Reproduce “Dream to Generalize: Zero-Shot Model-Based Reinforcement Learning for Unseen Visual Distractions” 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
Jeongsoo Ha, Kyungsoo Kim, Yusung Kim
This paper presents Dream to Generalize (Dr. G), a model-based reinforcement learning framework that achieves zero-shot generalization to unseen visual distractions by replacing traditional image reconstruction with self-supervised objectives. Dr. G demonstrates robust performance, improving by 117% over existing algorithms on challenging benchmark tasks like DeepMind Control and Robosuite under various visual noise conditions.
Key insights
- World models can learn robust, task-relevant features by replacing pixel-level reconstruction with self-supervised objectives, enabling zero-shot generalization to unseen visual distractions.
- A dual contrastive learning approach, combining reality-reality and dream-reality alignment, effectively trains invariant latent representations under varied visual augmentations.
- Integrating Recurrent State Inverse Dynamics (RSID) with contrastive learning enhances the temporal consistency of learned representations, further improving world model understanding for policy planning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.