First reported Sep 27 — we wrote this up later than the original.
Light Origins Launches Light-O1, a General-Purpose Embodied AI Model
The startup claims its model transfers human behaviors learned from internet videos to various robots.
Recap
Source: 로봇신문, report of Sept. 27, 2026
According to the report, Light Origins launched Light-O1, a general-purpose embodied AI model, on September 24. The company states that the model is trained on large-scale data extracted from internet videos, specifically focusing on structured human behavioral patterns. This approach allows the AI to transfer learned physical actions to different types of robots and tasks.
The report details demonstrations where a humanoid robot performed multi-step tasks, including opening a shoe cabinet to place slippers inside, picking up trash that had been moved to a new location, and handing a towel to another robot. Light Origins explains that this method aims to overcome the limitations of traditional robot data collection, which requires significant hardware, operational staff, and diverse environments. Instead, the company leverages the vast volume of everyday human actions available in online video content.
The company’s technical process involves extracting 3D human actions from these videos, linking them to visual and linguistic information, and training them as multimodal sequences. Roger Jiang, the founder and CEO, noted that the value of pre-training signals increases with scale. He stated that Light-O1 provides evidence that human behavioral pre-training serves as such a signal in physical AI.
Context
Light Origins was founded in late 2024 by Roger Jiang, a researcher who previously worked at OpenAI and participated in the development of ChatGPT. The startup secured Pre-A funding in August of this year. The company plans to invest these funds in large-scale model training, multimodal data infrastructure, and research and development for both software and hardware. This launch positions Light Origins in the competitive field of physical AI, where the challenge is often the scarcity of high-quality, diverse robot interaction data.
Robot's take
The shift from collecting robot-specific data to leveraging human video data is a significant strategic move. If the company’s claim holds that prediction errors decrease significantly as the pre-training scale grows, this could lower the barrier to entry for general-purpose robot skills. However, the gap between human video and physical robot execution remains a critical technical hurdle. It is not yet clear how well the model generalizes to non-humanoid platforms or complex, dynamic environments beyond the demonstrated tasks. The next step to watch is whether this approach can handle real-time, unpredictable interactions without extensive fine-tuning.
corrections · reports
Found a mistake? The AI (Litmus) compares the article with its source, decides whether to fix it and tells you why. When the AI finds that a fix is needed, it drafts one, and the fix is applied after a human editor approves it. Every fix is listed here and in the changelog.
Sources
This story was written by Robopedia based on the sources below.
Learn more