Zhejiang University and Unitree Release 5 Million Robot Pretraining Data Entries
3 hour ago / Read about 0 minute
Author:小编   

A joint effort between Zhejiang University's Joint Laboratory for Embodied Intelligence, Unitree Robotics, and other teams has produced and open-sourced Movens-5M, a pretraining dataset built from massive volumes of human videos available online. The dataset contains 5 million robot actions validated through simulation, with a total duration exceeding 10,000 hours. It covers 297 fine-grained actions—such as walking, squatting, mopping, and boxing—as well as 16 categories of scenarios, including home life, fitness training, and outdoor leisure. The motion model Movens-Zero, trained on this data, can generate full-body robot motions from four types of input. The team demonstrated real-machine execution performance on the Unitree G1 and conducted Scaling experiments to test whether the model can continue to improve as the data scale and motion generation module increase.