Robotic throwing based on deep reinforcement learning with predictive reward system

Through throwing action, the spatial reachability of a robot can be enhanced. However, the learning of the throwing skill is difficult and still faces some open issues. In this paper, we employ deep reinforcement learning to address the throwing task and acquire parametric settings for throwing skill. Considering in the training, the agent can only get the rewards from the environment at the end of the episode, these rewards are sparse and delayed which poses challenges for convergence efficiency of the learning. We propose a compound reward system based on predictive reward, which outputs rewards according to the state information of the environment with turning sparse reward into dense reward. Finally, both the feasibility and the superiority of the proposed method is verified through comparative tests.