Sim-to-real transfer of co-optimized soft robot crawlers

This work provides a complete framework for the simulation, co-optimization, and sim-to-real transfer of the design and control of soft legged robots. Soft robots have “mechanical intelligence”: the ability to passively exhibit behaviors that would otherwise be difficult to program. Exploiting this capacity requires consideration of the coupling between design and control. Co-optimization provides a way to reason over this coupling. Yet, it is difficult to achieve simulations that are both sufficiently accurate to allow for sim-to-real transfer and fast enough for contemporary co-optimization algorithms. We describe a modularized model order reduction algorithm that improves simulation efficiency, while preserving the accuracy required to learn effective soft robot design and control. We propose a reinforcement learning-based co-optimization framework that identifies several soft crawling robots that outperform an expert baseline with zero-shot sim-to-real transfer. We study generalization of the framework to new terrains, and the efficacy of domain randomization as a means to improve sim-to-real transfer.