Feedback Error Learning for Gait Acquisition

Robots execute a wide array of rhythmic behaviors like walking, running, dribbling, etc. Rhythmic movement primitives are parametric models that can be used to represent these behaviors. Previously, these models have been learned either by means of demonstrations or by means of reward signals. While executing a rhythmic behavior, a stabilizing controller is needed in most robots to prevent failure. An example for this controller can be that of a balance controller, needed to keep the torso of a biped upright while walking. The stabilizing controller uses a feedback mechanism to maintain stability. Learning a balanced forward trajectory or gait in case of a biped, which mutes the stabilizing controller would avoid the use of high gain feedbacks, and instead use the feedbacks only in case of emergencies. We propose to learn the balanced forward trajectory, i.e., its movement primitive using feedback error learning. Feedback error learning is a model-learning method that we modify to learn rhythmic movement primitives using the feedback of the stabilizing controller. This modification, along with the architecture of feedback error learning allows us to learn new trajectories without complicated reward functions, or training examples as are needed by reinforcement learning or supervised learning respectively. Moreover, this method is online. Thus, the trajectories can be learned by the system while executing them stably. We test our novel method, using simulations on two problems: forward trajectory learning in a two link manipulator and gait adaptation by a biped robot to confirm the method’s efficacy in learning trajectories using feedback errors.