据机器之心 11 月 18 日报道,浙江大学与宇树科技具身智能联合实验室等团队从海量互联网人类视频中构建并开源预训练数据集 Movens-5M,收录 500 万条经仿真执行验证的机器人动作,总时长超过 1 万小时,覆盖行走、深蹲、拖地、拳击等 297 种细分动作,涵盖居家、健身、户外等 16 类场景。Per Jiqizhixin on Nov 18, Zhejiang University and Unitree's joint embodied-AI lab built and open-sourced Movens-5M from vast internet human videos — 5 million simulation-verified robot motions totaling over 10,000 hours, covering 297 actions such as walking, squatting and mopping across 16 scenarios.

团队将文字指令、视频示范、语音和音乐与机器人动作配对,实现多模态对齐。基于这套数据训练的动作模型 Movens-Zero 能根据上述四种输入生成机器人全身动作,并在宇树 G1 上展示真机执行效果;Scaling 实验显示,数据更多、动作生成模块更大时模型性能持续提升。The team paired text instructions, video demos, speech and music with robot motions for multimodal alignment. The accompanying Movens-Zero model generates full-body motions from those four input types, demonstrated on the Unitree G1, with scaling experiments showing continued gains as data and model size grow.