Light Origins

Light Origins is releasing Light-O1, a whole-body action model for humanoid robots that treats internet video of people as the primary pretraining source, then adapts the resulting representation to specific robot bodies. The training pipeline segments videos, tracks people, reconstructs their movement in three dimensions and aligns those actions with visual observations and language annotations, encoding the result as discrete action tokens for an autoregressive Transformer that predicts language and action in the same sequence. In the launch demonstrations, the same starting model drives Light Origins' own LightBot and a Unitree G1 through movements including kneeling, balancing, wiping a blackboard and picking up an object. The company is led by founder Roger Jiang, who previously worked at OpenAI across pretraining, alignment, optimization and infrastructure for InstructGPT, ChatGPT and GPT-4, holds a physics Ph.D. from the University of Maryland and founded Light Origins in late 2024. His thesis for Light-O1 is that the scaling behavior seen in language modeling can be recreated for physical behavior by tapping the much larger pool of human video before spending scarce robot hours on embodiment-specific training. Light Origins reports power-law declines in next-action prediction loss and whole-body pose error as the pretraining budget increases across six budgets ranging from 3.75 billion to 120 billion multimodal tokens, with the largest run corresponding to roughly 100,000 hours of recovered human action. Alongside the full system, Light Origins is shipping an open preview for developers to work with directly. Light-O1-Preview is available on Hugging Face under the Apache 2.0 license with an inference repository on GitHub, listed as a 6 billion parameter checkpoint traced to Qwen3.5-4B-Base, accepting a text instruction and producing a short reasoning trace plus a sequence of whole-body actions at 20 frames per second across root movement, pelvis height, yaw, 22 joints and hand-open states. The timing matters because the launch lands roughly five weeks after a Pre-A round of several hundred million yuan led by CAS Investment on August 14th, with China Merchants Venture Capital and Xiang He Capital participating , giving humanoid teams a concrete way to test whether a shared human-video base can shorten the path to useful robot behavior.