Efficient Data Driven Algorithms for Tactile Data Processing in Embedded and Wearable Systems

Human interaction with the physical world relies heavily on the sense of touch, which provides information essential for manipulation and perception. Artificial tactile sensing systems have therefore emerged as a promising technology to enhance human–machine interaction, with growing interest in applications ranging from wearable devices to robotics and prosthetics. These systems typically consist of artificial tactile sensors that generate signals in response to physical contact, followed by embedded algorithms running on resource-constrained devices for signal acquisition and processing. However, achieving efficient tactile data processing on embedded and wearable platforms remains a challenge due to limitations in memory footprint, execution time, and power consumption. This thesis explores the development of data-driven algorithms and embedded imple- mentations for efficient tactile data processing across multiple use cases. The first part focuses on processing signals generated by our piezoelectric-based tactile sensors. First, the thesis outlines the analysis of the sensor response behavior, in which we integrated a setup comprising an array of piezoelectric sensors and interface electronics. Indenta- tion experiments were performed with variations in load and speed. Signal processing techniques were developed to detect contact events and reconstruct normal forces under controlled indentation. A lightweight machine learning model was then introduced for force estimation using a shallow artificial neural network trained on time-domain features extracted from the sensor signals. The trained neural network was deployed on a micro- controller to demonstrate real-time estimation of the peak of the normal force, confirming the efficiency of the proposed algorithm, with performance degrading at higher forces. The complete processing pipeline was evaluated on embedded hardware, achieving a total inference time of ≈ 6.23 ms and a total energy consumption of ≈ 528 𝜇J per inference. In parallel, modeling of the sensing system, including the transducer and the soft protective layer, was carried out to characterize its electromechanical behavior, aiming to generate tactile simulated data to train the algorithms with a wide range of forces and test the network experimentally, thereby reducing dependence on extensive experimental datasets and minimizing resource usage. The results indicated that the system exhibits a derivative-like response with resonance peaks above 350 Hz, likely originating from the soft cover layer. Accordingly, a phenomenological lumped-parameter model was developed to fit the sensor response function under indentation. The second part of this work addresses object recognition through the development of low-cost multisensory wearable gloves that combine tactile and proprioceptive sensors. Embedded machine learning techniques, including TinyML workflows for processing time-series data, were implemented to enable real-time recognition of twenty-eight daily- life objects. The results demonstrate that the system achieved high accuracy, with an inference time of 8.147 ms while consuming only ≈ 0.4875 mJ per inference. Furthermore, hardware-aware neural architecture search was applied to automate the design of the learning algorithms, thereby reducing resource usage. The results indicated an energy reduction of ≈ 27.92% without compromising performance. The glove was subsequently upgraded with a battery-powered embedded system, en- abling fully portable operation and real-time classification of 3D-printed objects spanning different categories of shape, size, and stiffness. The system was built using cost-effective components and validated through on-device experiments. The results demonstrated re- liable object-recognition performance, with an estimated battery lifetime of ≈ 26.5 hours under continuous operation.