Active Tactile Exploration With Emphasis on 3D Texture Modeling

Active tactile exploration effectively reconstructs 3D shapes, particularly when visual input is limited. However, current Gaussian process (GP)-based methods primarily capture global shapes, often neglecting local texture details. To overcome the limitation, a texture-guided active tactile exploration pipeline is proposed for accurate texture reconstruction. Specifically, the GelStereo Palm 2.0 sensor is leveraged to collect high-precision tactile point clouds. GP is used to model object shapes, complemented by a GP classifier to learn the distribution of texture regions based on a local texture intensity indicator. By integrating outputs from the regression model and the classifier, the pipeline devises effective exploration strategies. Experiments on different wave surfaces and irregular objects demonstrate that our method achieves a 15% accuracy and 42% speed increase in reconstructing texture region over baseline methods. This technique shows great potential for applications in underwater shape perception and the reconstruction of transparent objects. See supplementary video and results at https://sites.google.com/view/activetactileexploration