Reinforcement learning for human-robot shared control

This paper aims at proposing a general framework of shared control for human-robot interaction. Human dynamics are considered in analysis of the coupled human-robot system. Motion intentions of both human and robot are taken into account in the control objective of the robot. Reinforcement learning is developed to achieve the control objective subject to unknown dynamics of human and robot. The closed-loop system performance is discussed through a rigorous proof. Simulations are conducted to demonstrate the learning capability of the proposed method and its feasibility in handling various situations. Compared to existing works, the proposed framework combines motion intentions of both human and robot in a human-robot shared control system, without the requirement of the knowledge of humans and robots dynamics.