OpenAI's Astra scores 19/20 zero-shot on gripper task, crushing specialized robotics models
adam_karvonen · x · 2026-11-01
Adam Karvonen and Will Depue discuss a striking OpenAI robotics result: the Astra model scored 19/20 zero-shot (no sim training) on a gripper task, while specialized model fable scored just 1/20.
- Karvonen argues basic pick-and-place via a strong LLM is unsurprising: humans can do it with zero practice on a joystick, and it's easy to train in simulation.
- Dexterity tasks (screwing in a lightbulb, controlling 20 DOF in a human hand) remain where specialized models dominate, and he expects them to hold on for now.
- Depue thinks Astra's zero-shot performance is too strong for specialized models to last, suggesting frontier LLMs may end up issuing high-level instructions to specialized models.
More from Embodied
- Naval predicts 4-passenger drones will eventually end commercial air travel, with a long interregnum ahead — naval · 2026-11-01
- Meta Quest game scans your real room and turns it into a haunted house — Scobleizer · 2026-11-01
- RoboAuto: Robot Program Search With Multi-Level Physical Understanding — DJiafei · 2026-11-01
- Developer dresses robot 'Bob' as a ghost to hand out Halloween candy — chrismatthieu · 2026-11-01
- GRASP authors compare with SuperMimic: extracting more optimization signal from sim experience — hankyang94 · 2026-11-01
- Harvard's GRASP refines human demos in physics sim for dexterous hands and Unitree G1 — hankyang94 · 2026-11-01