GPT-6 Astra controls dexterous hands zero-shot at just 5.2% success in 500 trials
stingning · x · 2026-10-27
The HandsOn team evaluated GPT-6 Astra via moveeef, an interface for specifying palm poses and per-finger joint targets, receiving images, robot state, instructions and execution feedback—no demonstrations or fine-tuning. Across 660 trials: Core-50 saw 26/500 successes (5.2%) and Transfer-16 just 3/160 (1.9%). Saved action notes show task-aware planning even in failed trials. This is a separate controller setting with different observations, trial counts, and model-call budgets from the trained policies.
Related event: GPT-6 Astra Scores Only 5.2% on Zero-Shot Dexterous Manipulation Benchmark(2 posts)→
More from Embodied
- AnyRank ranks robot policies reliably with far fewer real-robot trials — pappasg69 · 2026-10-27
- Podcast: Is Tesla Building a Robot Army? The Optimus Legion Theory — MarwaEldiwiny · 2026-10-27
- Full Cybercab teardown video live: Tesla team walks through the engineering — elonmusk · 2026-10-27
- Open tactile skins like Johns Hopkins' OpenTac ease robot touch design, durability is next bottleneck — _krishna_murthy · 2026-10-27
- Farming robotics doc series debuts: five machines, one operator, wine grape growers signed — Scobleizer · 2026-10-27
- Agent Buddy: A Desk Gadget That Shows What Your Background AI Coding Agents Are Doing — DanWahlin · 2026-10-27