NVIDIA Open-Sources ASPIRE 1.0, Where Coding Agents Teach Robots New Skills, Including a Rubik’s Cube
NVIDIA’s GEAR robotics lab released ASPIRE v1.0 under Apache 2.0. Instead of training a neural policy, an AI coding agent writes robot control programs, watches them run, fixes failures and saves what works as reusable skills. The new release adds a real two-armed robot solving a physical Rubik’s Cube.
By 1EyeAI Desk ·
ASPIRE comes from NVIDIA’s GEAR lab, co-led by Jim Fan and Yuke Zhu, with collaborators from several universities. It first appeared as an arXiv paper in June; the v1.0 code release, dated Oct. 16, is Apache 2.0 licensed. The loop has three parts: an execution engine that logs perception, grasp and motion traces so the agent can find where things broke; a growing library of validated skills; and an evolutionary search over candidate programs.
NVIDIA’s claims: with Claude Opus 5.5 driving it, ASPIRE reaches 99% success across 80 LIBERO-Pro tasks and 97–100% on selected Robosuite and BEHAVIOR-1K tasks. On a 10-task subset, success swings from 15% to 99% depending only on which language model writes the code. These are the team’s own reported results, not independent benchmarks.
On real hardware, a two-armed YAM robot running the agent (GPT-6 Astra in NVIDIA’s demo) reconstructs all six cube faces from its cameras, executes layer turns and reorientations, and sets the solved cube down. The move sequence comes from a standard symbolic cube solver; what ASPIRE handles is the physical manipulation. Other demos include unscrewing a bottle cap, putting a pill in a bottle and stacking cups.
NVIDIA lists the limits itself on its project page: it is not yet a fully autonomous real-world learner (it still needs success detection, safe resets and safety monitoring), it depends on a frozen frontier model, and it is bounded by a fixed set of robot primitives.
What this means for you: If you work with robots or simulators, you can try it today. For everyone else, it is a concrete sign that better chatbots may translate fairly directly into more capable robots, since the same system got far better just by swapping in a stronger model.