Towards Morphology-Agnostic Control for Soft Robots

Abstract

The advent of soft robots promises to fundamentally shift the landscape of robotic systems as they offer several advantages over the current paradigm of rigid bodies. Most notably, they provide adaptability to uncertain environments and look to bridge the gap between humans and machines. However, determining the optimal structure of a soft robot for a given task is difficult and complicated by the fact that soft robots have a design-dependent control profile. Thus, existing approaches have relied on human intuition or biomimicry. Co-design has been introduced as an approach to developing soft robots and involves jointly optimizing over the design and control of compliant bodies. An iterative design optimization routine suggests new morphologies while a control optimization subprocess determines a controller for each unique body. However, in its current form, co-design is a lengthy process due to the control optimization step being computationally expensive. Moreover, this step must be carried out separately for every unique morphology. This thesis discusses the development of MANTIS: a Morphology-Agnostic Controller for Soft Robots. We evaluate MANTIS against expert controllers using a soft robotic benchmarking suite (EvoGym) and demonstrate proficiency in zero-shot generalization to unseen morphologies. Importantly, this work makes strides towards universal control for soft robots, an objective which will greatly accelerate the rate of research in soft robotics.