ASPIRE: AI system achieves 77% better robot skills via continual learning
Robots now write their own code and learn transferable skills across tasks
ASPIRE (Agentic Skill Programming through Iterative Robot Exploration) is a new continual learning system that lets robots autonomously discover, write, and refine control programs. Developed by researchers from institutions including UC Berkeley, Stanford, NVIDIA, and the University of Michigan, ASPIRE operates in an open-ended loop with three components: a closed-loop execution engine that diagnoses and repairs failures using fine-grained multimodal traces; a continually expanding skill library that distills fixes into reusable knowledge; and evolutionary search that generates diverse task sequences to explore beyond single-trajectory refinement.
ASPIRE demonstrates dramatic improvements over prior methods: 77% on LIBERO-Pro manipulation under perturbation, 72% on Robosuite bimanual handover, and 32% on BEHAVIOR-1K long-horizon household tasks. Even more impressively, its accumulated skill library enables zero-shot generalization to unseen long-horizon tasks—31% success on LIBERO-Pro Long vs. 4% for previous approaches despite their use of test-time reasoning. Simulation-discovered skills also show initial sim-to-real transfer, significantly reducing real-robot programming effort across different embodiments and robot APIs. The paper includes 43 pages, 12 figures, and 9 tables of detailed results.
- ASPIRE autonomously writes and refines robot control programs in a code-as-policy paradigm, compounding experience into a reusable skill library.
- Surpasses prior methods by 77% on LIBERO-Pro manipulation, 72% on Robosuite bimanual handover, and 32% on BEHAVIOR-1K long-horizon tasks.
- Zero-shot generalization on unseen tasks achieves 31% success vs 4% for baselines; sim-to-real transfer reduces real-robot programming effort.
Why It Matters
ASPIRE paves the way for robots that automatically learn and transfer skills, reducing manual programming and enabling faster deployment.