Behavior-based reinforcement learning control for robotic rehabilitation training

A behavior-based reinforcement learning controller (BRL) is developed for robotic rehabilitation training systems with time-varying properties. This adaptive BRL control system consists of an event-based planner layer, a behavior decision-making layer, and a control execution layer. In the adaptive BRL, the event-based planner layer is used to create a representative database and generate a real optimal desired trajectory for each patient, and then the behavior decision-making layer utilizes multiple behavior modules to select an optimal control behavior, which is transmitted to the control execution layer. In addition, to avoid the conflicts and the competition of different control behaviors, a self-adjusting shaping algorithm is proposed for BRL. Simulation experiments verify that the feasibility of the proposed BRL framework.