SuperTac - tactile data super-resolution via dimensionality reduction

TL;DR

SuperTac improves tactile sensor resolution in robotics by combining dimensionality reduction and upsampling, enhancing texture classification and enabling better tactile perception.

Contribution

SuperTac introduces a novel tactile super-resolution framework combining dimensionality reduction and Residual-In-Residual Blocks for real-time high-resolution tactile data.

Findings

SuperTac achieves a fourfold spatial resolution improvement in tactile images.

Texture classification accuracy improves by 17% with super-resolved tactile data.

The framework maintains computational efficiency for real-time applications.

Abstract

The advancement of tactile sensing in robotics and prosthetics is constrained by the trade-off between spatial and temporal resolution in artificial tactile sensors. To address this limitation, we propose SuperTac, a novel tactile super-resolution framework that enhances tactile perception beyond the sensor’s inherent resolution. Unlike existing approaches, SuperTac combines dimensionality reduction and advanced upsampling to deliver high-resolution tactile information without compromising the performance. Drawing inspiration from the spatiotemporal processing of mechanoreceptors in human tactile systems, SuperTac bridges the gap between sensor limitations and practical applications. In this study, an in-house-built active robotic finger system equipped with a 4 × 4 tactile sensor array was used to palpate textured surfaces. The system, comprising a tactile sensor array mounted on a spring-loaded robotic finger connected to a 3D printer nozzle for precise spatial control, generated spatiotemporal tactile maps. These maps were processed by SuperTac, which integrates a Variational Autoencoder for dimensionality reduction and Residual-In-Residual Blocks (RIRB) for high-quality upsampling. The framework produces super-resolved tactile images (16 × 16), achieving a fourfold improvement in spatial resolution while maintaining computational efficiency for real-time use. Experimental results demonstrate that texture classification accuracy improves by 17% when using super-resolved tactile data compared to raw sensor data. This significant enhancement in classification accuracy highlights the potential of SuperTac for applications in robotic manipulation, object recognition, and haptic exploration. By enabling robots to perceive and interpret high-resolution tactile data, SuperTac marks a step toward bridging the gap between human and robotic tactile capabilities, advancing robotic perception in real-world scenarios.

Figures