NVIDIA GEAR open-sources ASPIRE V1.0: coding agents write and debug robot control code and build a reusable skill library; a GPT-6 Astra-driven dual-arm robot solves a Rubik's Cube, and swapping models moves LIBERO-Pro success from 15% to 99%

Why it matters

【Key points】

Jim Fan, who co-leads NVIDIA's GEAR lab, announced on Oct 16 at 16:15 UTC that ASPIRE V1.0, an agentic robotics framework, is open source (Apache 2.0). Instead of training a neural policy, a coding agent writes robot control code, reads execution traces to find what failed, repairs and re-validates the code, and stores verified fixes in a reusable skill library that speeds up later tasks.

V1.0 adds online skill learning on a real robot: a dual-arm station driven by GPT-6 Astra solves a Rubik's Cube, unscrews a bottle cap, puts a pill in a bottle and screws it shut, and stacks cups upside down. The repo reports about 99% success on 80 LIBERO-Pro tasks plus selected Robosuite and BEHAVIOR-1K tasks. Holding the setup fixed, success on 10 LIBERO-Pro tasks ranges from 15% (GPT-6 Luna) to 99% (Claude Opus 5.5). Fan says the lab swapped in Astra the day it launched and saturated the benchmarks overnight.

【Verifiable】

Project page: https://research.nvidia.com/labs/gear/aspire/

Code: https://github.com/NVlabs/ASPIRE

Paper: https://arxiv.org/abs/2607.00272

Jim Fan's post: https://x.com/DrJimFan/status/2111128797181567466

【Our read】

1. For code-as-policy robotics, the ceiling is mostly the underlying model: same environment, 15% to 99% by swapping models. As frontier models improve, these robots improve with them.

2. The project's own limits matter: real-world use still needs success detection, safe resets and calibration; behavior is bounded by a fixed perception and control API; and the search loop needs many model calls, so it is costly.

3. Near-saturation says LIBERO-Pro-style tests are no longer hard enough, not that general robot skill is solved.