Geometry-Aware 6-DoF Grasp Learning From Stacked Point Clouds in Unstructured Scenes

Abstract

To address the challenge of robust 6-DoF grasp prediction in cluttered, unstructured environments, this paper proposes a geometry-aware two-stage grasping framework. First, a plug-and-play Structural Edge-aware Keypoint Extractor is introduced to selectively retain salient edge geometry while mitigating point loss during downsampling. Second, a graspable feature extractor of GraspLAKAN-Net is developed to extract graspable features from stacked point clouds by integrating local aggregation with nonlinear mapping via an improved KAN-based architecture. To further enhance cross-object generalization, a coarse-to-fine Geometry-Aware Grasp Pose Estimation strategy is designed, enabling accurate and explainable 6-DoF pose prediction across both surface-dominant and edge-dominant objects. The proposed method is evaluated on a newly constructed MSDRG dataset comprising both simulated and real-world multi-object scenes. Comprehensive experimental results indicate the effectiveness of the proposed framework, which significantly outperforms state-of-the-art baselines in both recognition precision and operational efficiency.