Reproduce “Learning to See through Action: Predictive Curiosity for Robots” 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

Stelios Zarifis, Panagiotis Filntisis, George Retsinas, Petros Maragos

The authors introduce DySCOVER, a robot vision method using reconstruction and action-conditioned prediction errors to guide exploration. An encoder, decoder, and recurrent world model provide the two curiosity signals that shape experience. In MuJoCo simulation, adjusted DySCOVER reduced evaluation RGB reconstruction MSE by 14.5% versus visual-only curiosity after 10 million steps; this measures reconstruction, not task or real-world performance.

Summary by alphaXiv.

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