The Vector Control Scheme for Amphibious Spherical Robots Based on Reinforcement Learning

Abstract

Due to variable underwater working conditions and unfavorable environments, it is difficult to design a controller suitable for underwater robots. This paper uses the adaptive ability of reinforcement learning to propose a two-layer network framework based on reinforcement learning to realize the control of amphibious spherical robots. The upper planning layer mainly plans the total torque of the robot at each moment according to the desired position and speed. The lower control layer mainly configures the parameters of the four machine legs according to the planning instructions of the upper planning layer. Through the cooperation of the planning layer and the control layer, the adaptive motion control of the amphibious spherical robot can finally be realized. Finally, the proposed scheme was verified on a simulated amphibious spherical robot.