Reproduce “SkiLD: Unsupervised Skill Discovery Guided by Factor Interactions” 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

Zizhao Wang, Jiaheng Hu, Caleb Chuck, Stephen Chen, Roberto Martín-Martín, Amy Zhang, Scott Niekum, Peter Stone

SkiLD is an unsupervised skill discovery framework that learns skills by targeting specific local dependencies between environmental factors. This approach enables the acquisition of complex multi-object interaction skills, leading to superior performance on challenging embodied AI tasks compared to methods focusing solely on state diversity.

Key insights

  • Skills that create diverse interactions between environmental factors are typically more valuable for solving downstream tasks than skills focused on simple state diversity.
  • Representing skills as desired local dependency patterns between state factors provides a structured inductive bias that guides exploration towards semantically meaningful behaviors.
  • Integrating a high-level policy for selecting target dependency graphs with a low-level policy for inducing these dependencies enables the learning of complex, reusable skills.

Summary by alphaXiv.

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