Reproduce “Few-Shot In-Context Imitation Learning via Implicit Graph Alignment” 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
Vitalis Vosylius, Edward Johns
A system for few-shot in-context imitation learning, Implicit Graph Alignment (IGA), enables robots to learn manipulation tasks from a limited number of demonstrations and generalize to novel objects and categories. IGA achieves an 80% average success rate on six real-world tasks and significantly outperforms baselines in simulated generalization scenarios, reaching 3.0 cm translation and 14.7 deg rotation error for unseen categories.
Key insights
- An energy-based model with heterogeneous graph representations can learn a conditional distribution of alignments, enabling few-shot generalization to unseen object categories.
- Increasing the diversity, rather than just the number, of demonstrations is essential for robust performance in inferring consistent task-relevant alignments.
- Data generation via correspondence-preserving shape augmentation with implicit models provides a scalable way to create diverse training data for category-level generalization without manual annotation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.