Briefs · Taipei time (GMT+8)

NVIDIA open-sources ASPIRE v1.0, a robot agent framework where coding agents build up skills as they work

一句重點：With ASPIRE, a coding agent writes a robot's control programs, uses the execution logs to find and fix its own mistakes, and files approaches that pass validation into a skill library. NVIDIA GEAR built it together with several US universities (ASPIRE project page, GitHub).

發生什麼事

Version 1.0, published on GitHub in October, has two headline results: a 99% success rate across 80 LIBERO-Pro tasks, and a Rubik's Cube solved (GitHub). The project page marks GPT-6 Astra as the model used for the real-robot Rubik's Cube (ASPIRE project page).

Which model sits underneath matters a lot: on 10 LIBERO-Pro tasks, Claude Opus 5.5 reaches 99%, while GPT-6 Luna manages only 15% (GitHub).

Correction (2026-10-22)

The original version labeled @DrJimFan's personal post as an NVIDIA GEAR announcement; GitHub and the project page now serve as the primary sources. A cost comparison between Opus 5.5 and Sonnet 5.5 could not be traced to a verifiable source and has been removed.

01Sources

  • ASPIRE project page: NVIDIA GEAR; writing and fixing control programs based on how runs go, skill library, real-robot Rubik's Cube (GPT-6 Astra), list of limitations, team and partner schools
  • GitHub NVlabs/ASPIRE: Apache 2.0; README News marks the v1.0 release in 2026-10, Rubik's Cube, 99% on 80 LIBERO-Pro tasks, 99% (Claude Opus 5.5) versus 15% (GPT-6 Luna) on 10 tasks, Claude Code / Codex agent skills
  • @DrJimFan: personal account of one of the ASPIRE project leads; 10/16 16:15 UTC (00:15 on 10/17, Taipei time); announcing the open-source V1.0

More briefs →