RoboBoom
Light Origins· September 21, 2026· 6:26

Introducing Light-O1#humanoid #robots #embodiedai

Watch on YouTube

Light Origins’ presenter explains how Light-O1 uses human video to train robot action models and demonstrates it on a Lightwheel setup and a Unitree G1.

In this talk

  1. 2:24

    Two robots share one home

    The team pairs Lightwheel, a video dataset, with a Unitree G1 to perform household chores. Both robots use Light-O1, a pretrained model built by Light Origins, and the demonstration was recorded in one continuous take.

  2. 2:53

    Learning actions from human video

    Light Origins converts everyday activity videos into robot-training data. It reconstructs actions, aligns them with video and language, and trains a model to predict what action should follow. The team says larger-scale pretraining on human activity improves the accuracy of robot action predictions during adaptation.

  3. 3:51

    Adapting the model to each robot

    A behavioral foundation model decides and coordinates actions for a particular robot. Light Origins uses a shared training framework, then post-trains Light-O1 with data from each robot so it can move and perform tasks in that robot’s body.

  4. 4:41

    Testing task and object changes

    The demonstration includes two cabinet-opening tasks, with the robot retrying and continuing a task. In another example, Light-O1 handles a water bottle, then the team changes the object to a cola can and moves both the can and trash can to different positions.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.

More talks