Light-O1 Robot Foundation Model Bets on Human Video Pretraining

What happened

Light Origins, a startup that is just under two years old, has released Light-O1. The company calls it its first general-purpose embodied foundation model. The core idea is to teach robots how bodies move by learning from people in internet video instead of relying only on demonstrations recorded on robots.1413 The model pulls structured 3D human motion out of online footage, pairs that motion with visual observations and language, and trains an autoregressive model on the combined sequences. The resulting "action prior" can then be fine-tuned for a particular robot and task.13

The headline result is a scaling study. Starting from one 4-billion-parameter base model, the company trained six separate versions on pretraining budgets from 3.75 billion to 120 billion multimodal tokens. The largest budget equals roughly 100,000 hours of human action.11 Each version was then adapted three ways: to public first-person human data, to public data from the Unitree G1 humanoid, and to data from Light Origins' own humanoid, LightBot.11 In all three settings, larger pretraining budgets led to lower held-out next-action and whole-body pose prediction errors, and the company says the gains followed power-law curves.11

Alongside the study, the company released Light-O1-Preview, a text-to-action model, with weights, code and a public playground.11 Given an instruction, it first writes out in plain language what the body needs to do, then generates a whole-body motion sequence.13 The GitHub documentation says the output is a 138-value-per-frame action representation at 20 frames per second, and that the preview covers text-to-action only. Vision and manipulation are planned for the full Light-O1.16

The founder and the money

The launch makes more sense with the founder's background in view. Light Origins was founded in late 2024 by Roger Jiang, known in Chinese-language coverage as Jiang Xu. He worked at OpenAI from 2019 to 2023 on InstructGPT, ChatGPT and GPT-4.1 Company materials say he is credited eight times in the GPT-4 technical report, tied with co-founder Greg Brockman for the most credits.5

In mid-August, Light Origins announced a Pre-A round of several hundred million yuan. China Merchants Venture Capital and Xiang He Capital took part.5 The lead investor's name varies across outlets: CAS Investment, Guoke Investment, CASIC Investment and China Science & Merchants Investment all appear.4231 These are probably different renderings of one Chinese firm, but the inconsistency shows how much early coverage of Chinese robotics depends on translation. PitchBook also lists Legend Star among the company's four investors, which the August announcements did not mention.7 No valuation has been disclosed.

One Dealroom headline describes a "quick hundred-million-yuan re-raise" ahead of Light-O1. The article underneath, however, only refers to the August Pre-A round.12 Nothing in the reporting shows a second financing, and the more likely explanation is that the headline overstates things. The August announcement had already promised large-scale pretraining results and cross-embodiment demos.25 Light-O1 delivers on that promise about six weeks later.

The company says it has more than 100 employees, over 90% of them in R&D, with hardware engineers from Huawei and DJI.15 Its listed locations are Singapore, Shenzhen and Beijing.5 That explains why the launch release carried a Singapore dateline while most outlets describe the company as Shenzhen-based or simply Chinese.144

Why it matters: a familiar recipe in a new medium

The main reason to pay attention is the strategy, not any one demo. Jiang is applying the language-model playbook directly to robotics: find a pretraining signal that keeps improving as it scales, then build alignment and deployment on top of it. His own statement at launch put it this way: the key question is whether physical AI has a pretraining signal whose value keeps growing with scale, and he argues human action data is one.13

The company's roadmap uses the same three stages as language models. Light-O1 covers pretraining. LightNav-0, released earlier in September, covers alignment. It uses a Real2Sim2Real pipeline that turned more than 2,000 real scenes from the internet into over 4,000 hours of simulated navigation experience.11 The company says LightNav-0 works zero-shot on humanoid, quadruped, aerial and wheeled robots.13 The third stage, Light REACT, uses recent physical interactions to infer external forces, hardware damage or environmental limits and adjust whole-body movement in response.11 Back in July, a widely shared clip showed a Light Origins robot still moving with one or both legs restrained, by hopping or crawling. One commentator compared it to Skild AI's "omni-bodied" brain and pointed to the large valuation gap between the two companies.9

The underlying bet is about data supply. Robot demonstrations are slow and costly to collect. Human video already exists in huge quantities. Light Origins says its data pipeline now runs at thousand-GPU scale and processes about 200,000 hours of video a week, up from 12,500 hours six months earlier.13 If prediction gains keep translating as the company hopes, that pipeline matters more than the current model.

Where the coverage agrees, and the caveat everyone repeats

The reporting is consistent on one limitation, and the company itself states it plainly. The scaling curves measure prediction quality, not whether robots actually complete tasks. Every robot result still relies on adaptation data from the target robot.1113 RuntimeWire took the hardest line, describing the result as "a scaling curve, not a success rate." It notes that the study does not include independent measurements of reliability, cost or commercial performance. The household demos are separate from the scaling analysis. In them, LightBot opens a shoe cabinet to put slippers away, picks up trash that gets moved partway through, and hands a towel to a Unitree G1 that is wiping a table.13

Pandaily added details that make the release more concrete. The preview checkpoint is about 6 billion parameters under an Apache 2.0 license, sits in the Qwen3.5-4B-Base model tree on Hugging Face, and needs a separate behavior model or low-level controller to run on hardware.17 The documentation confirms that pairing: Light Origins uses its own behavior foundation model for LightBot and NVIDIA-associated GEAR-SONIC as the controller for the G1.16 Pandaily also says the company reports a 79.3% macro success rate across 24 RoboCasa GR-1 kitchen simulation tasks, run by the company and not on an independent leaderboard.17 So the open release is a motion-prior layer that labs can download and test, not a turnkey robot system.17

The coverage differs mainly in tone. Funding-focused outlets read the quick progression from Pre-A to open model as evidence of how fast capital and technical claims are moving in physical AI.6 Explainer-style coverage puts Light-O1 inside a broader Chinese push to build "robot brains" and the data infrastructure behind them. That includes Beijing's nearly 6,000-square-meter humanoid data training base, which aims to produce a million hours of interaction data. The more technical write-ups stick to what the metrics can and cannot show.17

The competitive context

Light Origins is entering a crowded field. Robot foundation model efforts from Skild AI, Dyna Robotics, Generalist AI, NVIDIA and Google DeepMind each take a different route to the same goal. Generalist AI, for example, recently released GEN-1.5, which is built to learn tasks from one or a few demonstrations.4 In China, investment is spreading below the model layer: tactile sensing (ModuTech), force sensors (Bluepoint Touch), and physical-interaction data collection (Mifeng Technology, which is building toward tens of millions of hours).4 Light Origins' August round was reported as part of that same wave.4

Against that backdrop, learning from human video has a real structural advantage. It avoids the robot-data bottleneck that drives competitors to build collection centers and teleoperation fleets. Its open weakness is the one Light Origins acknowledges: human footage may not contain enough information about contact, balance and force to replace robot-specific data, as opposed to just reducing how much is needed.

The verdict

Light-O1 is a credible, carefully scoped first data point, not a breakthrough in general-purpose robots. The company deserves credit for separating its scaling evidence from its demos and for releasing weights that others can test.1317 What it has shown so far is narrower than the "Physical AGI" framing implies: more human-video pretraining gives better starting points for adaptation, as measured by prediction metrics. The next milestone that matters is a scaling curve of real-robot task success, ideally checked by someone outside the company. Until that exists, Light Origins mainly has a well-funded hypothesis, a fast-growing video pipeline and a founder applying a scaling playbook he helped build once before.