Reproduce “Debiased Offline Representation Learning for Fast Online Adaptation in Non-stationary Dynamics” 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

Xinyu Zhang, Wenjie Qiu, Yi-Chen Li, Lei Yuan, Chengxing Jia, Zongzhang Zhang, Yang Yu

DORA learns an offline reinforcement-learning policy that adapts to dynamics changing within an episode, using recent state-action history without a pre-collected calibration rollout. Its recurrent encoder groups representations by task and penalizes differences between successive history prefixes to reduce reliance on data-collection actions. Across six simulated MuJoCo settings, it leads methods in five; on Hopper-gravity, Offline ESCP scores higher.

Summary by alphaXiv.

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