Central Pattern Generators (CPGs) have several properties desirable for locomotion: they generate smooth trajectories, are robust to perturbations and are simple to implement. However, they are notoriously difficult to tune and commonly operate in an open-loop manner. This paper proposes a new methodology that allows tuning CPG controllers through gradient-based optimisation in a Reinforcement Learning (RL) setting. In particular, we show how CPGs can directly be integrated as the Actor in an Actor-Critic formulation. Additionally, we demonstrate how this change permits us to integrate highly non-linear feedback directly from sensory perception to reshape the oscillators’ dynamics. Our results on a locomotion task using a single-leg hopper demonstrate that explicitly using the CPG as the Actor rather than as part of the environment results in a significant increase in the reward gained over time (20x more) compared with previous approaches. Finally, we demonstrate how our closed-loop CPG progressively improves the hopping behaviour for longer training epochs relying only on basic reward functions.

[ pdf] [ DOI: 10.1007/978-3-030-89177-0_3] [ arXiv]