GPT-6 Astra on real robots: competence yes, practical intelligence not yet
Why it matters
In a technical report posted Oct 15 (arXiv 2610.22471), SenseNova and a joint research team of 60 authors tested GPT-6 Astra on 12 robot benchmarks and 8 real robots. Their verdict: embodied competence, largely yes; practical embodied intelligence, not yet. With in-context learning from training demos and up to 10x the official step limit, it scored above 95% on the LIBERO family and SIMPLE and around 90% on VLN-CE; zero-shot on RoboLab (10x budget) it reached 69.2% vs 42.9% for the best listed baseline. The harness mattered on RoboDojo: an unreviewed robot policy scored 5.83%, GPT-6 Astra reviewing four actions at a time 25.83%, and adding in-context learning 54.17% at about 6.9x the tokens. On real hardware, with no platform-specific training, one model drove all 8 robots: rigid pick-and-place 17/18, cloth folding 2/4, long-horizon navigation and search 9/23, precise plug insertion 0/11. Costs are high (about 1M tokens and nearly an hour per LIBERO-Long++ episode), and in a prototype home with a running world clock it turned off a burner about 90 s late and the egg charred. Caveats: technical report, not peer-reviewed; code release listed as coming soon.
Read the source: arXiv + SenseNova project page (opens in a new tab)
All hits from 19 Oct 2026More RoboticsGet the daily hits by email