Reproduce “Learning from Relative Action Credit for Escaping Local Minima” 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
Giovanni Minelli, Giulio Turrisi, Lorenzo Amatucci, Victor Barasuol, Claudio Semini
PPO+GRPO compares same-state actions using an action-conditioned critic, adding a local action-relative signal to PPO’s temporal signal and an exploration floor. In simulated peg insertion, it escaped a plateau after roughly 650 iterations and reached about 95% of maximum return, averaged over five seeds; this result does not establish transfer to other tasks or hardware.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.