Nvidia's Astranomy study reaches 74.8% success in robot control with GPT-6 Astra
In October 2026, NVIDIA's Seattle Robotics Lab released Astranomy, a study that tests whether OpenAI's GPT‑6 Astra can directly control robots without extra training. The system feeds camera images, robot state, and task instructions to the model and generates actions via API calls. Using the RoboLab‑120 benchmark of 120 manipulation tasks in the DROID simulation, Astranomy achieved a 74.8%…
Key points
- Astranomy achieved 74.8% success on the RoboLab-120 benchmark using GPT‑6 Astra, beating the best specialized model by 23.8 points.
- Agentic Policy with Codex used about 30.9 million tokens per successful episode, costing an estimated $36.40.
- Direct Policy inference took 18.56 seconds per call; simulation latency does not reflect real‑world speed constraints.
The research compared three control approaches. The Agentic Policy, which runs GPT‑6 Astra inside a persistent Codex session, reached the highest success at 74.8%, using about 30.9 million tokens per successful episode and costing an estimated $36.40. Direct Policy, which calls the model without history, scored 66.3% and required 18.56 seconds per inference. Orchestrated Policy, which supervises NVIDIA’s Cosmos3‑Nano‑Policy, achieved 66.5%, showing little gain from combining a specialized model. Motion‑representation tests showed EEF methods outperform joint‑position and B‑spline formats, especially under the Agentic setup.
Authors note the results are limited to simulation; real‑world latency and API costs could hinder high‑speed operation, and performance on assembly or rapid‑change tasks remains untested.
Model page: GPT-6 Astra →
The story so far
3 episodes →- Nvidia's Astranomy study reaches 74.8% success in robot control with GPT-6 Astrathis story
Overview of NVIDIA 'Astranomy' - Achieving a 74.8% Success Rate in Robot Control with GPT-6 Astra
note.com · 20 October 2026
Loading the full article…
This text was published by note.com and written by npaka. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- OpenAI claims AI-generated proof of quasi‑Riemann hypothesis · 2 src
- OpenAI's GPT-5.6 Sol and Google's Gemini 4 Argon find critical PHP flaw · 2 src
- PHP discloses PHP2Shell vulnerability affecting up to 70% of websites · 8 src
- Ggml-org adds fused top‑k mask operation to llama.cpp · 1 src
- Microsoft Research says Xinzhi Zhang and Zhao Song prove KLS conjecture with AI · 1 src
Comments
via GitHub Discussions