Reproduce “TARGO: Benchmarking Target-driven Object Grasping under Occlusions” 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
Yan Xia, Ran Ding, Ziyuan Qin, Guanqi Zhan, Kaichen Zhou, Long Yang, Hao Dong, Daniel Cremers
Researchers from Technical University of Munich, Peking University, and University of Oxford developed TARGO, the first dedicated benchmark for target-driven object grasping under occlusions, and introduced TARGO-Net. This transformer-based model significantly enhances robustness, showing only an approximate 7% decrease in grasp success rate from low to high occlusion levels in synthetic tests, outperforming existing baselines.
Key insights
- Existing state-of-the-art grasping models experience a significant degradation in performance (up to 30% in some cases) as occlusion levels increase, highlighting the severe challenge.
- The integration of a dedicated 3D shape completion module for the target object is crucial for maintaining grasp success under high occlusion, proving more effective than general scene completion.
- Augmenting training data with unoccluded single-target scenes substantially improves the robustness and generalizability of grasping models against occlusion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.