Data-Driven Compliance Discrimination via Biomimetic Soft Optical Tactile Sensors: Implementation and Benchmarking With a Model-Based Approach

Abstract

Humans can easily manipulate soft, deformable objects, relying on the intrinsic deformability of their fingerpads and their capabilities to infer item compliance. Transferring these skills into robotic systems is still an open challenging task. Recently, the introduction of soft biomimetic tactile sensors such as the TacTip, which aims at mimicking the structure of human skin layers, has represented an attempt to bridge this gap. However, while the advancement in endowing artificial systems with bio-aware embodied intelligence, the computational aspects of compliance estimation are still largely unexplored, mostly relying on Data-Driven (DD) methods. In a previous work [1], a model-based approach to compliance estimation by combining TacTip with a computational model of human tactile perception has been proposed. However, both these categories of techniques suffer from limitations (generalizability issues for the model-inspired; the curse of data for DD), suggesting that a hybrid approach could represent a more suitable solution. The first step in exploring the feasibility of a hybrid framework is to develop robust neural network architectures and then benchmark them against the results achieved through the analytical method. In this manner, we can assess their complementary strengths and potential integration. This work specifically addresses this step by proposing a novel approach for compliance estimation via TacTip using deep learning techniques. The neural network architecture was validated, and a comparative analysis of its performance against our previously developed model-based approach was performed.