Light Origins launches Light-O1 to turn internet video into robot actions

Why it matters

Robot learning is constrained by expensive hardware-collected data. Light-O1 tests whether internet video can supply a reusable pretraining layer before robot-specific adaptation.

Roger Jiang (@RogerXJiang)'s Light Origins launched Light-O1 on September 21st, pitching human video as the pretraining layer that can give different humanoid robots a shared base of whole-body skills.

https://x.com/LightOrigins_/status/2101854146026717416

Light Origins said in a four-post thread on X that Light-O1 learns structured human actions recovered from internet videos, then adapts those actions to specific robot bodies. Demonstrations show the same starting model driving Light Origins' LightBot and a Unitree G1 through movements including kneeling, balancing, wiping a blackboard and picking up an object.

Jiang previously worked at OpenAI across pretraining, alignment, optimization and infrastructure for InstructGPT, ChatGPT and GPT-4. He holds a physics Ph.D. from the University of Maryland and founded Light Origins in late 2024. His bet is that the scaling playbook he worked on for language models can be applied to physical behavior, using the much larger pool of human video before spending scarce robot hours on embodiment-specific training.

Training on people before training on robots

The Light-O1 technical report describes a pipeline that segments videos, tracks people, reconstructs their movement in three dimensions and aligns those actions with visual observations and detailed language annotations. Light Origins encodes the resulting motion as discrete action tokens and trains an autoregressive Transformer to predict language and action in the same sequence.

The model represents an action through root trajectory, body pose and hand state. After pretraining, Light Origins adapts that shared representation to robot-specific data and low-level controllers. This separation lets a higher-level model produce an intended movement without learning every motor command again for each robot body.

Light Origins tested six pretraining budgets ranging from 3.75 billion to 120 billion multimodal tokens. The largest experiment corresponded to 100,000 hours of recovered human action. Light Origins reported power-law declines in next-action prediction loss and whole-body pose error as the pretraining budget increased, including after adaptation to public egocentric-human data, Unitree G1 teleoperation data and Light Origins' own LightBot data.

The disclosed scaling study therefore reaches 100,000 video hours, even as Light Origins describes the broader data source as internet-scale. Light Origins has not attached a total video-hour figure to the production model beyond the measured budgets in the report.

That distinction matters because robot-learning developers have increasingly turned to demonstrations, simulation and human video to escape the cost of collecting every action on physical hardware. Light Origins' approach treats robot data as an adaptation layer rather than the sole source of physical knowledge.

Jiang laid out that thesis before the launch. At a July 17th industry event, he argued that predicting actions matters more for robotics than generating plausible future video frames, and pointed to the estimated billions of hours of video already available online as the largest potential source of embodied training data.

The open model is a narrower preview

Light Origins released Light-O1-Preview on Hugging Face with weights under the Apache 2.0 license, alongside an open inference repository on GitHub. Hugging Face lists the checkpoint at 6 billion parameters and traces its model tree to Qwen3.5-4B-Base.

The public checkpoint accepts a text instruction, writes a short reasoning trace and outputs a sequence of whole-body actions. Its 138-value-per-frame representation covers root movement, pelvis height, yaw, 22 joints and hand-open states at 20 frames per second. Developers can inspect generations in a browser playground or run the model locally on a Linux system with a compatible Nvidia GPU.

The public release is narrower than the full Light-O1 system shown in Light Origins' household demonstrations. Light-O1-Preview takes text context and generates motion; deploying that output on a robot still requires a separate behavior foundation model or low-level controller. Light Origins provides a Unitree G1 example using GEAR-SONIC, though the released setup runs in MuJoCo simulation and requires an additional policy checkpoint.

Light Origins also reported a 79.3% macro success rate on 24 simulated RoboCasa GR-1 kitchen tasks, using 50 evaluation episodes per task. Light Origins trained the GR00T N1.7 and pi0.5 baselines used in the comparison and took DIAL's result from its published paper, making the ranking a Light Origins-run evaluation rather than a single independently administered leaderboard.

The model follows the money

The launch comes five weeks after Light Origins announced a Pre-A round of several hundred million yuan. CAS Investment led the August 14th financing, with China Merchants Venture Capital and Xiang He Capital participating. Light Origins said the capital would fund large-scale model training, multimodal data infrastructure and software-hardware development.

Light-O1 is the clearest technical expression of Jiang's argument to those investors: humanoid developers can start with a reusable model of human movement, then spend robot-specific data and engineering on the final distance between a person in a video and a machine operating in a room. The preview gives developers access to the motion-generation layer. The harder work demonstrated in the launch videos still depends on proprietary adaptation data, controllers and robot hardware.