Generalizing Grasping via Shape Matching and Learning

Abstract

This thesis introduces a hybrid framework for robust robotic grasp synthesis from partial point clouds. We first formulate grasping as rigid shape matching, using the gripper's Signed Distance Field (SDF) for collision avoidance and a parallelized Stochastic Gradient Descent ICP (SGD-ICP) for efficient optimization. We extend this with Stein Variational Gradient Descent (SVGD) to generate diverse grasp distributions. Our core innovation integrates Energy-Based Models (EBMs) with analytical ICP gradients within the SVGD pipeline, fusing learned priors with geometric optimization. Experiments show targeted dataset curation shapes the energy landscape effectively, and our synergistic approach improves grasp quality, generalization, and robustness over purely analytic or data-driven methods.