This paper presents an approach of sensory feedback integration into a CPG model under a hierarchical control architecture for adaptive caterpillar-like locomotion. Motivated by the simplicity of CPG models in gait generation, a sinusoidal generator is employed as the low level controller of a caterpillarlike robot. To regulate the behavior of the robot with respect to the changes of environmental conditions, a behavior adaptor is designed to integrate sensory feedback into the sine-based model, so that sensory input can be transferred into the model and further affect the output of the model to some extent. Meanwhile, a policy gradient based reinforcement learning method is adopted at the high level, aiming to learn the mapping between sensory input and reasonable responses of the robot. The optimized policy could be obtained after episodic learning. Simulation results show the caterpillar-like robot can climb over uneven terrains with the help of the sinusoidal generator and the learned policy, which verifies the effectiveness of the proposed approach.
展开▼
