Control strategies for physically simulated characters performing two-player competitive sports

In two-player competitive sports, such as boxing and fencing , athletes often demonstrate efficient tactical movements during a competition. this paper, we develop learning framework that generates control policies for physically simulated who have many degrees-of-freedom. Our uses two step-approach, basic skills bout-level strategies, with deep reinforcement learning, which is inspired by the way people how to learn sports. We policy model based on an encoder-decoder structure incorporates autoregressive latent variable, mixture-of-experts decoder. To show effectiveness of our framework, implemented learned can generate both natural-looking behaviors. also evaluate comparisons other configurations ablation studies.