Abstract: For the grasping of computer character animation, it is difficult to guarantee the naturalness, stability and adaptability of the generated action sequence at the same time. In other words, the natural and stable grasping controller are often limited in generalization and cannot be applied to other types of grabbing tasks. A virtual hand adaptive grasping controller was constructed based on deep reinforcement learning by introducing hand teaching data corresponding to the grasping types and by designing the reward function. Experimental results show that the designed controller can generate a grasping motion sequence with both naturalness and stability, and are also highly adaptive for different sizes and types of primitive objects in the material library. 

Key words: deep reinforcement learning, demonstration learning, motion generation, virtual hands, mocap data