Online Series-Parallel Reinforcement-Learning- Based Balancing Control for Reaction Wheel Bicycle Robots on a Curved Pavement

Abstract

The reaction wheel bicycle robot is a kind of unmanned mobile robot with great potential. However, the control of such bicycle robots on a curved pavement under inaccurate model parameters, model uncertainties and disturbances is challenging due to the lateral instability and udneractuated characteristic. Applying conventional control methods to this problem often results in brittle and inaccurate controllers. In this paper, an online serial-parallel combination reinforcement learning with conventional control methods is designed to achieve the path tracking and banlancing control for a reaction wheel bicycle robot on curved pavements. The parallel part of the controller refers to compensate the equilibrium point and the serial part of the controller refers to adjust the parameters of a sliding mode controller that tracks the target roll equilibrium point. The comparison between the proposed controller and several existing controllers in experimental test built in Matlab Simscape illustrates stronger robustness and better control performances.INDEX TERMS reaction wheel bicycle robot, reinforcement learning, robustness, sliding model control.