$ cat wiki/papers/2026/2607.00272-aspire-robot-skills.md
ASPIRE: Agentic Skills Discovery for Robotics
TL;DR
A continual learning system (NVIDIA GEAR Lab + UMich/UIUC/Berkeley/CMU) that lets robots autonomously write and accumulate reusable control programs. As the skill library grows, the robot generalizes zero-shot to long-horizon unseen tasks — +670% on LIBERO-Pro Long vs. prior SOTA with test-time reasoning.
Authors & Org
- NVIDIA GEAR Lab (Jim Fan's group)
- University of Michigan
- UIUC
- UC Berkeley
- Carnegie Mellon University
Jim Fan (@DrJimFan) announced on X: https://x.com/DrJimFan/status/2072004192294830583
Method
ASPIRE operates in an open-ended learning loop with three components:
- Closed-loop execution engine — generates fine-grained multimodal traces (visual + proprioceptive + language) that expose why a skill failed
- Continually expanding skill library — distills validated fixes into transferable, cross-embodiment skills; the library persists across different robot hardware
- Evolutionary search procedure — explores diverse task sequences and control programs, self-repairs from execution trace diagnostics
The system uses a code-as-policy paradigm: skills are Python programs, which enables programmatic composition, inspection, and versioning.
Results
| Benchmark | Improvement over prior SOTA |
|---|---|
| LIBERO-Pro manipulation (under perturbation) | +77% |
| Robosuite bimanual handover | +72% |
| BEHAVIOR-1K long-horizon household | +32% |
| LIBERO-Pro Long (zero-shot) | 31% vs 4% (+670%) |
| The zero-shot result is the most significant: ASPIRE's accumulated skill library allows generalization to unseen long-horizon tasks without any additional training, whereas prior methods (which used test-time reasoning + retries) achieved only 4%. |
Significance
Physical scaling law: parallels the data-flywheel in language models — the more tasks a robot attempts, the more transferable skills it accumulates, and generalization improves with skill-library depth. This is empirical evidence that continual robot learning compounds.
Code-as-policy advantage: writing skills as code (rather than learned weights or trajectories) makes skills composable, inspectable, and debuggable — a prerequisite for safe deployment in real environments.
Cross-embodiment claim: the paper claims skills transfer across different robot hardware. If this holds, a centrally maintained skill library could serve a heterogeneous fleet — reducing per-platform retraining costs.
Open Questions
- Do gains hold on hardware not seen during training? (Cross-embodiment claim needs broader testing)
- How does skill-library depth scale? Does performance plateau or continue improving?
- Computational cost of the evolutionary search procedure at deployment time?
Cite
@article{aspire2026,
title={ASPIRE: Agentic Skills Discovery for Robotics},
author={...},
journal={arXiv:2607.00272},
year={2026}
}
Related
- Embodied Agents — the broader category this paper advances
- ENPIRE: Agentic Robot Policy Self-Improvement in the Real World — related agentic robot self-improvement paper (NVIDIA/CMU/Berkeley, June 2026)
- Jim Fan — lead researcher at NVIDIA GEAR Lab
- NVIDIA — institutional affiliation