Reproduce “Autonomously Acquiring Robot Manipulation Skills with Language-Driven Quality-Diversity” 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

Émiland Garrabé, Mahdi Khoramshahi, Stéphane Doncieux

This research introduces a framework that enables autonomous robot manipulation skill acquisition by leveraging Large Language Models to infer Quality-Diversity algorithm components from natural language instructions. The method generates diverse motion primitive archives, achieving up to 85518 successful solutions for tasks like "Slide the mug," outperforming both expert-defined and naive LLM-inferred baselines in terms of success archive size and ground-truth behavior descriptor coverage.

Key insights

  • LLMs can autonomously design the functional components (success conditions, fitness, BDs) required for QD algorithms, significantly enhancing robot autonomy in skill acquisition.
  • A structured, split-space exploration for behavior descriptors (e.g., separating timestamps and measured quantities) and an iterative fitness selection process are crucial for LLMs to infer effective QD parameters.
  • Employing a multi-BD archive structure within MAP-Elites Success, where multiple behavioral dimensions are considered concurrently, proves effective for generating large, diverse skill archives, especially with LLM-inferred BDs.

Summary by alphaXiv.

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