The Pulse
Light Origins Trains Light-O1 on 100,000 Hours of Human Action
Light Origins says its Light-O1 model shows a power-law relationship between human-action pretraining scale and post-adaptation prediction accuracy across robot embodiments. No attributable quotation appears in the technical post; the page
AI.info Team ·
120 Billion Tokens Behind Light-O1
Light Origins says it trained its largest Light-O1 model on 120 billion multimodal tokens, a budget corresponding to approximately 100,000 hours of human action. The company’s technical post, dated September 21, 2026, presents the work as a study of whether human-action pretraining can provide a transferable foundation for whole-body robot intelligence.
The central result is a cross-embodiment transfer scaling law. As the human-action pretraining budget increased, next-action-token prediction loss and whole-body pose prediction error declined after the models were adapted to new datasets, robot bodies and tasks. Light Origins characterizes those declines with power-law fits.
Light-O1 is described as the company’s first general-purpose embodied foundation model. Its training approach begins with human videos, from which Light Origins recovers structured 3D actions, aligns movement with language and visual observations, and converts continuous actions into discrete tokens. The company then trains an autoregressive Transformer on multimodal sequences before adapting the resulting models to particular robot embodiments.
Six Training Budgets, Three Adaptation Settings
Light Origins began with a 4-billion-parameter base model and trained independent versions at six pretraining budgets ranging from 3.75 billion to 120 billion multimodal tokens. The budgets count language, vision and discrete action tokens processed during pretraining.
To test transfer, the company adapted each pretrained checkpoint to three settings: public egocentric human-action data, public Unitree G1 teleoperation data, and in-house LightBot teleoperation data. The evaluations measured next-action-token prediction loss and whole-body pose prediction error on held-out data.
The pose measurements used open-loop evaluation. Given a language prompt, preceding visual observations and an action prefix, the model predicted the next one-second action chunk. Light Origins compared the generated movement with recorded actions using mean per-joint position error. The company reported the best result across the evaluated post-training configurations and checkpoints at each pretraining budget.
That distinction limits what the scaling result establishes. The reported measurements concern held-out prediction and decoded pose accuracy, not a claim that a single unadapted model can perform every physical task. Light Origins presents the findings as evidence that larger human-action pretraining can create a stronger starting point for later adaptation across embodiments.
From Human Video to Whole-Body Motion
Light Origins says its data pipeline segments internet videos, detects and tracks people, reconstructs three-dimensional actions and adds language annotations. The resulting records connect visual context, semantic descriptions and physical movement in a shared temporal sequence.
The company’s unified action representation has three components: root trajectory, body pose and hand state. Those continuous representations are tokenized into discrete action tokens and interleaved with language and visual observations. The resulting sequences are used to train a model on how observations, instructions and environmental changes relate to action.
Light-O1 has two highlighted capabilities. In loco-manipulation, an instruction defines a task and the model determines how to act in the observed environment. In expressive whole-body skills, the instruction specifies a movement and the model generates a corresponding full-body action.
The technical post shows demonstrations involving humanoid robots from Light Origins and Unitree. Examples include picking up trash, wiping a table, opening a shoe cabinet and placing slippers inside, picking up headphones and placing them on a table, kneeling to propose, swinging a golf club and standing on one leg.
Reported Evaluation Results
Light Origins reports results in humanoid manipulation and human-action generation. In the simulated RoboCasa GR-1 benchmark, which covers 24 kitchen tabletop tasks on a bimanual humanoid with dexterous hands, the company reports a 79.3% macro success rate for Light-O1 over 50 episodes per task.
For human-action generation, the company says it used human ratings over more than 30,000 prompts. Judges compared Light-O1 with HY-Motion-1.0 and Kimodo on semantic following, expressiveness and acceptability. Light Origins reports an Elo score of 1,472.8 for Light-O1, compared with 1,078.3 for HY-Motion-1.0, while Kimodo was fixed at 1,000 as the anchor.
The post also reports results on the SSAE evaluation of HY-Motion-Bench. Light-O1 scored 78.0 overall, compared with 74.7 for HY-Motion-1.0 and 61.4 for Kimodo. The company says Light-O1 led the cited models across all six prompt categories in that evaluation.
Infrastructure and Open Release
Light Origins says its data infrastructure operates at the thousand-GPU scale and processes 200,000 hours of video each week. The company describes that throughput as 16 times higher than the 12,500 hours processed six months earlier.
The technical post also lists an open release for Light-O1. The release materials include Light-O1-Preview on Hugging Face, code on GitHub and a public playground. The page does not present the open materials as a complete substitute for the full system used in the company’s robot demonstrations; instead, it lists them as the currently available release components.
The clearest research claim is the reported relationship between human-action pretraining scale and post-adaptation prediction accuracy across different embodiments. The demonstrations and benchmark results extend the story to robot manipulation and human-motion generation, while the open release provides access to selected Light-O1 materials for further use and evaluation.