Rhoda AI Raises $450 Million to Teach Robots Physics Using Internet Video

A robotics startup has emerged with millions in funding to tackle a major hurdle in physical artificial intelligence. While language models quickly advanced by learning from the entire text of the internet, roboticists lacked a similar dataset for the physical world. Many past lab demonstrations failed to work reliably in messy real-world settings because they relied on small data collections.

Training on Internet Video

Rhoda AI uses massive amounts of internet video to pretrain its software. Instead of focusing only on clips of human hands or specific tasks, the system watches diverse footage like sports, waves, and trees in the wind. This broad approach teaches the model basic physics, motion, and physical interaction before it ever encounters a specific robotic task.

How the Direct Video Action Model Works

The company deploys a system called a direct video action model, which combines movement and manipulation. The software analyzes past video frames and predicts what will happen in the scene over the next fraction of a second. An action converter then translates that visual prediction into commands for a robot. After the machine executes the movement, the system captures a new observation and repeats the loop.

Real-World Deployment

After the initial training phase, operators add a small amount of task-specific training to adapt the robot to a particular job. The technology is designed to work across different hardware bodies and is currently focused on solving labor shortages in manufacturing, logistics, automotive, and e-commerce.