This paper aims to develop a robust robotic grasping framework for cluttered scenes under single-view constraints by addressing the challenges of feature efficiency, physical feasibility and perceptual incompleteness. The primary objective is to build a system that integrates multiscale feature fusion, collision-aware reasoning and scene-completion-guided optimization to ensure stable and physically consistent grasp execution.
The authors propose a framework integrating three synergetic components. A collision-aware grasp prediction module incorporates collision reasoning to enhance physical feasibility, while a multiscale seed sampling feature fusion module uses density-adaptive cylindrical sampling to balance geometric precision and computational efficiency, coupled with a local attention mechanism to explicitly suppress feature over-smoothing. Finally, a scene-completion-guided grasp refinement module leverages geometric priors to hallucinate occluded surfaces, guiding contact-based optimization to ensure grasp stability and physical consistency.
The proposed method outperforms existing approaches by a clear margin on the GraspNet-1Billion benchmark, achieving an improvement of 5.53 AP and is further validated in feasibility verification experiments in real-world cluttered scenes.
This research introduces an integrated modular approach that decouples collision reasoning, feature extraction and refinement. By integrating auxiliary collision learning and pretrained geometric priors, our system significantly enhances grasping robustness and physical consistency in cluttered scenes.
