Tumbling Robot Optimization Using a Learned Control Policy

Abstract

Tumbling robots are simple, robust, and able to overcome large obstacles relative to their size. The specific body dimensions and leg lengths of tumbling platforms are often an uninformed choice. We introduce a framework for optimizing the parameters of a tumbling robot by referencing a control policy trained with deep reinforcement learning. We find a globally optimal tumbling robot in simulation, and provide proof for the real-world applicability of our framework by constructing and testing a physical robot based on our optimization results.