Jim Fan's lab open-sources ASPIRE after Astra swap
The benchmark jump came from a stronger backbone model, not from new robot code.
Jim Fan's lab has open-sourced V1.0 of ASPIRE, an agentic robotics framework, with what Fan calls full real robot skill discovery and self improvement capabilities. The headline detail is how the numbers got there. On the day GPT-6 Astra arrived, the lab switched the backend model and, in Fan's words, saturated all the benchmarks to around 99% overnight.
A rising frontier tide lifts all boats, but Astra was a 0->1 tsunami on many tasks.
On the day Astra arrived, our lab switched the backend model and saturated all the benchmarks to ~99% over night.
Guanzhi Wang posted the breakdown. 99% on LIBERO-Pro across 80 tasks, 100% on Robosuite across 7 representative tasks, and 100% on BEHAVIOR-1K across 2 representative tasks. Wang also lists real-world robot manipulation with Astra and continual skill refinement on physical robots as progress since the first release.
What actually moved
Read those three lines carefully. The saturated scores are simulation benchmarks, and two of the three sets are tiny, 7 tasks and 2 tasks, picked as representative. The jump came from a model swap, not from rewriting the robot stack. Fan says ASPIRE and its /robot skill library landed in June, before the big models could reliably do robotics, which is the point he is making. The framework stayed put and the backbone got better.
That is the part worth sitting with. Fan calls it the Physical Bitter Lesson, the robotics version of the old argument that general methods riding compute beat hand-built structure. His own framing is that a rising frontier tide lifts all boats, but that Astra was a step change on many tasks rather than a gentle lift.
Model shopping
Fan also posted a quick comparison across frontier models on the lab's robot simulation bench. Opus 5.5 slightly edges out Astra. Sonnet 5.5 sits on the Pareto frontier of cost versus robot capability at $12 per task, which is the first time anyone in this orbit has put a per task dollar figure on running an agentic robot policy off a frontier model.
Shown versus claimed. The near perfect numbers are in simulation. The real robot manipulation and continual skill refinement are listed as capabilities in the release, and the sources do not give success rates for either on hardware. The code and project site are both public, so anyone can check the sim results themselves.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.