Summary

ASPIRE (Agentic Skill Programming through Iterative Robot Exploration) is a continual learning system that autonomously writes and refines robot control programs in a code-as-policy paradigm while compounding experience into a reusable skill library. Three components operate in an open-ended loop: (1) a closed-loop robot execution engine with multimodal traces enabling autonomous failure diagnosis and repair, (2) a continually expanding skill library distilling validated fixes into reusable transferable knowledge, and (3) evolutionary search generating diverse task sequences to explore beyond single-trajectory refinement.

Key Contributions

  • Closed-loop execution engine with multimodal trace collection for automated failure diagnosis and repair synthesis
  • Continually expanding skill library that distills validated code fixes into reusable primitives
  • Evolutionary task sequence generation for broad exploration
  • 77% improvement over prior methods on LIBERO-Pro under perturbation
  • 72% improvement on Robosuite bimanual handover
  • 32% improvement on BEHAVIOR-1K long-horizon household tasks
  • Zero-shot generalization to unseen long-horizon tasks: 31% vs. 4% for prior methods on LIBERO-Pro Long
  • Collaboration: NVIDIA, UMich, UIUC, UC Berkeley, CMU

Significance

ASPIRE represents a major step toward autonomous robot skill acquisition—a system that improves itself through exploration, failure analysis, and skill reuse, with dramatic gains on challenging long-horizon manipulation benchmarks from top research institutions.