In batch-size-one part handling scenarios, robotic grasping of diverse and previously unseen objects remains a considerable challenge. This task requires robots not only to perceive and identify objects in unstructured environments autonomously but also to infer the grasp configurations that are optimally adapted to the geometric characteristics of both the target object and the specific end effector. To address this problem, the present work proposes a modular and extensible framework for optimal grasp point estimation based on 3D point cloud input and predefined gripper configurations. The system is designed for vision-guided robotic manipulation and aims to detect, evaluate, and compare feasible grasp candidates across multiple gripper types to determine the most suitable grasp configuration. The framework supports gripper-specific grasping strategies and introduces a unified Grasp Suitability Score GSS that enables consistent evaluation and direct cross-gripper comparability. In its current implementation, the method focuses on parallel grippers. It reduces the 3D grasp evaluation problem to 2D contour-based analysis through plane segmentation and PCA-based projection of object point clouds. Candidate test points are generated via grid sampling within the 2D contour of the projected point cloud and are subsequently filtered based on collision constraints and minimum contact area criteria. The resulting feasible grasp points are scored based on their alignment with the object's center of mass and the extent of contact surface coverage, thereby producing a ranked list of grasp configurations. The configuration with the highest GSS score is selected as the globally optimal solution. Experimental validation was conducted based on both synthetic and real-world point clouds, including models derived from open-source grasping datasets and their corresponding 3D-printed parts. The evaluation covered several key performance dimensions: functional correctness, localization accuracy of the estimated grasp point, repeatability under identical input conditions, robustness to varying point cloud densities, and computational efficiency. Results demonstrate that the proposed method reliably identifies the most suitable grasp configuration, with localization deviations of less than 3 mm, and maintains consistent performance across various scenarios. Furthermore, the framework exhibits extensibility toward suction-based and adaptive grippers. Future work will focus on improving plane segmentation accuracy, incorporating inner contour features, enhancing parameter adaptability, and enabling deployment in fully reconstructed real-world scenes. The complete implementation, including models, point clouds, evaluation results, and all associated executable scripts, is publicly available in an open-source repository to support reproducibility and facilitate further research.
Hanyu Liu (Thu,) studied this question.