NVIDIA report: agentic setup lets GPT-6 Astra hit 74.8% on RoboLab-120 robot manipulation
YuXiang_IRVL · x · 2026-10-15
NVIDIA's tech report "Astranomy" empirically evaluates GPT-6 Astra as a native policy for language-conditioned manipulation on RoboLab-120 across three roles: Direct Policy (raw action generation), Agentic Policy (Codex as agent with persistent context and tool use — best at 74.8%), and Orchestrated Policy (Astra delegates action generation to a robot foundation model). Ablations cover action representations, geometry info, observation history, and task specification. Astra can natively handle general pick-and-place tasks, but averages 18.56 seconds per inference call, making latency and token cost the central obstacles to real-time closed-loop robot control.
More from Embodied
- Robotics will recapitulate AI: robot firms profit by replacing you, not selling to you — broodsugar · 2026-10-15
- Dyna Robotics' Napkin-Folding Robots Win Bigger Din Tai Fung Contract via Repeat Orders — JasonMa2020 · 2026-10-15
- Sipeed shows $50 stereo depth camera with NPU outputting depth maps and point clouds — AIFlow_ML · 2026-10-15
- IRVL is evaluating frontier AI models for manipulation, report coming soon — YuXiang_IRVL · 2026-10-15
- Rhoda unveils fingertip-like tactile sensor letting robots identify objects while blindfolded — CSProfKGD · 2026-10-15
- Infinite Machine's Olto e-scooter spotted roaming NYC streets twice in 4 hours — soleio · 2026-10-15