On September 21, Liangyuan Xinchuang released Light-O1, a basic model for full-body intelligence, which first verified the cross-body migration expansion rules of human full-body movement pre-training. The model learns human movements from massive Internet videos, combines language and visual information to migrate to robots, breaking through bottlenecks such as high remote operation costs and insufficient diversity. The pre-training of human movements reached 100,000 movement hours. After the data was expanded, motion prediction losses adapted across the body and perspective showed a downward trend in power law, and total body posture errors decreased simultaneously. In the same period, the Light-O1-Preview model was open source on the company's official website, and its three-stage technical achievements of “large-scale pre-training - large-scale alignment - large-scale deployment” were all implemented.

On September 21, Liangyuan Xinchuang released Light-O1, a basic model for full-body intelligence, which first verified the cross-body migration expansion rules of human full-body movement pre-training. The model learns human movements from massive Internet videos, combines language and visual information to migrate to robots, breaking through bottlenecks such as high remote operation costs and insufficient diversity. The pre-training of human movements reached 100,000 movement hours. After the data was expanded, motion prediction losses adapted across the body and perspective showed a downward trend in power law, and total body posture errors decreased simultaneously. In the same period, the Light-O1-Preview model was open source on the company's official website, and its three-stage technical achievements of “large-scale pre-training - large-scale alignment - large-scale deployment” were all implemented.