ICLR 2026

Self-Improving Skill Learning

A robust skill-based meta-reinforcement learning framework for learning and refining reusable skills from noisy or suboptimal offline demonstrations.

Sanghyeon Lee · Sangjun Bae · Yisak Park · Seungyul Han

Abstract

Meta-RL enables fast adaptation to unseen tasks, but long-horizon problems remain challenging when the underlying skill library is learned from noisy or suboptimal offline demonstrations. SISL addresses this issue by iteratively improving skills with decoupled high-level and skill-improvement policies. It also uses maximum return relabeling to prioritize task-relevant trajectories, leading to more stable skill learning and robust adaptation across long-horizon tasks.

Citation

@inproceedings{lee2026sisl,
  title     = {Self-Improving Skill Learning for Robust Skill-based Meta-Reinforcement Learning},
  author    = {Lee, Sanghyeon and Bae, Sangjun and Park, Yisak and Han, Seungyul},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year      = {2026},
  url       = {https://openreview.net/forum?id=g6MncErbEb}
}