Light Origins Launches Light-O1 Foundation Model for Robot Learning

Light Origins positions Light-O1 as the "first general-purpose embodied foundation model," with its core selling point being that it uses human motion recovered from internet video to provide a reusable learning starting point for different tasks and different robot embodiments. The company says it trained six versions of the same 4-billion-parameter base, with multimodal token usage ranging from 3.75 billion to 120 billion, the largest run roughly equivalent to 100,000 hours of human motion.

But the most noteworthy thing about this release is not the scale, but the boundary the company itself draws: robot-side results still use target-specific data, and the scaling tests measure held-out next-action prediction and whole-body pose prediction, not real-world task success rates. In other words, what can currently be verified is that "prediction error falls as pretraining scale rises," not that "robots can do more work." This distinction is provided by the company itself and is one of the hardest constraints in this release.

According to the company, the six versions were adapted to three types of data: public first-person human data, public Unitree G1 robot data, and data from Light Origins' own LightBot humanoid robot. The company says that across all three datasets, prediction error fell as pretraining scale increased, showing a power-law trend; larger-scale human motion pretraining provides a stronger starting point for downstream adaptation.

These conclusions are currently all company self-reports, from press-release-style coverage, with no third-party reproduction or independent evaluation. What can be confirmed is the release itself and the disclosed training configuration; whether the "stronger starting point" translates into cross-embodiment general capabilities remains an unverified inference. The tasks mentioned in the demos include LightBot opening a shoe cabinet and putting in slippers, picking up different types of trash (including objects moved during the task), handing over a towel, and Unitree G1 wiping a table and taking a towel in the same demo—these are demo descriptions, not success-rate data.

Released alongside Light-O1 is Light-O1-Preview, which the company describes as a text-to-motion model: it accepts natural-language instructions describing the desired body motion and generates the corresponding whole-body motion sequence; model weights, code, and a public demo environment are provided with the release.

The company organizes the advancement of physical AI into three stages: scalable pretraining, alignment, and deployment. Light-O1 corresponds to pretraining; LightNav-0, launched earlier this month, is for alignment, and its Real2Sim2Real system is said to convert more than 2,000 internet-sourced scenes into simulation environments, producing over 4,000 hours of navigation experience, with the resulting model generalizing to humanoid, quadruped, aerial, and wheeled robots without additional training; on the deployment side is Light REACT, used to let robots adjust behavior based on recent physical interactions, estimating external forces, hardware damage, and environmental constraints before adjusting whole-body motion. These are all company self-reports, with no independent verification yet.

Supporting the above training is what the company describes as its data infrastructure: it has reached thousand-GPU scale, processing about 200,000 hours of video per week, compared with 12,500 hours six months ago. This growth rate is one of the few comparable figures in this release, but it likewise comes from company self-reports.

On the funding side, this Chinese company says it completed a Pre-A round of several hundred million yuan in August 2026. This information appears as background in the coverage, without investors or a specific amount, so it can only be treated as a company-level statement, not a confirmed financing fact.

What can be confirmed: Light Origins released Light-O1 and Light-O1-Preview, disclosed the training configuration (six 4-billion-parameter versions, 3.75 billion to 120 billion tokens, the largest about 100,000 hours of human motion), the three types of adaptation data, and the opening of weights and code. These are statements at the level of release actions and configuration.

What must be held in reserve: the scaling conclusions, the power-law trend, cross-embodiment generalization, the thousand-GPU scale and 200,000 hours of video per week, and the several-hundred-million-yuan Pre-A round all come from company self-reports or secondhand retellings, lacking first-party verifiable materials and third-party reproduction. The most critical gap is that the company itself admits it measures prediction error rather than task success rate—meaning that from "more accurate prediction" to "robots can do more work," there remains a distance not yet covered by this evidence.