Reproduce “GLUE: Global-Local Unified Encoding for Imitation Learning via Key-Patch Tracking” 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
Ye Chen, Zichen Zhou, Jianyu Dou, Te Cui, Yi Yang, Yufeng Yue
GLUE addresses the covariate shift problem in robotic imitation learning by introducing a global-local unified encoding framework that efficiently defines and tracks task-relevant "key-patches" and fuses them with global scene context. The framework significantly improves policy robustness and generalization, achieving an average success rate of 85.0% in real-world tasks and outperforming baselines by 58.3% in out-of-distribution scenarios.
Key insights
- Combining global scene context with moderately granular, dynamically tracked "key-patches" effectively mitigates covariate shift by focusing attention on task-relevant regions.
- Using multi-head cross-attention to distill local features from global context creates a unified, less heterogeneous feature space that enhances robustness and generalization.
- Text-guided initialization of keypoints offers an efficient and adaptive method for defining task-relevant visual cues, reducing reliance on extensive manual annotation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.