Reinforcement Learning Based Character Controlling

Character controlling is a longstanding problem in understanding the behavior of human. This task aims to generate various and high quality human motion in the simulated environment as in the real world controlled by human. Therefore, how to use the captured real human motions is the crucial component to solve this problem. Rather than regressing the human motion directly in previous motion prediction methods, in this project, a reinforcement learning method is adopted in to our framework to learn robust control policies capable of imitating a broad range of example motion clips. Besides, we also explore the new reward functions to encourage the motion similarity between the real human and the virtual character. With these new rewards, the algorithm will convergence faster than recent advances.