Object Handle Segmentation in 3D Point Cloud for Robot Grasping Using Scale Invariant Heat Kernel Signature With Optimized XGBoost Classifier

Abstract

Segmenting graspable regions is crucial for robotic manipulation tasks like pick-and-place and pouring. This study proposes a robust method for detecting handle-like regions in common objects, focusing on slender handles distinct from the main body. This characteristic is prevalent in many daily-use objects that are often manipulated. Our method employs the scale-invariant heat kernel signature (SI-HKS) descriptor to capture local and global shape features of 3D objects. By utilizing SI-HKS properties, we extract meaningful geometric information. Points are classified into segments using the XGBoost classifier, known for its efficiency and accuracy, while hyperparameters are optimized through random search. A post-processing step refines handle detection by filtering out non-graspable regions based on geometric skeleton curvature. The proposed approach is evaluated on a custom dataset in two configurations: five categories of handle-equipped objects and extended version with eleven categories. In the 5-class setup, the method achieves a mean intersection-over-union (mIoU) of 97.6%, outperforming leading deep learning models like PointNet, PointNet++, and DGCNN with statistically significant improvements confirmed by t-tests. In the extended 11-class setup, the method maintains a strong performance with a mean IoU of 97.5%. The use of intrinsic geometric features enhances rotation invariance, ensuring consistent segmentation across different orientations.