Light Origins Debuts Light-O1 to Scale Physical AI via Human Action

The core of Light-O1 lies in its ability to translate structured 3D human motion into whole-body action sequences for machines. In experiments ranging from 3.75 billion to 120 billion multimodal tokens, the company observed that larger pretraining budgets consistently reduced prediction errors across both egocentric human data and robotic platforms like the Unitree G1. This power-law scaling suggests that massive human-action datasets serve as a reliable foundation for downstream robot adaptation, bypassing the bottleneck of collecting bespoke interaction data for every new hardware embodiment.

Beyond technical benchmarks, the model demonstrates practical utility in multi-step environments. Real-world tests show the LightBot humanoid navigating household chores—such as organizing slippers or managing trash—even when objects shift during execution. To support this rollout, Light Origins is providing a preview version of the model, which interprets natural-language instructions to generate corresponding physical movements. Founder and CEO Roger Jiang emphasizes that this approach is one of three pillars in the company’s roadmap, alongside alignment and deployment, aimed at creating more generalized physical intelligence.