NVIDIA and Universities Introduce ASPIRE Self-Improving Robotics Framework

NVIDIA worked with researchers at the University of Michigan, UIUC, and UC Berkeley to release ASPIRE. The framework lets robots generate and refine their own control code across repeated attempts.

ASPIRE coordinates a programming agent, a closed-loop executor, a growing skill library, and an evolutionary search process. On the LIBERO-Pro benchmark the system raised success rates by as much as 77 points over the prior best baseline.

The same approach lifted bimanual handover success in Robosuite from 20 percent to 92 percent and radio pickup performance in BEHAVIOR-1K from 56 percent to 88 percent. When skills learned on LIBERO-90 were applied zero-shot to longer LIBERO-Pro tasks, ASPIRE reached roughly 31 percent success while earlier methods remained near 4 percent.

How the coordinator and executor split work

The programming agent, built on Claude Opus 4.6 with a one-million-token context window, writes Python control scripts. The executor runs each script on the robot and returns sensor traces plus task outcome. When a script fails, the agent receives the full trace and must propose a revised version.

The skill library stores every successful script along with its measured success rate. Evolutionary search ranks stored scripts, recombines useful segments, and mutates parameters to produce new candidates. Only scripts that pass closed-loop verification enter the library.

This loop continues without human-written code updates. Each cycle adds either a new skill or an improved version of an existing one.

Performance numbers across three benchmarks

On LIBERO-Pro the largest recorded lift versus the strongest reported baseline reached 77 points. The improvement appeared on long-horizon tasks that combine multiple sub-goals and require precise spatial reasoning.

In Robosuite the bimanual handover task moved from 20 percent success to 92 percent after the agent reused and refined two-handed grasp sequences stored in the library. In BEHAVIOR-1K the radio pickup task climbed from 56 percent to 88 percent once the agent discovered consistent drawer-opening and radio-removal primitives.

Zero-shot transfer from LIBERO-90 skills produced a 31 percent success rate on LIBERO-Pro Long tasks. Earlier methods that lacked the closed-loop refinement step stayed near 4 percent on the same suite.

Where prior approaches reached limits

Earlier robot learning systems typically trained a single policy or relied on a fixed set of human-authored primitives. Once the policy encountered an unseen sequence of sub-tasks, performance collapsed. ASPIRE avoids that wall by treating the control program itself as the object of optimization.

Because the agent can read full execution traces and edit code directly, it can recover from failures that defeat gradient-based policies. The skill library supplies reusable building blocks that accumulate across tasks rather than resetting with each new scene.

Remaining questions on scaling and verification

It is still unclear how far the approach extends when task diversity grows beyond the current benchmarks. The evolutionary search step consumes multiple trials per improvement, and the cost of those trials in real hardware remains unmeasured at larger scale.

No independent lab has yet reproduced the full pipeline on new robot platforms. The reported numbers rest on the environments and task definitions supplied by the original authors.

Future work will need to track how quickly the library grows and whether success rates plateau once the most frequent failure modes have been addressed.