2.8 hours of inference for 30 seconds of walking: GPT-6 Astra as a robot policy tested

Willing_Amphibian825 · reddit · 2026-11-13

Are frontier multimodal models about to make dedicated robot models (VLAs, world models) obsolete? A Galbot test says not yet: using GPT-6 Astra as a direct robot controller, a 30-second simulated locomotion run needed 250 inference calls averaging 40s each — roughly 2.8 hours of model time — and 1.132B tokens across 50 RoboDojo instances; none of five attempts on one obstacle course reached the goal. A separate academic group got 22.48% average success zero-shot across all 42 RoboDojo tasks (above every public policy on the leaderboard), but the authors call it "a heavily imbalanced policy" — good on semantic tasks, weak on precision and contact-rich work. Meanwhile dedicated models actually run on robots: NVIDIA's GR00T 2 (reportedly 400k hours of egocentric human video, weights expected December), Apache-2.0 LingBot-VLA 2.0 (20 body types from 17 manufacturers, company figures), and open-weight pi0.5. Author's read: fast embodied models handle motor control locally, frontier models plan on top — though the "just scale the general model" camp still has Sutton's Bitter Lesson.

Original post →

2.8 hours of inference for 30 seconds of walking: GPT-6 Astra as a robot policy tested

More from AGI Musings