Inside Rhoda AI: Teaching Robots From Internet-Scale Video

For decades, autonomous-looking robots performed well in demo reels but never reached the factory floor. People still don't have humanoid robots at home. After all the backflip demos, the missing piece wasn’t the hardware, rather an AI model good enough to handle edge cases in the real world. Rhoda AI extracts physical intelligence from internet-scale video and uses it to teach robots tasks that customers thought were too difficult to automate. Training a model like this meant doubling compute at a critical moment, adding petabytes of storage next to dedicated GPUs, and finding an infrastructure partner reliable enough to stop thinking about. Hear from Rhoda's technical leadership about what it took to change the way robots learn and reach real customers. Chapters 0:00 — Demos vs. reality 0:24 — Robotics is a data problem 0:43 — Why teleoperation breaks down 1:15 — The data pyramid 1:43 — Internet-scale video, doubling compute 2:26 — Kubernetes, Slurm, and petabyte storage 3:24 — Robots that ship to real customers Links Rhoda AI: https://www.rhoda.ai/ Nebius AI Cloud: https://nebius.com/ See how other teams build on Nebius: https://www.youtube.com/playlist?list=PLL30rV4wd9h4