Light Origins open-sources Light-O1-Preview, a 6B whole-body motion model for robots

Light Origins has open-sourced Light-O1-Preview, a roughly 6-billion-parameter whole-body action model released under the Apache 2.0 license on Hugging Face and GitHub. According to Pandaily, citing the company's September 21 technical blog, the checkpoint is positioned as the public preview slice of Light-O1, Light Origins' embodied foundation model line, and is distinct from the firm's earlier LightNav-0 navigation stack.

The core idea is to turn human video into robot motion. Light Origins segments people in video, reconstructs their three-dimensional motion, and tokenizes root trajectory, body pose and hand state into discrete action tokens. Those tokens are interleaved with language and vision and used for autoregressive Transformer pretraining, so the model absorbs action priors from human footage rather than from robot teleoperation alone.

The company reports scaling experiments spanning roughly 3.75 billion to 120 billion multimodal tokens, with the largest measured budget covering up to about 100,000 hours of recovered human action. In those runs, next-action loss and whole-body pose error reportedly fall in power-law fashion after adaptation to egocentric human data, Unitree G1 teleoperation sets and Light Origins' own LightBot corpora.

Launch materials show the same motion prior transferred onto LightBot and a Unitree G1 for behaviors such as kneeling, balancing, wiping and picking up objects. Cross-embodiment transfer is the central claim: one pretrained prior, several different robot bodies.

On the interface side, the Hugging Face preview lists a 6B checkpoint inside the Qwen3.5-4B-Base model tree. It accepts a text instruction, emits a short reasoning trace and returns whole-body action sequences at 20 frames per second, using a 138-value-per-frame layout that covers root motion, pelvis height, yaw, 22 joints and hand-open states.

This is not a turnkey robot operating system. Deploying the trajectories on hardware still requires a separate behavior model or low-level controller, and Light Origins documents a Unitree G1 example path that currently runs through simulation with an extra policy checkpoint. The company's RoboCasa GR-1 kitchen simulations are reported at a 79.3% macro success rate across 24 tasks, which is a company-run result rather than an independent public leaderboard.

For robotics teams, the practical value is a downloadable motion-prior layer under a permissive license. Open weights let labs test the human-video pretraining claim directly and fine-tune on their own embodiments. What to watch next is whether the adaptation data and low-level controllers stay partly proprietary, and whether independent evaluations reproduce the reported success rates on real hardware.