Generic Neural Locomotion Control Framework for Legged Robots

In this article, we present a generic locomotion control framework for legged robots and strategy policy optimization. The is based on neural black-box combines central pattern generator (CPG) radial basis function (RBF) network to create CPG-RBF network. acts as produce arbitrary rhythmic trajectories the joints of robots. main features are: 1) it since can be applied with different morphologies; 2) has few parameters, resulting in fast learning; 3) scalable, both terms policy/trajectory complexity number legs that controlled using similar trajectories; 4) does not rely heavily sensory feedback generate thus less prone faults; 5) once trained, simple, minimal, intuitive use analyze. These will lead an easy-to-use convergence ability encode complex policies. work, show successfully three simulated varying morphologies and, even broken joints, learn We also after learning, policies transferred real-world robot without any modifications. We, furthermore, scalability by implementing controller all decentralized individual leg pairs. By investigating correlation between morphology encoding type, are able Finally, how integrated into enable online adaptation.