Event

GES Framework Co-Design Optimizes BALLU Robot Morphology and Control

1 reports · 1 sources

Ankit Sinha and other authors propose a reinforcement-learning-based co-design framework, "Gaussian Evolution of Expert Strategies" (GES), for simultaneously optimizing robot morphology and control policy. The framework maintains a set of expert policies, letting their respective design-space territories evolve with control performance, in order to alleviate the behavioral diversity collapse problem in which a single shared policy converges to the same control strategy across different morphologies. The authors evaluate GES on the buoyancy-assisted lightweight legged robot BALLU, with the tasks being obstacle crossing and slope walking, and a design space of up to four dimensions. The authors report that in teleoperation hardware experiments, the optimized BALLU morphology could cross obstacles up to 24 cm high, compared with 8 cm for the original design.