On September 21, embodied-intelligence startup LightInspire (Liangyuan Xinchuang) released Light-O1, a whole-body intelligence foundation model that learns human motion from massive internet video, combining language and vision to transfer action knowledge to robots—breaking through the industry bottlenecks of costly teleoperation and insufficient data diversity.
The company says the model reveals, for the first time, the scaling law of cross-embodiment transfer from human whole-body action pre-training: the more human action data used, the more accurate action prediction becomes after adapting to different robot bodies and viewpoints. Its human-action pre-training dataset has reached 100,000 action-hours, and full-body pose error falls as data scale grows.
The model connects instruction understanding, environment perception and whole-body action, driving its LightBot to perform continuous tasks such as handing over a towel, putting away slippers and picking up trash, and making a Unitree G1 water flowers and wipe tables. The company also open-sourced a general action-generation model, Light-O1-Preview.
Source: Tencent News / Phoenix / People's Finance
