Josef Chen on X: "Agentic Robotics Flips the Build Order" / X

Why it matters

This piece argues that generalist AI models can now write robot control programs directly, upending the strategy of training visuomotor policies on massive human demonstration datasets. At the University of Illinois, a coding agent given wrist force-sensor feedback iterated on its Python until a robot arm could reliably open a transparent drawer, and the finished program kept working without the agent. The author frames this against Hans Moravec's paradox and SayCan's gap between planning and execution, suggesting labs have prioritized the hands over the brains when the build order should be reversed. RoboDojo's benchmark adds weight: GPT-6 Astra, with no robot-specific fine-tuning and a single gripper action, beat every trained policy on the leaderboard, and a harness prompting it to generate code did better still. The stakes are commercial as well as technical, since Physical Intelligence recently raised a large funding round premised on the demonstration-collection approach. The author's rule: when brain and hands communicate in words or code, the brain can be swapped without retraining the hands.