Reproduce “Pretrained Features as Teachers, Not States: Compact Latent World Models for Visual Planning” 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
Yanbo Wang, Nan Lin, Wendelin Böhmer, Javier Alonso-Mora
Anchor-WM uses pretrained DINO features as a training-time teacher to train a compact action-conditioned world model. A head reconstructs DINO's spatial patch features, while inference discards teacher and head and plans with compact latent rollouts. Across four visual-only tasks, it averages 84.7% success, versus 79.9% for LeWM and 76.6% for DINO-WM, but does not lead on every task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.