tasks. Additionally, even though robot designs are evolved autonomously from scratch without prior knowledge, they often grow to resemble existing natural creatures while outperforming hand-designed robots. Nevertheless, all tested algorithms fail to find robots that succeed in our hardest environments. This suggests that more advanced algorithms are required to explore the high-dimensional design space and evolve increasingly intelligent robots -- an area of research in which we hope Evolution Gym will accelerate progress. Our website with code, environments, documentation, and tutorials is available at http://evogym.csail.mit.edu.

1 Introduction

Co-optimizing robot structure and control remains challenging, and comprehensive benchmarks for evaluating these methods are lacking. Evolution Gym addresses this gap with a large-scale soft-robot benchmark and baseline algorithms whose results show gains on easier tasks but failures in the hardest environments.

1 Introduction

  • Earlier work often focused on control for fixed robot structures or co-optimization on only a few simple tasks.The cited examples concern control algorithms for existing structures [1] [19] and co-optimization for tasks such as running and jumping [16].Source: §1 · p. 1
  • The benchmark gap and the complexity of bilevel co-design, including long inner control-optimization loops, are identified as reasons the area remains under-explored.Source: §1 · p. 2
  • Evolution Gym provides a large-scale benchmark for co-optimizing soft-robot structure and control across more than 30 tasks.Its tasks span locomotion on varied terrain and manipulation; robots use soft, rigid, and actuator voxels.Source: §1 · p. 2
  • Baseline algorithms evolve robots that outperform hand-designed robots on easier tasks, but none completes the hardest environments.The authors present these results as evidence for the value of joint structure-and-control optimization and the need for more capable methods.Source: §1 · p. 2
  • The paper combines design optimization methods with reinforcement learning to develop and evaluate several co-design baselines.Source: §1 · p. 2

2 Related work

Prior robot co-design work explored graph-based rigid robots and voxel-based soft robots, but faced limits in compliant design, control, or task breadth. Existing robotics-learning benchmarks focus on control, leaving co-design studies to build or modify testbeds and often evaluate only a few tasks.

2 Related work

  • Rigid-robot co-design represents structure as graphs, but rigid-only components can prevent feasible designs for tasks needing compliance.Early evolutionary graph design is followed by graph-search approaches [15] [31].Source: §2 · p. 2
  • Voxel-based soft-robot methods co-optimize structure and control, but periodic open-loop actuation limits learning complex non-periodic tasks.Other work optimizes material parameters and neural-network policies while keeping robot shape fixed.Source: §2 · p. 3
  • Existing robotics-learning benchmarks focus on control, so co-design studies often build their own testbeds and cover only a few tasks.That independent testbed development also leads to indirect comparisons among algorithms; Evolution Gym instead spans locomotion and manipulation at varying difficulty levels.Source: §2 · p. 3

3 Evolution Gym

Evolution Gym is a benchmark for co-designing voxel-based soft robots, combining a task-specific environment with a soft-body simulator and interfaces for design and control optimization. Its modular representation, task suite, and simulation support evaluation across varied robot forms and tasks.

3.1 Overview

  • Evolution Gym connects a task-specific environment and soft-body simulator to user-defined co-design algorithms that propose structures and optimize their controllers.The design optimizer receives the maximum reward achieved for each proposed structure.Source: Figure 1 · p. 3, §3.1 · p. 4

3.2 Multi-material voxel-based representation

  • A small set of voxel types creates a modular design space with diverse morphologies, while mass-spring simulation supports efficient training.The authors report that robots with fewer than 100 voxels can perform complex motions and that training can take minutes.Source: §3.2 · p. 4

3.3 Task representation

  • The task interface specifies robot structure, observations, actions, and task-specific rewards.A robot is encoded by a voxel material matrix and a list of adjacent-voxel connections; observations include robot state, terrain information when relevant, and task-specific goal information.Source: §3.3 · pp. 4–5

3.4 Simulation engine

  • The C++ simulator models voxelized objects and environments as a mass-spring system, with penalty-based contact and friction forces.It advances the simulation using symplectic RK-4 integration.Source: §3.4 · p. 5

3.5 Benchmark environment suite

  • The benchmark has over 30 tasks across locomotion and manipulation, with difficulty levels based on baseline performance, and users can add tasks.Figure 2 presents selected examples from the suite.Source: §3.5 · p. 5, Figure 2 · p. 5

3.5.1 Locomotion tasks

  • Locomotion tasks range from flat-ground walking and rope-bridge traversal to climbing stairs or walls and crossing a pit.Source: §3.5.1 · pp. 5–6

3.5.2 Object manipulation tasks

  • Manipulation tasks include carrying or throwing an object, sliding a beam, catching a falling object, and lifting an object from a hole.Source: §3.5.2 · p. 6

4 Evolving soft robots

Robot evolution is framed as a two-level optimization process: design methods propose physical structures, and reinforcement learning optimizes a controller for each design. The section compares three design optimizers under physical constraints and uses PPO for control optimization.

4 Evolving soft robots

  • Design optimization proposes robot structures in an outer loop, while control optimization trains a controller for each structure in an inner loop.The algorithm evaluates design-controller pairs and uses their rewards to update the design population.Source: §4 · p. 6

4.1 Design optimization

  • Design search maximizes reward while requiring each robot body to be connected and to contain actuators.The section evaluates three instantiations of the design-optimization step.Source: §4.1 · p. 6
  • The genetic algorithm keeps top designs as survivors and mutates them, including voxel changes that can alter body topology; it does not use crossover.The survivor fraction decreases from 60% to 0% over generations, and each voxel has a 10% probability of changing during mutation.Source: §4.1 · pp. 6–7
  • Bayesian optimization uses Gaussian-process surrogates and batch Thompson sampling, with batch size matched to the other population-based methods.L-BFGS optimizes the acquisition function.Source: §4.1 · p. 7
  • CPPN-NEAT maps each voxel’s spatial coordinates to its type and evolves the network that generates those mappings.NEAT applies mutation, crossover, and selection to the CPPN structure.Source: §4.1 · p. 7

4.2 Control optimization

  • PPO trains controllers with reinforcement learning, enabling non-periodic control patterns beyond the periodic actuation used in prior approaches.The paper motivates this choice by noting that periodic control prevents learning complex non-periodic tasks such as walking on uneven or varying terrain.Source: §4.2 · p. 7

5 Experiments and results

Across the selected benchmark tasks, GA performs best overall, though no algorithm leads on every task. The experiments also show evolved designs developing task-specific structures, outperforming hand-designed robots on each task in at least one algorithm, while leaving the hardest settings unsolved.

5 Experiments and results

  • The baseline evaluation combines GA, BO, or CPPN-NEAT for design optimization with PPO for control and reports results on 10 selected tasks.The paper directs readers to Appendix E for complete results on all benchmark tasks.Source: §5 · p. 7

5.1 Comparisons among baseline algorithms

  • GA outperforms the other baselines overall across the 10 selected tasks, although no single algorithm performs best on every task.Figure 3 plots each algorithm’s best robot performance against evaluations, averaging curves across six random seeds with variance shaded.Source: §5.1 · p. 7, Figure 3 · p. 8
  • CPPN-NEAT generally performs well on locomotion but poorly on complex manipulation, while BO performs poorly on most tasks amid high-dimensional categorical search and noisy RL evaluations.The authors suggest NEAT’s preference for simpler CPPNs yields regular patterns, whereas complex manipulation may require agile substructures in irregular designs.Source: §5.1 · pp. 7–8

5.2 Evolution analysis

  • Across Carrier, Lifter, and Bridge Walker, evolved designs develop task-specific structures, including object-holding legs, a gripper-like shape, and a large front foot.Carrier survivors become faster while retaining the object; the Bridge Walker’s large front foot increases surface area and friction for crossing the soft rope bridge.Source: §5.2 · pp. 8–9, Figure 4 · p. 8

5.3 Comparison against hand-designed robots

  • On every tested task, at least one algorithm outperforms the hand-designed robots, though a natural hand-designed Carrier performs almost as well as the best optimized designs.For Climber, GA develops legs and a flat back that balance forward progress with wall contact and friction.Source: §5.3 · p. 9, Figure 5 · p. 9
  • Beam Slider remains a hard case: no tested hand-designed or algorithm-produced robot achieves satisfying performance, and the GA design falls short of optimal behavior.The authors conclude that further work is needed on co-optimization algorithms for these hard tasks.Source: §5.3 · p. 9

6 Conclusion and future work

Evolution Gym supports systematic study of soft-robot co-design, with robots evolved from scratch accomplishing complex tasks. The authors identify unresolved algorithmic and task-scope challenges and outline directions for future work.

6 Conclusion and future work

  • Evolution Gym enabled study of current co-design algorithms across diverse tasks, and robots evolved from scratch could accomplish complex tasks.The authors also report discovering limitations in existing techniques for evolving more intelligent embodied systems.Source: §6 · p. 10
  • Existing methods cannot address some difficult tasks, motivating more advanced co-design algorithms; the framework is not tied to the baselines’ nested optimization routine.Proposed alternatives include concurrent co-optimization, neuroevolution, morphogenetic development, gradient-based design optimization, and decentralized controllers.Source: §6 · p. 10
  • The authors suggest studying multi-task or multi-objective co-design because the benchmark provides a comprehensive set of tasks.They regard a robot as more successful if it can perform multiple tasks.Source: §6 · p. 10
  • The baseline algorithms use direct encoding and CPPN, while exploring other morphological representations remains future work.The authors note that more intelligent encodings could lead to better performance.Source: §6 · p. 10
  • The authors propose extending the benchmark beyond its current locomotion and manipulation tasks to categories such as flying and swimming.Such extensions would require incorporating new simulation capabilities.Source: §6 · p. 10
  • The authors view Evolution Gym as a flexible test bed that could support future improvements to co-design methods and robot intelligence.They describe the benchmark as filling a missing piece in soft robotics and robotic evolution research.Source: §6 · p. 10

Societal Impact

The authors describe the work as preliminary and identify potential ethical risks from large-scale real-world application. The simulator represents objects as connected voxels converted into a mass-spring system, with dynamics, contact handling, and additional techniques.

Societal Impact

  • Large-scale real-world application could raise ethical issues around automatically designing biological creatures, while user-defined pernicious rewards could encourage biased results.The authors characterize the work as preliminary and say they are still far from causing societal harm.Source: Societal Impact · p. 10

A Simulation engine

  • The simulator section describes object representation, dynamics, contact forces, quality-improvement techniques, scaling analysis, and implementation hyperparameters.Source: A Simulation engine · p. 14

A.1 Representation

  • Objects are initialized as connected, non-overlapping voxels; a material matrix assigns voxel types and adjacency connections determine object identity.The simulation converts each voxel into a cross-braced square of point masses and springs, with spring constants varying by material.Source: §A.1 · p. 14

A.2 Simulation Dynamics

  • The simulator advances positions and velocities using symplectic RK-4 while accounting for gravitational, contact, viscous-drag, and spring forces.Control runs at 50 Hz, with 30 simulation sub-steps at 1500 Hz per control input to improve stability.Source: §A.2 · p. 14

A.3 Collision detection and contact forces

  • Collision handling uses surface-voxel bounding-box trees and penetration-based penalty and friction forces, with contact forces held constant during each RK-4 step.The contact forces are precomputed because their computation is often expensive.Source: §A.3 · p. 15

A.4 Other techniques

  • Strain limiting reduces spring deformation, self-folding triggers a one-time −3 reward and reset, and delayed actuation favors longer, smoother motions.Strain limiting is not a hard constraint and can be overcome by sufficiently strong actuation; rigid-cell springs use a 3% threshold.Source: §A.4 · p. 15

A.5 Scalability and speed analysis

The simulator’s scaling is evaluated alongside its hyperparameters, benchmark suite, notation, and three locomotion tasks. The suite’s task definitions specify each robot’s objective and reward conditions.

A.5 Scalability and speed analysis

  • Hundreds of simulation steps per second and parallel training in minutes distinguish the framework from 3D FEM simulations, while its speed is slower than rigid-body simulators and comparable to other 2D voxel-based simulations.Figure 7 graphs simulation steps per second and that rate multiplied by voxel count; the FEM comparison describes a single robot trained on a 4-core machine.Source: §A.5 · pp. 15–16, Figure 7 · p. 16

A.6 Hyperparameters

  • The simulation hyperparameters govern point-mass behavior, contact forces, and spring rigidity, with spring constants differentiated by voxel material.The listed categories include mass, viscous drag, gravity, contact stiffness and friction, and rigid, soft, and actuator spring constants.Source: §A.6 · p. 17

B Full benchmark suite

  • Evolution Gym implements 32 tasks, with ten named reference tasks spanning locomotion and object manipulation.The listed reference tasks are Walker-v0, BridgeWalker-v0, UpStepper-v0, Traverser-v0, Climber-v0, Carrier-v0, Thrower-v0, Catcher-v0, Lifter-v0, and BeamSlider-v0.Source: B Full benchmark suite · p. 17

B.1 Notation

  • The notation defines object position, velocity, and orientation from point masses, alongside relative point-mass coordinates and local terrain elevations.Position and velocity are centers of mass; orientation averages point-mass angles weighted by their initial distances from the center of mass.Source: §B.1 · p. 17

B.2.1 Walker-v0

  • Walker-v0 rewards positive x-direction travel on flat terrain and adds a one-time reward for reaching its end.The task is described as easy and runs for 500 steps.Source: §B.2.1 · p. 18

B.2.2 BridgeWalker-0

  • BridgeWalker-v0 rewards positive x-direction travel across a soft rope bridge and gives a one-time reward for reaching the terrain’s end.The task is described as easy.Source: §B.2.2 · p. 18

B.2.3 BidirectionalWalker-v0

  • BidirectionalWalker-v0 rewards movement toward a randomized x-position goal that changes during the task.The task is described as medium and runs for 1000 steps.Source: §B.2.3 · p. 19

B.3 Object manipulation tasks

The section surveys soft-robot tasks in object manipulation, climbing, and terrain traversal, spanning challenges that vary in difficulty and involve diverse objects, obstacles, and terrain. The terrain-focused tasks reward forward movement and use task-specific completion or failure conditions.

B.3 Object manipulation tasks

  • Carrier-v0 has the robot catch a box and carry it as far as possible, while harder Carrier-v1 requires placing the box on a table with terrain-dependent height constraints.Carrier-v0 is described as easy and Carrier-v1 as hard.Source: §B.3 · pp. 19–20

B.3.3 Pusher-v0 / B.3.4 Pusher-v1

  • Pusher-v0 has the robot push a box initialized in front of it; Pusher-v1 is medium difficulty and requires pushing or dragging a box initialized behind it forward.Pusher-v0 is easy; Pusher-v1's reward penalizes robot-box separation in the x-direction.Source: §B.3.3 · p. 20, §B.3.4 · p. 21

B.3.4 Pusher-v1

  • Thrower-v0, a medium-difficulty task, asks the robot to throw a box initialized on top of it while limiting how far the robot moves from x = 0.The reward favors positive x movement by the box and penalizes excessive robot movement from x = 0.Source: §B.3.4 · p. 21

B.3.6 Catcher-v0

  • Catcher-v0 is a hard task in which the robot catches a fast-moving, rotating box, with penalties for letting it fall below a threshold height.The environment runs for 400 steps.Source: §B.3 · p. 20, §B.3.6 · p. 22

B.3.7 BeamToppler-v0 / B.3.8 BeamSlider-v0

  • BeamToppler-v0 has the robot knock over a beam on two pegs from underneath, whereas hard BeamSlider-v0 requires sliding a beam across a line of pegs from underneath.Both tasks reward moving the robot toward the beam; toppling also rewards moving the beam and making it fall.Source: §B.3.7 · pp. 22–23, §B.3.8 · p. 23

B.3.9 Lifter-v0

  • Hard Lifter-v0 requires lifting a box out of a hole while keeping it near a goal x-position and avoiding a fall below a threshold height.The goal-position penalty encourages lifting the box straight up.Source: §B.3.9 · pp. 23–24

B.4.1 Climber-v0 / B.4.2 Climber-v1 / B.4.3 Climber-v2

  • Climber-v0 climbs through a flat vertical channel, Climber-v1 through a vertical channel of mixed rigid and soft materials, and hard Climber-v2 through a narrow stepwise channel.All three reward upward movement; Climber-v2 also rewards positive x movement.Source: §B.4.1 · p. 24, §B.4.2 · p. 25, §B.4.3 · p. 26

B.5.1 UpStepper-v0 / B.5.2 DownStepper-v0

  • UpStepper-v0 climbs stairs of varying lengths, DownStepper-v0 descends them, and ObstacleTraverser-v0 crosses increasingly bumpy terrain.The tasks are respectively medium, easy, and medium difficulty; the stepper tasks also penalize excessive rotation and reward reaching the terrain's end.Source: §B.5.1 · p. 26, §B.5.2 · p. 27

B.5.4 ObstacleTraverser-v1

  • The passages describe hard tasks involving very bumpy terrain, tall obstacles, and floating platforms at different heights; snippets also specify positive-x movement rewards and outcomes that trigger resets.One passage gives a one-time reward of 2 for reaching the terrain’s end before reset. Other passages describe a one-time penalty of −3 for rotation more than 90 degrees from the original orientation in either direction, or for falling off platforms, before reset.Source: §B.5.4 · p. 28

B.5.7 GapJumper-v0

  • GapJumper requires traversing spaced-out floating platforms at the same height, with falling off penalized.The robot is rewarded for moving in the positive x-direction, receives a one-time penalty of −3 for falling off, and the environment runs for 1000 steps.Source: §B.5.7 · p. 29

B.5.8 Traverser-v0

  • Traverser challenges the robot to cross a pit of rigid blocks without sinking, rewarding forward progress and end-of-terrain completion.The robot receives a one-time reward of 2 for reaching the terrain’s end, and the environment runs for 1000 steps.Source: §B.5.8 · p. 29

B.5.9 CaveCrawler-v0

  • CaveCrawler is a medium-difficulty task requiring the robot to squeeze through caves and low-hanging obstacles.It rewards movement in the positive x-direction and gives a one-time reward of 1 for reaching the terrain’s end; the environment runs for 1000 steps.Source: §B.5.9 · p. 30

B.6 Shape change tasks

The shape-change tasks reward robots for changing their area, width, or height. Their task descriptions distinguish easy expansion objectives from medium difficulty shrinking and height objectives.

B.6 Shape change tasks

  • AreaMaximizer-v0 rewards the robot for occupying the largest possible surface area.The task is described as easy.Source: §B.6 · p. 30
  • AreaMinimizer-v0 rewards the robot for shrinking to occupy the smallest possible surface area.The reward is based on the convex hull area of the robot’s point masses; the task is described as medium.Source: §B.6 · p. 30

B.6.3 WingspanMaximizer-v0

  • WingspanMaximizer-v0 rewards growth in the x-direction to make the robot as wide as possible.The task is described as easy.Source: §B.6.3 · p. 31
  • HeightMaximizer-v0 rewards growth in the y-direction to make the robot as tall as possible.The task is described as medium and runs for 500 steps.Source: §B.6.3 · p. 31, §B.7.2 · p. 32

B.7 Miscellaneous tasks

The miscellaneous tasks test flipping, jumping, and balancing, with balancing objectives differing in whether the robot starts on or beside the pole. The task descriptions state the rewarded motion and, for two tasks, the episode length.

B.7 Miscellaneous tasks

  • Flipper rewards repeated counter-clockwise rotations on flat terrain.The task is described as easy and runs for 600 steps.Source: §B.7 · p. 32

B.7.2 Jumper-v0

  • Jumper rewards jumping upward in place on flat terrain and penalizes horizontal motion.The task is described as easy and runs for 500 steps.Source: §B.7.2 · p. 32

B.7.3 Balancer-v0

  • Balancer-v0 rewards movement toward a goal atop a thin pole, where the robot starts already on the pole.The task is described as easy.Source: §B.7.3 · p. 32

B.7.4 Balancer-v1

  • Balancer-v1 rewards movement toward a goal atop a thin pole after the robot jumps onto it from beside the pole.The task is described as medium and runs for 600 steps.Source: §B.7.4 · p. 33

C Optimization methods

This section describes three design-optimization methods used in the co-design algorithms, alongside the general experiment settings and PPO control hyperparameters. The methods range from voxel-level genetic mutation to surrogate-guided search and evolving network-encoded designs.

C.1 Genetic algorithm (GA)

  • GA retains top-performing designs and mutates voxels; its survivor fraction decreases as generations progress, with a minimum of two survivors.The algorithm uses elitism, samples survivors as parents, and changes each voxel with 10% probability.Source: §C.1 · p. 33

C.2 Bayesian optimization (BO)

  • Bayesian optimization fits a Gaussian-process surrogate and uses expected improvement to propose designs, with batch Thompson sampling and L-BFGS for acquisition optimization.The implementation uses a Matérn 5/2 kernel.Source: §C.2 · p. 34

C.3 CPPN-NEAT

  • CPPN-NEAT encodes robot designs with a CPPN and evolves its network structure using NEAT mutation, crossover, and selection.The paper uses standard CPPN and NEAT implementations.Source: §C.3 · p. 34

D.1 General hyperparameters

  • The experiments generally use a 5 × 5 design grid, increasing to 5 × 7 for Lifter, and compare algorithms by unique robots trained, up to 750 evaluations.Each robot receives 512,000 total RL steps for task evaluation.Source: §D.1 · p. 35

D.2 Design optimization hyperparameters

  • Algorithm-specific design-optimization settings are tabulated; BO uses default GPyOpt settings, while CPPN-NEAT settings are documented in the accompanying configuration reference.The BO passage explains the surrogate kernel and L-BFGS optimization settings.Source: §D.2.2 · p. 35, §D.2.3 · p. 35, Table 3 · p. 35

D.3 Control optimization hyperparameters

  • PPO uses Generalized Advantage Estimation optionally, Adam with linear learning-rate decay, and parallel processes that sample a specified number of steps per iteration.Other PPO parameters use their defaults.Source: §D.3 · p. 36, Table 6 · p. 36

E Complete experiment results

Across the benchmark, GA outperformed the other baselines on most of the ten highlighted tasks, while the full 32-task suite includes both successful designs and environments where evolved robots remain incomplete. The results show task-specific forms and behaviors, alongside concrete failure boundaries.

E Complete experiment results

  • GA outperformed the other baseline algorithms on most of the ten highlighted tasks, and its reward curves are reported across all 32 benchmark tasks.Figure 40 plots the best performance against evaluations, averaging curves over three random seeds with variance shaded.Source: E Complete experiment results · pp. 36–37, Figure 40 · p. 38
  • Balancer robots evolved actuated arm-like counterweights, while CaveCrawler robots converged on short, horizontally actuated snake-like forms.Figure 41 shows top-performing robots from three generations for the selected tasks.Source: Hurdler-v0 · p. 37, Figure 41 · p. 39
  • PlatformJumper designs evolved multiple actuators oriented to spring between platforms at different heights.The optimized designs nevertheless can get stuck in a gap between platforms.Source: Hurdler-v0 · p. 37, Figure 41 · p. 39
  • DownStepper robots evolved bipedal forms with a horizontal actuator on one foot and a vertical actuator on the other for stair traversal.Figure 42 shows step-by-step sequences of optimized designs.Source: Hurdler-v0 · p. 37, Figure 42 · p. 39
  • BeamToppler produced two surviving designs with hand-like arms; one also evolved a hook-like gripper to push the beam from above.The other design pushes the beam from underneath.Source: Hurdler-v0 · p. 37, Figure 41 · p. 39
  • Near-optimal CaveCrawler robots often fail at the final low obstacle, while Hurdler designs can be caught by a wider barrier.Some CaveCrawler designs clear every section, but the baseline algorithms do not produce fully successful robots in some harder tasks.Source: Hurdler-v0 · p. 37
Loading 2201.09863v1…