Learning to Crawl

We combine a mechanical model for a segmented, soft-bodied crawler (Paoletti and Mahadevan, 2014) with a reinforcement learning algorithm for choosing the neuronal excitations. The crawler chooses the neuronal excitations based on a minimal description of the state of its own body, with the goal of moving forward in 1-D. The gait achieved by learning neuronal excitations in this manner depends on the mechanical properties of the crawler. For a regime of properties, the crawler achieves forward locomotion by means of a peristaltic wave, qualitatively similar to what is observed in D. melanogaster larvae. This provides a mechanism for how organisms may learn to achieve locomotion in the absence of a central pattern generator, or recover from injury. This work also suggests a way to explore actuation patterns for soft-robots in cases where the optimal actuation patterns may not be intuitive and provides a means for online learning while exploring an unfamiliar terrain.