Learning to extend legs of a transformable-wheel robot

James Clinton, Willy Li, Max Linden, Anthony J. Clark, Elliot Hawkes

Keywords: proprioceptive locomotion, mechanism design

Abstract: Designing robots for dynamic, unstructured environments is a challenging task. Both the robot's mechanical design and control policy must be optimized for the environment. In this study, we explore the design of a transformable-wheel robot that can drive on smooth wheels and extend its legs to crawl over obstacles. We use reinforcement learning to train a policy for the robot that uses proprioception, namely motor velocity, body acceleration, and body orientation, to determine when to extend its legs. Our results indicate that this proprioception alone may be sufficient to control transformable wheels. In simulation, the robot was able to climb over a single step obstacle using only proprioception, though it was better able to do so with a collision sensor.

Submission Number: 16