Research of reinforcement learning based share control of walking-aid robot

In this paper we developed a new reinforcement learning based share control algorithm for walking-aid robot. We use a group of one-dimensional push-pull force sensors to estimate human walking intention, from which the user's desired moving velocity of robot is obtained. At the same time, the robot itself also plans a desired moving velocity. A weighted sum of the two desired velocities is taken as the real reference velocity and fed into the motion controller. The Sarsa-learning algorithm dynamically adapts the weights of user's control according to the control efficiency, the robot state and the environment. As a result, an optimal share control for walking-aid robot is realized in a certain environment. Finally experiments are performed to verify the effectiveness of algorithm.