GPT-6 Astra does physical in-context learning: drives a robot arm from one human demo
ChongZzZhang · x · 2026-09-10
A developer reports that GPT-6 Astra performs physical in-context learning out of the box: dropping a recording of a human doing a novel task into the Codex app and prompting it to drive a robot arm the same way worked on the first pass, no fine-tuning required — a striking zero-shot transfer result for robot manipulation.
More from Embodied
- Apple A20 Pro details leak: 32-core Neural Engine, ~115 GB/s memory bandwidth — Balance- · 2026-09-10
- NVIDIA open-sources ENPIRE, letting coding agents self-improve real robot policies — GuanyaShi · 2026-09-10
- Astra shows off its tendon-driven robot hand design — keerthanpg · 2026-09-10
- Foldables remain a 1.5% niche: 4% of Samsung's sales but 16% of its $800+ phones — firstadopter · 2026-09-10
- Scoble shares a robotics insider's parking-lot test: beater cars mean real scaling — Scobleizer · 2026-09-10
- IShowSpeed meets Apple CEO John Ternus and tries the new foldable 'iPhone Duo' first-hand — bugKrusha · 2026-09-10