Light Origins Releases Light-O1 Embodied Foundation Model Trained on Internet Human Motion

Light Origins announced the release of Light-O1, describing it as its first general-purpose embodied foundation model. The model was trained on human motions recovered from internet videos, aiming to give robots a reusable starting point for adapting to different tasks and physical embodiments.

The company trained six versions with 4 billion parameters each. Training used up to 120 billion multimodal tokens, and the largest run represented roughly 100,000 hours of human motion. This scale is intended to help the model generalize across robots rather than remain tied to a single platform.

Alongside the foundation model, Light Origins released Light-O1-Preview. This text-to-motion model accepts natural-language instructions and generates whole-body motion sequences. The company said model weights, code, and a public playground are available with the release, allowing researchers and developers to test the system.

Light Origins positions Light-O1 as a step toward reusable embodied intelligence, where robots can learn from human motion data found online and transfer that knowledge to varied bodies. The preview release focuses on text-driven motion generation, while the broader Light-O1 family is presented as a foundation for future robotics tasks. The release also highlights how internet video can supply scalable motion data for embodied AI.