Abstract
Traditional character animation often relies on kinematic-based methods (key frame, motion capture, etc.), resulting in less flexibility and authenticity. To solve this problem, we propose a walking control method based on the reinforcement learning algorithm. Compared with the traditional control method, besides using motion capture data as a reference, we adopt a multi-reward function training to generate more realistic motion. At the same time, the step update problem of the learning process is avoided, and the training is more stable. For verification, we employ the bullet physics engine to build a human motion model called RLM, which combines motion data with reinforcement learning to generate fast and stable motion. The experiment shows that the learning efficiency and the final reward are evidently improved; the RLM model achieves a stable walking in the physical simulation, being practically consistent with the real captured trajectory.