Reproduce “HiRE: Hindsight Reward Editing for Policy Finetuning” 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
Haoyi Niu, Zhengtao Han, Yufeng Ji, Zhongyu Li, Koushil Sreenath
HiRE edits visual rewards for robot reinforcement learning using examples of successful and failed outcomes in a frozen visual representation, without training an auxiliary reward model. It rewards states resembling success and penalizes states resembling failure. On Tool_Hang, it reaches about 95% success versus none with sparse-reward finetuning, within the reported DICE-RL setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.