Zhejiang University and Unitree Open-Source Movens-5M Dataset and Motion Model
What happened
Teams including the Zhejiang University–Unitree Embodied Intelligence Joint Laboratory built and open-sourced the pretrained dataset Movens-5M from internet human videos, containing 5 million robot actions validated by simulated execution, totaling over 10,000 hours and covering 297 fine-grained action types and 16 scene categories; they also released the motion model Movens-Zero trained on this dataset, with the project homepage providing access to the dataset, model, and code. In the processing pipeline, the system first estimates human 3D motion from videos, then adapts it to the robot's body structure, and uses physical refinement to fix issues such as jitter and instability, finally verifying in simulation whether the actions can be executed stably. The team demonstrated real-robot execution on the Unitree G1 and conducted scaling experiments to examine whether the model continues to improve as the data volume and the scale of the motion generation module increase.
AI 根据报道生成 · 4 days ago更新
Reporting timeline
Follow the reporting to explore different perspectives.
- 机器之心 · 微信公众号FeaturedZhejiang University and Unitree Open-Source Movens-5M Robot Pretraining Dataset and Movens-Zero Action Model
The joint embodied intelligence laboratory of Zhejiang University and Unitree Technology, along with other teams, built and open-sourced the Movens-5M pretraining dataset from internet human videos. It contains 5 million robot actions verified through simulation execution, totaling more than 10,000 hours, covering 297 fine-grained action types and 16 categories of scenes.
Activity for this event
- Current
- 3
- Peak
- 6Nov 19 02:00
- Past 24 hours
- -51%
The trend compares participants observed throughout the same period, so its range may be smaller than the current interest count. Move or tap to inspect each hour; use the arrow keys on a keyboard.