Light Origins ships Light-O1 robot model after quick hundred-million-yuan re-raise

What's the deal? China-based Light Origins has launched Light-O1, its first general-purpose embodied foundation model, which uses human actions recovered from internet video to give robots a reusable starting point for learning new tasks. The company raised a Pre-A round of several hundred million yuan in August 2026.

Why now? The launch follows quickly on the heels of that funding, part of a fast fundraising cadence in physical AI. Light Origins says its data infrastructure now runs at the thousand-GPU scale and processes about 200,000 hours of video each week, up from 12,500 hours six months ago.

How it works: Rather than relying only on robot-generated data, Light Origins extracts structured 3D human actions from internet videos and aligns them with visual information and language. It trained six versions of the same 4-billion-parameter base model on between 3.75 billion and 120 billion multimodal tokens, with the largest run representing about 100,000 hours of human action.

What's the endgame? The company organises its work around three areas: scalable pretraining, alignment, and deployment. Light-O1 covers pretraining; LightNav-0 handles alignment; and Light REACT lets robots adjust their behaviour based on recent physical interactions.

Alongside the model, Light Origins released Light-O1-Preview, a text-to-action model that takes a natural-language instruction and generates a whole-body action sequence. Model weights, code, and a public playground ship with the release. In demonstrations, its LightBot humanoid opened a shoe cabinet, sorted trash, and handed a towel to a Unitree G1Dealroom has a profile for this one. Try Dealroom → that wiped a table.

What could go wrong? Light Origins stressed that its robot results still use target-specific data, and that scaling tests measured held-out next-action and whole-body pose prediction — not real-world task-success rates. Prediction errors declined as pretraining increased, following power-law trends, according to the company.

The signal: The pairing of a large early round with an open model release shows how quickly capital and technical claims are moving in physical AI. Learning robot behaviour from internet video, rather than costly robot data alone, is emerging as a favoured route to scale.

Read more: theaiinsider.tech

Image credit: Light Origins