Reinforcement learning from expert demonstrations with application to redundant robot control

Current methods of reinforcement learning from expert demonstrations require humans to give all possible demonstrations in the learning phase, which is very difficult for continuous or high-dimensional spaces. In this paper, we proposed biased exploration reinforcement learning to avoid the exploration of unnecessary states and actions of the expert demonstrations. We present a convergence analysis of the novel method. This method is applied to learn the control of a redundant robot manipulator with 7-degree-of-freedom. The experimental results demonstrate that the proposed method accelerates the learning phase. The obtained policy can successfully achieve the pretended task.