Ensuring Safety in Upper-Limb Prostheses: Tactile Sensor and Machine Learning for Risk Prediction

Upper limbs play an important role in everyday life, enabling a variety of activities beyond mere object manipulation or grasping, as allowing communication, productivity, creativity, and physical health through writing, drawing, sports and recreation. It is evident, then, that the loss of upper limb functions deeply impacts individuals' daily lives and quality of life. Despite extensive research to replicate upper limb capabilities with prosthetic devices, users often face challenges in adapting to their prosthesis, leading to high rejection rates. Addressing this challenge requires prosthetics that not only restore functionality but also offer natural control and autonomy for daily activities. In particular, prostheses often demonstrate inadequate sensory feedback and limited proprioceptive information. This deficiency in sensory perception leads to poor slip control, difficulties in adjusting grip force, complications with object manipulation, and decreased dexterity. These factors, together with a heavy reliance on visual cues, prevent an easy integration of prosthetic devices into daily routines and challenge users' acceptance. In line with these goals, this master’s thesis seeks to enhance grasping safety by introducing a machine-learning algorithm capable of interpreting sensory information from tactile sensors embedded within the prosthetic hand. The work involves a comprehensive overview of existing non-invasive feedback mechanisms and tactile sensor technologies, alongside an in-depth exploration of experimental methodologies for detecting and predicting slippage. Afterwards, to build and train the machine learning algorithm, a comprehensive dataset was collected, incorporating various actions categorized into three main groups: grasp, risky and non-risky. These actions involved interacting with objects of different shapes and textures. The data was gathered using a commercially available prosthesis, specifically the Michelangelo hand, equipped with six tactile sensors embedded on its fingers. Finally, an experimental validation was conducted, involving external participants interacting with the prosthetic hand. This validation served to evaluate the accuracy of the algorithm's predictions and gather feedback for potential enhancements. Through this iterative process of data collection, algorithm development, and experimental validation, this thesis aims to predict slippage to ensure grasping safety.