yan0116/SMPL_Humanoid_offline_dataset

UniPhys数据集是一个用于物理基础角色控制的大规模离线数据集，支持基于SMPL-like角色的扩散策略训练。数据集包含由PULSE运动跟踪策略生成的状态-动作对，分为AMASS状态-动作对和BABEL状态-文本-动作对两种类型。AMASS状态-动作对包含关节位置、旋转、速度等信息，BABEL状态-文本-动作对在此基础上增加了帧级别的文本注释。UniPhys Dataset is an offline dataset for physics-based character control, supporting large-scale training of diffusion policies using SMPL-like characters. The dataset contains state-action pairs generated by the PULSE motion tracking policy, divided into two types: AMASS state-action pairs and BABEL state-action-text pairs. AMASS state-action pairs include joint positions, rotations, and velocities, while BABEL state-action-text pairs add frame-level text annotations to this information.