TacLoc: Global Tactile Localization on Objects from a Registration Perspective

Zirui Zhang, Fumin Zhang, Huan Yin

Keywords: Perception for Grasping and Manipulation, Force and Tactile Sensing, Localization

TL;DR: TacLoc performs fast, accurate, and robust 6DoF object localization based on only touch sensing.

Abstract: Existing tactile-based methods generally rely on rendering data or pre-trained models, which limits generalizability and efficiency. In this study, we propose TacLoc, a novel tactile localization framework that addresses these challenges by formulating the problem as a one-shot point cloud registration task. TacLoc introduces a graph-theoretic partial-to-full registration method, leveraging tactile sensor-generated dense point clouds and surface normals for efficient and accurate pose estimation. The approach eliminates the need for extensive training data. It achieves robust performance through normal-guided graph pruning and a hypothesis-and-verification pipeline. Evaluations on the YCB-based dataset demonstrate the superiority of TacLoc in terms of accuracy, efficiency, and generalization. We also perform the real-world tests using Gelsight Mini sensor.

Submission Number: 10