Light Origins Launches Light-O1 Foundation Model for Robot Learning
aiinsider ·
Light Origins has launched Light-O1, its first general-purpose embodied foundation model designed for robot learning. The company trained six versions of the 4-billion-parameter model using up to 120 billion multimodal tokens derived from human actions extracted from internet videos. These larger pretraining runs significantly reduced prediction errors for next-action and whole-body pose tasks across human, Unitree G1, and LightBot datasets. In addition to the main model, Light Origins released Light-O1-Preview, a text-to-action system that generates body movements based on natural language instructions. Demonstrations showed robots performing complex tasks such as opening cabinets, picking up trash, and handing over towels. The company also outlined a broader physical AI roadmap including LightNav-0 for alignment and Light REACT for adapting to recent physical interactions. Data infrastructure now processes approximately 200,000 hours of video weekly, supporting scalable pretraining efforts.
aiinsider