Rhoda AI has developed a robot whose control model was trained on millions of internet video clips to learn physics for robotics tasks.
This approach represents a shift in how machines acquire motor skills. By leveraging existing public data, the company aims to bypass the slow process of collecting specialized data in controlled laboratory settings.
The system utilizes a model called Direct Video Action. This technology allows the robot to observe how objects move and interact in the real world by analyzing footage from the internet. Rhoda AI said that using massive amounts of publicly available video can teach robots physics more efficiently than traditional methods.
There are varying reports on the exact scale of the training data. One source said the robot was trained on 100 million videos [1], while another describes the dataset as consisting of hundreds of millions of clips [2]. This scale allows the AI to encounter a vast array of physical scenarios and object interactions that would be difficult to replicate in a lab.
Traditional robotics often relies on teleoperation or meticulously scripted movements. By contrast, the Direct Video Action model attempts to generalize physical laws from visual observation. This method allows the robot to understand the properties of materials, and the mechanics of movement, through passive observation of human-centric videos.
The company believes this scalability is the key to creating more versatile robots. By absorbing the visual logic of the physical world from the web, the AI can potentially adapt to new environments without requiring thousands of hours of manual training for every new task.
“Rhoda AI believes that using massive amounts of publicly‑available internet video can teach robots real‑world physics”
This development signals a move toward 'foundation models' for physical movement, similar to how large language models were trained on the written web. If robots can learn physics from observation rather than manual programming, the speed of deployment for general-purpose robotics could increase significantly, reducing the cost and time required to train machines for diverse real-world environments.
