Nvidia’s GEAR lab has released ASPIRE 1.0 as open-source software for controlling robot arms with code written by a large language model. Startup Fortune reports a 99% success rate across 80 manipulation tasks, including solving a physical Rubik’s Cube, offering a concrete glimpse of AI moving from chat windows into physical work.
NVIDIA Watch analysis
What happened
ASPIRE is described as a code-as-policy system: an AI model writes, runs and debugs Python control code for a robot arm, rather than learning each skill through trial and error. The report says the framework is available on GitHub under the Apache 2.0 licence.
Startup Fortune reports the 99% result across 80 manipulation tasks and says the system solved a physical Rubik’s Cube. It also reports that switching the system’s backend to the Astra model brought benchmark performance to about 99%. These are reported benchmark results, not a guarantee that a robot will manage every task or setting equally well. Read Startup Fortune’s report.
Why it matters
Writing and debugging control code gives a model a route from a requested task to actions a robot can execute. That is a different proposition from asking a chatbot for instructions: the output has to work in the physical world, where a mistimed movement is more than an awkward answer.
Open-sourcing the system under Apache 2.0 could also make it easier for researchers and developers to inspect and adapt the framework. The reported results are promising, but the 80-task benchmark is a bounded test, not evidence of general-purpose robotic competence.
Our read
The interesting part is not the Rubik’s Cube’s long career as a technology-news prop. It is the combination of AI-written control code and reported success across a set of physical manipulation tasks. If developers can reproduce and build on the work, ASPIRE could be a useful tool for exploring how robots acquire new skills. The next test is whether that performance travels beyond a benchmark and into messier, less obliging tasks.
What to watch
- Whether the released code and benchmark details let others reproduce the reported results.
- Which robot setups and manipulation tasks ASPIRE supports in practice.
- How performance changes across models and tasks outside the reported benchmark.
Discussion spark: Should robot-control systems built around AI-written code be open by default, or does access need to be limited until their reliability is clearer?
Sources and evidence
- Nvidia's open-source ASPIRE system taught a robot to solve a Rubik's Cube (16 October 2026, 18:38 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.