Physically-Grounded 3D Point Cloud Filtering and Clustering Based on Tactile Sensor Specifications

Abstract

When robotic systems are operated in unstructured environments, they often rely on combining depth cameras and tactile sensors to achieve reliable contact with objects of diverse geometries, typically requiring precise alignment with planar surfaces. However, with traditional 3D point cloud filtering and planar clustering methods, extensive manual tuning of abstract mathematical parameters may be necessary, which are often unintuitive and disconnected from the physical properties of sensing hardware. To address this, this paper introduces a novel point cloud processing framework that directly integrates the physical characteristics of capacitive tactile sensors—specifically their spatial resolution and sensor area—into the filtering and planar segmentation stages. By aligning algorithmic parameters with sensor capabilities, the proposed method achieves physically interpretable parameterization, facilitating effective adaptation to real-world scenarios. In the filtering stage, the original object point cloud is processed using thresholds based on the sensor’s resolution, with geometric structure retention levels, such as edges and protrusions determined by user-specified roughness tolerance. Compared to voxel-based filtering methods, this approach effectively eliminates surface artifacts while significant geometric features are preserved. The clustering stage employs a sliding-window mechanism analogous to convolutional kernels, with both window size and flatness thresholds defined by the sensor’s coverage area and the acceptable surface roughness specified by the user. Experiments using a UR5 robotic arm equipped with capacitive tactile sensors of different resolutions demonstrate that the proposed framework robustly achieves precise surface alignment and generalizes effectively across different sensor types by simply adjusting parameters according to sensor specifications. This approach significantly reduces parameter tuning complexity, offering a physically grounded, low-overhead solution for geometry-aware robotic perception.