Abstract:The automatic design of robots has existed for 30 years but has been constricted by serial non-differentiable design evaluations, premature convergence to simple bodies or clumsy behaviors, and a lack of sim2real transfer to physical machines. Thus, here we employ massively-parallel differentiable simulations to rapidly and simultaneously optimize individual neural control of behavior across a large population of candidate body plans and return a fitness score for each design based on the performance of its fully optimized behavior. Non-differentiable changes to the mechanical structure of each robot in the population -- mutations that rearrange, combine, add, or remove body parts -- were applied by a genetic algorithm in an outer loop of search, generating a continuous flow of novel morphologies with highly-coordinated and graceful behaviors honed by gradient descent. This enabled the exploration of several orders-of-magnitude more designs than all previous methods, despite the fact that robots here have the potential to be much more complex, in terms of number of independent motors, than those in prior studies. We found that evolution reliably produces ``increasingly differentiable'' robots: body plans that smooth the loss landscape in which learning operates and thereby provide better training paths toward performant behaviors. Finally, one of the highly differentiable morphologies discovered in simulation was realized as a physical robot and shown to retain its optimized behavior. This provides a cyberphysical platform to investigate the relationship between evolution and learning in biological systems and broadens our understanding of how a robot's physical structure can influence the ability to train policies for it. Videos and code at

DiffuseBot: Breeding Soft Robots With Physics-Augmented Generative Diffusion Models

Creation of Novel Soft Robot Designs using Generative AI

Scalable sim-to-real transfer of soft robot designs

A Soft Artificial Muscle Driven Robot with Reinforcement Learning

A Morphological Transfer-Based Multi-Fidelity Evolutionary Algorithm for Soft Robot Design

Evolution and learning in differentiable robots

From Bioinspiration to Computer Generation: Developments in Autonomous Soft Robot Design

DiffGen: Robot Demonstration Generation via Differentiable Physics Simulation, Differentiable Rendering, and Vision-Language Model

SoftZoo: A Soft Robot Co-design Benchmark For Locomotion In Diverse Environments

The Ingredients for Robotic Diffusion Transformers

Robot Shape and Location Retention in Video Generation Using Diffusion Models

Soft Robots Learn to Crawl: Jointly Optimizing Design and Control with Sim-to-Real Transfer

Dynamic simulation of articulated soft robots

Efficient and Diverse Generative Robot Designs using Evolution and Intrinsic Motivation

Robots As Models of Evolving Systems.

Reality-assisted evolution of soft robots through large-scale physical experimentation: a review

Multi-environment robotic transitions through adaptive morphogenesis

Evolution-based Shape and Behavior Co-design of Virtual Agents

Physics-Guided Deep Learning Enabled Surrogate Modeling for Pneumatic Soft Robots

Efficient automatic design of robots