A Behavior-Based Reinforcement Learning Approach to Control Walking Bipedal Robots Under Unknown Disturbances

Abstract

A new approach is developed to control 3-D bipedal motion and balance under disturbance, called the behavior-based locomotion controller (BBLC). Bipedal walking is divided into various task motions and optional control behaviors, which are utilized by a behavior-based controller to generate new balancing strategies (i.e., combinations of behaviors resulting in the balance of the robot) that are more robust to unknown external disturbances. A reinforcement learning (RL) algorithm, namely Q-learning, is used to determine which behavior combinations result in new balancing strategies. The controller is implemented on ABL-BI, a 13-DOF bipedal robot. Three different disturbance cases are examined: a push, step, and slope disturbance. For each case, the BBLC is able to generate a new balancing strategy that increases the robustness of the system to the disturbance. The BBLC framework also provides the ability to interpret the RL agent’s actions, due to the combination of discrete behaviors that the agent deals with. Additionally, an evaluation of the selected balancing behaviors is completed using a stability analysis of the linear inverted pendulum.