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

Scalable Co-Optimization of Morphology and Control in Embodied Machines

Investigating Premature Convergence in Co-optimization of Morphology and Control in Evolved Virtual Soft Robots

Modular Controllers Facilitate the Co-Optimization of Morphology and Control in Soft Robots

The dynamic of body and brain co-evolution

Deep Reinforcement Learning Based Co-Optimization of Morphology and Gait for Small-Scale Legged Robot

Environmental Adaptation of Robot Morphology and Control Through Real-World Evolution

Evaluation of Frameworks That Combine Evolution and Learning to Design Robots in Complex Morphological Spaces

Brains, Bodies, and Beyond: Competitive Co-Evolution of Robot Controllers, Morphologies and Environments

Efficient and Diverse Generative Robot Designs using Evolution and Intrinsic Motivation

An Investigation of the Factors Influencing Evolutionary Dynamics in the Joint Evolution of Robot Body and Control

Competitive Co-evolution of Predator and Prey Sensory-Motor Systems

The Role of Morphological Variation in Evolutionary Robotics: Maximizing Performance and Robustness

N-LIMB: Neural Limb Optimization for Efficient Morphological Design

Multi-embodiment Legged Robot Control As a Sequence Modeling Problem

Simultaneous Optimization of Discrete and Continuous Parameters Defining a Robot Morphology and Controller

Evolution and learning in differentiable robots

Morphological Development at the Evolutionary Timescale: Robotic Developmental Evolution

Embodied intelligence via learning and evolution

On Designing a Learning Robot: Improving Morphology for Enhanced Task Performance and Learning

Artificial evolution of robot bodies and control: on the interaction between evolution, learning and culture

Data-efficient Co-Adaptation of Morphology and Behaviour with Deep Reinforcement Learning