A heuristic reinforcement learning for robot approaching objects

Autonomous approaching objects for an arm-hand robot is a very difficult problem because the possible arm-hand configurations are numerous. In this paper, we propose a modified reinforcement learning algorithm for a multifingerd hand approaching target objects. The proposed approach integrates the heuristic search information with the learning system, and solves the problem of how an arm-hand robot approaches objects before grasping. In addition, this method also overcomes the problem of time consuming of traditional reinforcement learning in the initial learning phase. The algorithm is applied to an arm-hand robot to approach objects before grasping, which can enable the robot to learn approaching skill by trial-and-error and plan its path by itself. The experimental results demonstrate the effectiveness of the proposed algorithm.