Light Origins Launches Light-O1 Model for Robot Tasks

Light Origins announced in a press release the launch of Light-O1, a general purpose embodied foundation model for physical AI. The model learns a reusable human action prior from structured actions recovered from internet videos, then adapts it to different robots and tasks.

The company tested six pretraining budgets ranging from 3.75 billion to 120 billion multimodal tokens. It said larger budgets consistently reduced next action prediction loss and whole body pose prediction error after adaptation across human data, Unitree G1 robot data and its own LightBot data.

Demonstrations showed robots opening a shoe cabinet, placing slippers inside, collecting moved trash items, handing over a towel and wiping a table. Light Origins also released Light-O1-Preview, a reasoning text to action model, along with model weights, code and a public playground.