Reproduce “Behavior Retrieval: Few-Shot Imitation Learning by Querying Unlabeled Datasets” 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
Maximilian Du, Suraj Nair, Dorsa Sadigh, Chelsea Finn
Researchers at Stanford University introduced Behavior Retrieval, a method for few-shot imitation learning that selectively leverages large, unlabeled offline datasets using a learned state-action similarity metric. This approach consistently improved success rates by over 20% on real robot tasks and more than 30% in simulations compared to standard pre-training and fine-tuning baselines, especially under distribution shifts.
Key insights
- A small amount of task-specific expert data can serve as a query to identify and retrieve relevant behaviors from vast, unlabeled offline datasets.
- Functional similarity between state-action pairs, rather than just state similarity, is crucial for effectively identifying useful prior data segments.
- Behavior Retrieval can extract useful sub-trajectories from otherwise irrelevant demonstrations, allowing it to leverage shared components across diverse tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.