Grasping by parallel shape matching

TL;DR

This work reframes grasp generation as rigid shape matching between a gripper and an object, solved efficiently with parallel Annealed Stein ICP (AS-ICP) and SGD-ICP on GPU. By incorporating the gripper's tool center point relative to the object's center of mass and leveraging a gripper SDF for collision checking, the method achieves robust, real-time capable grasps without optimizing finger joints. It demonstrates 87.3% average success over 11 objects with the Kinova KG3 and a per-grasp time of 0.926 s in real experiments, outperforming several data-driven baselines in terms of robustness to noise and partial observations. The approach is gripper-agnostic and training-free, offering practical utility for real-world robotic manipulation while suggesting future gains from integrating preshape selection learned from data.

Abstract

Grasping is essential in robotic manipulation, yet challenging due to object and gripper diversity and real-world complexities. Traditional analytic approaches often have long optimization times, while data-driven methods struggle with unseen objects. This paper formulates the problem as a rigid shape matching between gripper and object, which optimizes with Annealed Stein Iterative Closest Point (AS-ICP) and leverages GPU-based parallelization. By incorporating the gripper's tool center point and the object's center of mass into the cost function and using a signed distance field of the gripper for collision checking, our method achieves robust grasps with low computational time. Experiments with the Kinova KG3 gripper show an 87.3% success rate and 0.926 s computation time across various objects and settings, highlighting its potential for real-world applications.

Figures (8)

• Figure 1: An illustration of the optimization process. Green, blue and black point clouds are the initial poses of three preshapes of the KG3 gripper. Blue plots show the optimization process to find the grasp pose of a single preshape.
• Figure 2: (a) Initializations sampled from a mixture of Gaussian to provide some prior knowledge of the object. (b) Initializations sampled from the Fibonacci sequence and projected onto a quarter of the sphere, facing the direction of the robot arm.
• Figure 3: Preshapes used for simulation in this paper. On the left, we have two preshapes of Barrett Hand. On the right, we selected two preshapes from the ten used for Franka Hand. The full point cloud generates SDF, and the partial point cloud is used to optimize.
• Figure 4: Plots of success rate and computation time with an increasing number of initializations for different sampling methods: (a) Gaussian and (b) Fibonacci. Sampling from a mixture of Gaussian leads to a higher success rate and faster computation time with fewer initializations as it provides some prior knowledge of objects.
• Figure 5: Best grasp poses with Franka Hand for 50 trials. On the left is the pose generated by our algorithm and on the right is the pose generated by AnyGrasp.