Reproduce “Hierarchical Reinforcement Learning By Discovering Intrinsic Options” 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

Jesse Zhang, Haonan Yu, Wei Xu

HIDIO, a hierarchical reinforcement learning method, learns task-agnostic options self-supervisedly and jointly trains a high-level scheduler to solve complex sparse-reward tasks. The framework demonstrates superior sample efficiency and higher success rates across robotic manipulation and navigation environments compared to state-of-the-art baselines.

Key insights

  • Jointly training task-agnostic option discovery with the scheduler is more stable and effective than sequential pre-training, ensuring learned skills are directly relevant to the task.
  • Using sub-trajectory features (e.g., 'StateAction', 'StateDiff') for the discriminator is more effective for learning distinct and identifiable options than relying on instantaneous states.
  • Temporal abstraction through discovered options significantly enhances exploration and learning in sparse-reward environments, enabling solutions for tasks beyond the reach of non-hierarchical methods.

Summary by alphaXiv.

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