MoCapAct

该数据集名为MoCapAct，包含了超过50万个专家MoCap追踪策略生成的模拟人形角色的运动轨迹，总计涉及6700万个环境转换。这些策略涵盖了广泛的人类运动行为，如走路、跑步以及更复杂的动作，如侧手翻等。该数据集被用于评估H-GAP在表示运动先验以及在模仿学习和控制任务中的性能。其规模宏大，包含超过50万个回放和6700万个环境转换，旨在应对模仿学习及相关下游控制任务。

The dataset is named MoCapAct. It contains over 500,000 motion trajectories of simulated humanoid characters generated by expert MoCap tracking policies, totaling 67 million environment transitions. These policies cover a wide range of human locomotor behaviors, including walking, running, and more complex movements such as cartwheels. This dataset is utilized to evaluate the performance of H-GAP in representing motion priors, as well as in imitation learning and control tasks. Boasting its large scale with over 500,000 replays and 67 million environment transitions, it is designed to address imitation learning and related downstream control tasks.