Liangyuan Xinchuang releases Light-O1, a full-body AI model that learns robot actions from internet videos
On September 21, Chinese embodied AI company Liangyuan Xinchuang (Bright Source Innovation) released Light-O1, a full-body intelligent foundation model that learns human actions from internet videos and transfers them to robots. The company claims it is the first to validate cross-embodiment transfer scaling laws for full-body human motion pre-training, using 100,000 action hours of video data. The company also open-sourced the Light-O1-Preview model.
Editorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Cross-source coverage
Reporting timeline
Liangyuan Xinchuang Releases Full-Body AI Model Light-O1 for Robot Learning from Videos
On September 22, embodied AI company Liangyuan Xinchuang (Bright Source Innovation) announced the release of Light-O1, a full-body intelligent foundation model. The model learns from massive human actions in internet videos, combining language and visual information to transfer human action knowledge to robots. Light-O1 is the first to validate cross-embodiment pre-training scaling laws for full-body human actions: more pre-training data leads to more accurate action predictions when adapting to different robot bodies and perspectives. The model expands video-based human action pre-training to 100,000 action hours, using internet video data to supplement costly specialized data collection and restructure data cost structures. In terms of deployment, the company's self-developed humanoid robot LightBot, equipped with Light-O1, can autonomously perform tasks such as handing over towels, tidying slippers, and picking up trash. The company has also open-sourced a general action generation model, Light-O1-Preview. The report is sourced from Time Weekly.
Read sourceLiangyuan Xinchuang Releases Full-Body AI Foundation Model Light-O1 for Robot Learning
On September 21, Chinese media outlet Cailianshe reported that Liangyuan Xinchuang (Bright Source Innovation) has released a full-body intelligent foundation model called Light-O1. The company claims this is the first verification of cross-embodiment transfer scaling laws for human whole-body motion pre-training. The model learns human actions from massive internet video data, combining language and visual information to transfer skills to robots. The announcement states that Light-O1 overcomes bottlenecks such as high teleoperation costs and insufficient diversity in robot training data. The model represents a step toward enabling robots to learn complex human movements from widely available video sources.
Read sourceLiangyuan Xinchuang Releases Full-Body AI Foundation Model Light-O1
On September 21, Liangyuan Xinchuang (Bright Source Innovation) released the full-body intelligent foundation model Light-O1, claiming to be the first to validate the cross-embodiment transfer scaling law for human full-body motion pre-training. The model learns human actions from massive internet videos and transfers them to robots by combining language and visual information. Concurrently, the company open-sourced the Light-O1-Preview model on its official website, with all technical achievements of its 'scaled pre-training, scaled alignment, scaled deployment' three-stage paradigm fully implemented. The report originates from People's Financial Information and was published on East Money's company news section.
Read sourceShow 2 older updatesHide older updates
Liangyuan Xinchuang Releases Whole-Body AI Foundation Model Light-O1
On September 21, Liangyuan Xinchuang (Bright Source Innovation) released the whole-body intelligence foundation model Light-O1, claiming to have validated for the first time the cross-embodiment transfer scaling law of human whole-body motion pre-training. The model learns human actions from massive internet videos and transfers them to robots by combining language and visual information, overcoming bottlenecks such as high teleoperation costs and insufficient diversity. The company stated that its human motion pre-training reached 100,000 action hours, and as data scaled up, the action prediction loss for cross-embodiment and viewpoint adaptation showed a power-law decreasing trend, with whole-body pose errors simultaneously reduced. Concurrently, the company open-sourced the Light-O1-Preview model on its official website, with all technical achievements of its three-stage paradigm—'scaled pre-training, scaled alignment, scaled deployment'—fully implemented.
Read sourceLiangyuan Xinchuang Releases Full-Body AI Foundation Model Light-O1, Open-Sources Preview
On September 21, Liangyuan Xinchuang (Bright Source Innovation) released the full-body intelligent foundation model Light-O1, claiming to have verified for the first time the cross-embodiment transfer scaling law for pre-training on full human body movements. The model learns human actions from massive internet videos and combines language and visual information to transfer to robots. Concurrently, the company open-sourced the Light-O1-Preview model on its official website, stating that all technical achievements of its 'scaled pre-training—scaled alignment—scaled deployment' three-stage paradigm have been implemented. The report originates from Securities Times and was republished by East Money's industry economy channel.