Reproduce “Learning Multiple Timescales for Goal-Conditioned Reinforcement Learning” 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

Pedro Robles Dutenhefner, Dikshant Shehmar, Wagner Meira Jr., Marlos C. Machado

GITA addresses offline goal-conditioned reinforcement learning by learning one value function queried at multiple temporal scales and training a high-level policy with subgoal advantages aggregated across those scales. Across 23 evaluation tasks, it raises average success by 25 percentage points over HIQL and 7 over OTA. It does not win every task, and the comparison retunes discount per task.

Summary by alphaXiv.

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