Morphology-Control Co-design

Optimizing a robot's body and its control policy together, instead of fixing the body first.

Morphology-control co-design means searching for a robot's body (number of limbs, limb lengths, joint layout, and so on) and the policy that controls it at the same time. The traditional approach fixes the body first and then writes or trains a controller for it; co-design argues that the best controller depends on the body, and the best body depends on the controller, so both should be optimized in a single loop. Karl Sims's 1994 SIGGRAPH paper “Evolving Virtual Creatures” used simulated evolution to co-evolve the morphology and neural controllers of virtual creatures. Stanford's Gupta and colleagues built on this with DERL (2021), which evolves the body while using reinforcement learning to train the controller; ICLR 2022's Transform2Act instead treats “changing the body” as just another action for the policy to learn.

ExampleIn DERL, virtual agents assembled from different limbs learn to walk and manipulate objects in simulation; the better-performing individuals are kept and have their body structure mutated, and after many generations, more stable, more energy-efficient, faster-learning morphologies emerge.

Also called

Morphological Evolution, Brain-Body Co-design, Co-design of Morphology and Control

Related

Morphological Computation · Reinforcement Learning · Embodiment · Simulator · Soft Robot · Embodied AI

Sources

Evolved Virtual Creatures (Karl Sims, 1994)
Embodied Intelligence via Learning and Evolution (DERL)
Transform2Act: Learning a Transform-and-Control Policy for Efficient Agent Design