NVIDIA study: GPT-6 Astra hits 74.8% on RoboLab-120 as agentic robot policy
ZeeshanZiaML · x · 2026-11-10
NVIDIA's CoRL 2026 empirical study "Astranomy" tests whether GPT-6 Astra can serve as a native robot policy for language-conditioned manipulation. On the full RoboLab-120 benchmark it compares three setups: Direct Policy (raw action generation), Agentic Policy (Astra as a Codex agent with persistent context and tools), and Orchestrated Policy (Astra monitoring while a robot foundation model acts).
Key results: the Agentic Policy achieves 74.8%, showing Astra can handle generalist pick-and-place tasks — but inference averages 18.56 seconds per call, making latency and token cost the central practical bottleneck. The paper also ablates action representation, geometry info, observation history, and task specification across all three policy types. The poster confirms Astra is a VLM trained on robot-specific data, with headline results likely coming from a tool-equipped agentic harness.
More from Embodied
- Johns Hopkins lab hiring postdocs in embodied intelligence, applications due Dec 11 — mangahomanga · 2026-11-10
- HARP-Bench: learned robot policies score 6-19% on 37 tasks involving humans — gan_chuang · 2026-11-10
- Prediction: a major lab will solve robotic intelligence by end of 2027 — alexmacgregor__ · 2026-11-10
- One Euler step suffices: VLAs get 1.6-4.8x faster with no retraining — kaixhin · 2026-11-10
- 6-8K CAD Local LLM Rig: M5 Ultra 96GB, M5 Max 128GB, or a Big-RAM Custom PC? — EfficientDivide1572 · 2026-11-10
- ODE Policy learns action dynamics via neural ODEs for multimodal robot trajectories — ntumars · 2026-11-10