Identifying and Exploiting Structure in Robot Co-Design

原文アブストラクト

Co-design of a robot's morphology and control is a high-dimensional search problem. Efficient search depends on exploiting the structure shaped by the interaction between morphology and control. In this paper, we analyze co-design landscapes of two soft locomotion and two manipulation tasks. We identify three patterns that are consistent across regions of their co-design spaces: 1) Within a region, quality varies along a low-dimensional manifold, with minimal variation orthogonal to it, reducing the effective search space dimensionality. 2) In higher-quality regions, the variance in quality is spread across more dimensions, necessitating search to expand dimensionality as quality improves. 3) In higher-quality regions, quality varies along joint morphology-control dimensions, requiring search along them. We leverage these insights to devise an efficient co-design algorithm. This algorithm yields co-designs $36\%$ better than those of state-of-the-art algorithms. Moreover, these benchmark algorithms required about an order of magnitude more function evaluations to achieve co-designs of comparable quality to ours. We inspect the exploration patterns of all algorithms and ablate our algorithm, demonstrating that exploiting the identified structure enables efficient co-design.