Notes on Physical AI
We gave Codex and Claude Code control of a Vega 1 Pro and recorded what happened between the task description and the result: what the agents build, where their time goes, and where they still fall short.
How Dexmate robot and Omniteleop help you collect teleoperation data for physical AI model training
A new benchmark from Dexmate and Carnegie Mellon asks which kind of memory a robot agent should keep, and finds that saving the camera frame beats describing it.
What Industrial Physical AI actually requires, and why most robots aren’t ready.