LESS Improves Tactile Imaging with Local Scene Representations.

Tactile imaging aims to reconstruct the internal structure of soft objects through touch, with applications ranging from medical diagnosis to robotic manipulation. While recent self-supervised learning methods have shown promise, they often rely on global, unstructured representations and robot-controlled sensing, which limits their generalization and practical utility. This research introduces Local Encoder for Spatial Sensing (LESS), an object-centric tactile representation specifically designed to leverage the local nature of touch. LESS models the tactile scene as a grid of recurrent encoders, each with a local receptive field. The states from these encoders are then fused to reconstruct 2D or 3D images of the object's internal structure. The compositional design of LESS allows for robust generalization; models trained on simple single-inclusion phantoms can accurately image objects with multiple inclusions and varying sizes. Furthermore, the local structure facilitates spatial uncertainty estimation. The method also extends to hand-held tactile imaging using external pose tracking and human-like palpation data, enabling full 3D reconstruction.