On September 21, LightSource Innovation announced the launch of its groundbreaking full-body intelligent foundation model, Light-O1. This innovation marks a significant milestone, as it is the first to successfully validate the cross-ontology transfer and scaling laws within the realm of human full-body action pre-training. Leveraging the power of extensive internet video data, the model adeptly learns human movements and integrates both language and visual cues. It then effectively transfers this accumulated knowledge to the robotics domain, offering a robust solution to the longstanding challenges of high teleoperation costs and limited action diversity.

The model has undergone rigorous pre-training, encompassing a staggering 100,000 action-hours of human motion data. As the volume of data continues to grow, the model demonstrates a power-law decrease in action prediction loss for both cross-ontology and cross-view adaptations. This trend is paralleled by a corresponding reduction in full-body posture errors. In a move to foster collaboration and innovation, LightSource Innovation has also open-sourced the Light-O1-Preview model on its official website. This release signifies the complete realization of all technical achievements within the three-stage paradigm: "large-scale pre-training—large-scale alignment—large-scale deployment."