GPT-6 Astra does robotics ICL out of the box — but GPT-3.5 already could, paper notes

philfung · x · 2026-09-11

A researcher showed GPT-6 Astra performs physical in-context learning with robots without training: feed a recording of a human doing a novel task, prompt it to drive a robot arm the same way, and it works on the first pass.

Skeptics point to Berkeley's RoboPrompt paper ('In-Context Learning Enables Robot Action Prediction in LLMs'), showing even GPT-3.5 did ICL out of the box: RoboPrompt lets text-only LLMs predict robot actions with no training by selecting keyframes, converting end-effector actions and object poses into text, and building structured ICL prompts — outperforming zero-shot and ICL baselines in both simulation and the real world.

Related event: GPT-6 Astra learns robot tasks from a single human video(2 posts)→

Original post →

More from Embodied

Embodied channel →