How NVIDIA launched ASPIRE on X

NVIDIA has open-sourced ASPIRE v1.0, a framework that lets coding agents write robot control programs, inspect how those programs fail on a robot, and save the repairs that succeed. It is for robotics labs and engineers who already use agents such as Claude Code or Codex and want that same loop on simulated and physical robots. Jim Fan, NVIDIA's director of robotics and co-lead of the GEAR lab, says that switching the backend to GPT-6 Astra pushed their benchmarks to about 99 percent overnight. Claude Opus 5.5 is slightly ahead on the lab's simulation bench, and Claude Sonnet 5.5 is reported at $12 per task. The system, Agentic Skill Programming through Iterative Robot Exploration, builds a library of text and code skills that later tasks can retrieve. A closed-loop engine logs perception, grasp, and motion traces. The agent uses those traces to find the failure, rewrite the program, and confirm the fix by running it again, while an evolutionary search tries several programs in parallel. With Claude Opus 5.5, the project page reports 99 percent average success across all 80 LIBERO-Pro tasks and 100 percent on representative Robosuite and BEHAVIOR-1K tasks. Skills discovered in simulation also transfer onto a different real robot and API, which reduces the programming effort on hardware. v1.0 is the open-source release that adds full real-robot skill discovery and self-improvement to the system first described in June 2026. Public demos on a YAM station include a Rubik's Cube solved through online practice with GPT-6 Astra, table bussing carried over from simulation, and contact tasks such as unscrewing a bottle cap, placing a pill in a bottle, and stacking cups. The repository is Apache 2.0, with separate instructions for simulation experiments and physical hardware. Project leads are Runyu Lu, Yuke Zhu, Jim Fan, and Guanzhi Wang, working with collaborators from NVIDIA, the University of Michigan, UIUC, UC Berkeley, and Carnegie Mellon.