Reproduce “Enhancing Visual Domain Robustness in Behaviour Cloning via Saliency-Guided Augmentation” 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
Zheyu Zhuang, Ruiyu Wang, Nils Ingelhag, Ville Kyrki, Danica Kragic
A saliency-guided augmentation method, RoboSaGA, improves the visual domain robustness of vision-based behavior cloning policies by dynamically applying per-pixel augmentation based on task-critical saliency. This approach, which integrates without architectural changes, reduced performance degradation due to visual shifts by 41.6% to 72% over existing methods in simulated and real-world robot manipulation tasks.
Key insights
- Policy-driven saliency maps enable dynamic, pixel-level augmentation that maintains task-critical visual information while introducing diverse visual elements, crucial for robotic manipulation.
- Direct training with saliency-guided augmented samples offers comparable or better robustness against visual domain shifts than more complex representation learning methods like SODA, without additional architectural changes.
- Superimposition-based augmentation methods, particularly RoboSaGA, generalize visual robustness improvements across various behaviour cloning policy types, including BC-MLP, BC-RNN, and Diffusion Policy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.