Reproduce “Agentic Skill Discovery” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xufeng Zhao, Cornelius Weber, Stefan Wermter

Researchers at the University of Hamburg developed the Agentic Skill Discovery (ASD) framework, enabling Large Language Models (LLMs) to autonomously propose novel robotic tasks, guide reinforcement learning, and evaluate skill acquisition. This system acquired 17 atomic skills from 22 learnable tasks and completed complex multi-step instructions through decomposition, leveraging a Vision-Language Model to filter false positive skill detections.

Key insights

  • The coordinated 'fast' (LLM-generated) and 'slow' (VLM-verified) success determination mechanism is crucial for filtering false positives and ensuring the reliability of learned skills, preventing contamination of the skill library.
  • LLMs can effectively drive autonomous skill discovery by proposing novel tasks and specifying learning objectives, significantly reducing reliance on human supervision for task definition and reward engineering.
  • Retrieval Augmented Generation (RAG) by prompting the LLM with previously verified success and reward functions enhances efficiency in generating effective learning objectives, while sequential policy initialization aids in learning complex, multi-step tasks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.