Overview of NVIDIA 'Astranomy' - Achieving a 74.8% Success Rate in Robot Control with GPT-6 Astra
Overview of NVIDIA 'Astranomy' - Achieving a 74.8% Success Rate in Robot Control with GPT-6 Astra
NVIDIA's 'Astranomy' research investigates the use of GPT-6 Astra for robot control without additional training, achieving a 74.8% success rate on the RoboLab-120 benchmark. The study evaluates three control methods—Direct, Agentic, and Orchestrated—demonstrating the impact of AI agent history management on performance. Challenges include inference speed, API costs, and stability in object manipulation.
NVIDIA's 'Astranomy' research explores using GPT-6 Astra for robot control without additional training, achieving a 74.8% success rate on RoboLab-120. The study compares three control methods—Direct, Agentic, and Orchestrated—highlighting the effectiveness of history management in AI agents and the challenges of real-time inference and API costs.
Source: Robotics — Robot learning news discovery · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
74.8%
View original evidence
Overview of NVIDIA 'Astranomy' - Achieving a 74.8% Success Rate in Robot Control with GPT-6 Astra
Open source S1Reported success rate
74.8%
View original evidence
ranomy provides GPT-6 Astra with camera images, robot state, and task instructions to generate actions directly. In the robot manipulation benchmark ' RoboLab-120 ', it achieved a 74.8% success rate , outperforming the comparative robot-specific models. The main features are as follows: ・Robot control without additional training Directly generating robot actions with GPT-6 Astra ・Three types of control methods Compar
Open source S4
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
s based on new observations This configuration delegates basic movements to a specialized model while a general-purpose AI oversees the progress of the entire task. 3. Benchmark Results Astranomy uses ' RoboLab-120 ', which includes 120 types of object manipulation tasks. The tasks include operations such as placing objects into containers, moving multiple objects in sequence, and positioning them at specified locati
Open source S10
Overview of NVIDIA 'Astranomy' - Achieving a 74.8% Success Rate in Robot Control with GPT-6 Astra
Open source S1
ranomy provides GPT-6 Astra with camera images, robot state, and task instructions to generate actions directly. In the robot manipulation benchmark ' RoboLab-120 ', it achieved a 74.8% success rate , outperforming the comparative robot-specific models. The main features are as follows: ・Robot control without additional training Directly generating robot actions with GPT-6 Astra ・Three types of control methods Compar
Open source S4
ons. The results of executing each task five times in the DROID robot simulation environment are as follows: ・GPT-6 Astra (Agentic): 74.8% ・GPT-6 Astra (Orchestrated): 66.5% ・GPT-6 Astra (Direct): 66.3% ・Lakeside: 51.0% ・Cosmos3-Nano-Policy: 36.8% ・π0.5: 28.0% The highest success rate was recorded by the Agentic Policy using Codex. It outperforms Lakeside, the best of the comparison specialized models at 51.0%, by 23
Open source S11
Source：Robotics — Robot learning news discovery · note.com