Traditional programming method can achieve certain manipulation tasks with the assumption that robot environment is known and structured. However, with robots gradually applied in more domains, robots often encounter working scenes which are complicated, unpredictable, and unstructured. To overcome the limitation of traditional programming method, in this paper, we apply deep reinforcement learning (DRL) method to train robot agent to obtain skill policy. As policy trained with DRL on real-world robot is time-consuming and costly, we propose a novel and simple learning paradigm with the aim of training physical robot efficiently. Firstly, our method train a virtual agent in an simulated environment to reach random target position from random initial position. Secondly, virtual agent trajectory sequence obtained with the trained policy, is transformed to real-world robot command with coordinate transformation to control robot performing reaching tasks. Experiments show that the proposed method can obtain self-adaptive reaching policy with low training cost, which is of great benefits for developing intelligent and robust robot manipulation skill
system.