Rhoda AI bets $450 million on video-trained robot models

Rhoda AI's pitch rests on a comparison to how large language models advanced: ChatGPT and its peers were trained on essentially every word ever put on the internet, while physical AI has lacked an equivalent dataset. According to CEO Jagdeep Singh, prior efforts to collect robot data by teleoperation produced lab demonstrations — robots folding T-shirts or making coffee — but nothing robust enough for the real world. The company's answer is to pretrain on hundreds of years of video, from baseball games to waves crashing, so the model picks up a general sense of how the world works before any task-specific tuning.

The model works in a loop: it observes a history of video frames and action tokens, predicts the next half second of video, converts that scene into robot actions, executes them, and then re-observes. Rhoda AI says the broadest training data — not just human-centric manipulation clips — produced the best-performing models, which is also what lets them generalize across different robot bodies. The hardware only enters the picture during post-training, when a small amount of task-specific data teaches the workflow.