Physical AI News (Japan)

Physical AI, robotics and compute — reported from Japan.

日本語
· Industry· 日本語で読む

Rhoda AI to use video data to train general-purpose AI models for robots

Rhoda AI plans to develop general-purpose AI models for robots using video data. The company’s co-founder and CEO said its DVA model will require large-scale training data to handle variations in real-world environments. The article does not specify how data linked to robot actions is generated from video or provide performance figures.

What happened

The Robot Report reported on November 13, 2026, that Rhoda AI is developing general-purpose AI models for robots using video data. The company aims to apply its direct video action (DVA) model to real-world robots. Co-founder and CEO Jagdeep Singh said robust models require large-scale training data because conditions and configurations vary from one task to the next in the real world.

Why it matters

In Rhoda AI’s framing of the challenge, large-scale data is key to handling variations in real-world working conditions. How video is linked to robot action and whether robots can reproduce actions when environments change will be important factors in assessing the company’s DVA concept.

Analysis

The report focuses on Rhoda AI’s DVA concept, which uses video as an entry point for training data rather than collecting robot operation data individually. Singh said that environments and configurations vary from task to task in the real world, pointing to a data challenge between demonstrations of actions under specific conditions and models that can handle different conditions. However, the article does not explain how action labels are generated from video or how video is matched to robot actions. Using video alone does not guarantee that it contains enough information for operation or that a model can adapt to unfamiliar environments. We have previously reported on Rhoda AI in Forbes reports on stealth development by Rhoda AI and Genesis AI (October 15, 2025) and Rhoda AI raises $450 million in Series A at a $1.7 billion valuation (April 2, 2026). This report adds a technical challenge to the previously reported business background by covering the company’s approach to developing models using video as a source of training data, rather than its fundraising or the existence of its development efforts. However, the report covers the company’s plans and the CEO’s explanation; it does not provide the volume of video data, training methods, target tasks or evaluation results in real-world settings. The previously reported fundraising total therefore cannot be taken as evidence of the model’s practicality or development results. A key structural question is whether the process can function as an integrated system: ingesting video, linking it to robot actions, executing those actions on physical robots and feeding the results back into subsequent training. The article cites Singh’s view that conventional teleoperation-based data collection has sometimes been limited to narrow demonstrations, but provides no comparative data showing how far DVA addresses that limitation. The article also does not provide enough detail to assess how DVA differs from vision-language-action (VLA) models. Points to watch include the sources of video and rights clearance, methods for converting video into action data, the volume and diversity of training data, reproducibility across different robot configurations and environments, and safety measures in the event of failure. Without this information, it is not possible to assess whether using video has improved real-world generality or autonomous operation.

Sources

Translated from the Japanese original: /news/the-robot-report-rhoda-ai-no-douga-deeta-katsuyou-wo-houdou-2e6b4a. The Japanese version is authoritative if they differ.