Motion Generation for Physics-Based Character by Clustering Selection

Motion generation for physics-based character is a difficult problem due to high dimensionality, non-linearity and strong coupling among joints. On the premise of analysis of the solution space of human physical character, we propose an optimization algorithm by combining spatial partition selection and intelligent evolutionary strategy. Firstly, we pre-processed the captured motion and sampled in each solution space randomly to acquire the initial population; secondly, the individuals of the initial population were optimized separately by utilizing multiple covariance matrix evolution strategies simultaneously. Then, subspace division was applied in the obtained solutions, and the optimum solution was selected in each subspace to construct the initial solution of the next stage. After several iterations, we utilized the obtained physical control trajectory to generate human physical motion. The experimental results show that the our algorithm not only makes human physical model track the motion data better, but also achieves a good improvement in robustness, time performance and stability compared with the previous work.