Light Origins trains Light-O1 on human video before adapting it to robots
Founder Roger Jiang says larger pretraining runs reduced held-out prediction errors after adaptation, though the study did not measure real-world task success at scale.
By RuntimeWire Staff · Published
Primary source: PR Newswire
Why it matters
Light-O1 tests whether internet video can reduce the cost of building robot foundation models. Its current evidence shows better held-out prediction after adaptation as pretraining scales, while real-world task success at scale remains unmeasured.
Light Origins, a company building foundation models for Physical AI, has released Light-O1, a model pretrained on structured human actions recovered from internet videos and then adapted to different robot bodies and tasks. The company's technical report, published September 21st, puts founder and CEO Roger Jiang (@RogerXJiang) behind a specific bet: human video can provide a broader starting point for robot learning than robot demonstrations alone.
Jiang brings a language-model scaling background to the problem. Before founding Light Origins, he worked at OpenAI from 2019 to 2023, contributing to pretraining, alignment, optimization and infrastructure for InstructGPT, ChatGPT and GPT-4, according to Light Origins' founder materials. The GPT-4 technical report lists him among its authors. He has argued that internet video could supply movement and interaction data at a scale that dedicated robot collection cannot easily match.
Light-O1 is the first product in Light Origins' scalable pretraining pillar. The company's other announced work covers scalable alignment and deployment: LightNav-0 uses simulated navigation experience, while Light REACT is designed to adapt robot behavior to physical changes and constraints. Together, those projects show Jiang is building a full training-and-deployment program around the idea that foundation-model methods can carry into the physical world.
The result is a scaling curve, not a success rate
For its scaling study, Light Origins started from the same 4-billion-parameter base model and trained six independent models with pretraining budgets ranging from 3.75 billion to 120 billion multimodal tokens. The largest run represents about 100,000 hours of human action, the company says. Researchers then adapted those models separately to public egocentric human data, public Unitree G1 robot data and Light Origins' in-house LightBot data.
On held-out data in all three settings, the company reports that increasing the human-action pretraining budget reduced next-action-token prediction loss and whole-body pose prediction error. Light Origins says the measured trends followed power-law fits. The robot results still involve adaptation using data from the target domain; the experiments test whether larger human-action pretraining gives the model a stronger starting point, not whether a robot can transfer without robot-specific training.
That distinction sets the limits of the evidence. The scaling analysis reports prediction and pose metrics, including open-loop pose evaluation. It does not provide a scaling curve for successful task completion on real robots, or independent measurements of reliability, cost or performance in commercial settings. The company's videos show LightBot opening a shoe cabinet, picking up moved trash and handing over a towel; a Unitree G1 demonstrates table wiping and a towel handoff. Those demonstrations are separate from the reported scaling experiments.
Light Origins is also releasing Light-O1-Preview, a reasoning text-to-action model. Given a natural-language instruction, it first describes what the body needs to do, then generates a whole-body action sequence. The public code repository says its output is a 138-dimensional action representation at 20 frames per second, and includes a Unitree G1 simulation example. The company says weights, code and a public playground are available.
The costly part is building the data pipeline
The approach moves some of the data burden upstream. Rather than relying only on demonstrations recorded for a particular robot, Light Origins extracts structured 3D actions from existing video, aligns them with visual observations and language, and trains an autoregressive model on those sequences. Robot data remains necessary for adaptation, but the company is trying to make it less responsible for supplying the entire foundation.
That plan depends on substantial infrastructure. Light Origins says its data operation runs at thousand-GPU scale and processes about 200,000 hours of video a week, up from 12,500 hours six months earlier, according to the company's release. These are company-reported throughput figures, not evidence by themselves that the resulting actions transfer reliably to deployed robots. The key test will be whether gains in prediction metrics translate into repeatable task performance across hardware and environments.
The timing also follows a financing milestone. On August 14th, Light Origins said it had closed a Pre-A round of several hundred million yuan, led by CAS Investment, with China Merchants Venture Capital and Xiang He Capital participating. The company said the funding would support model training, data infrastructure, software and hardware research, and hiring. It has not announced a valuation. Light Origins was founded in late 2024 and says it now employs more than 100 people, with over 90% in research and development.
Jiang's experience at OpenAI helps explain the structure of the wager: scale a broad pretraining signal, then use narrower downstream data to adapt the model. In language models, the relevant substrate is text. For Light-O1, it is human movement converted into action tokens. The hard question is whether that substrate carries enough information about contact, balance and changing physical conditions to reduce the amount of robot-specific data required.
The first results support a narrower conclusion: on the reported held-out prediction tasks, increasing the human-action pretraining budget improved the adapted models' metrics across three domains. Turning that relationship into reliable robot work will require evidence beyond the scaling curves and selected demonstrations. Light-O1 makes that gap explicit by showing where the company's pretraining thesis has quantitative support, and where deployment still has to prove it.