Reproduce “Nightmare Dreamer: Dreaming About Unsafe States And Planning Ahead” 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
Oluwatosin Oseni, Shengjie Wang, Jun Zhu, Micah Corah
Nightmare Dreamer is a model-based Safe Reinforcement Learning algorithm that achieves near-zero safety violations while maximizing cumulative rewards, demonstrating up to a 20x improvement in sample efficiency compared to model-free baselines, especially when using image observations.
Key insights
- Proactively predicting future unsafe states through world model rollouts allows agents to prevent violations rather than react to them.
- Decoupling reward maximization and safety constraint satisfaction into a bi-actor architecture facilitates efficient multi-objective learning.
- Employing a discriminator to regularize the safe actor's policy provides stable training and improved generalization compared to standard behavior cloning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.