Reproduce “Learning Action-based Representations Using Invariance” 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

Max Rudolph, Caleb Chuck, Kevin Black, Misha Lvovsky, Scott Niekum, Amy Zhang

Action-bisimulation extends visual reinforcement learning representations by recursively combining single-step action prediction with predicted future state differences under the same actions. It is pretrained on reward-free random transitions before policies are trained with task rewards. In Habitat navigation, it reports 0.7754 ± 0.005049 on Success (SPL) versus 0.7374 ± 0.0065 for ACRO, while final scores favor single-step on some other tasks.

Summary by alphaXiv.

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