Carl Bettosi

Upper-limb rehabilitation is vital to patients recovering from critical injuries such as fractures, ensuring the best possible health outcomes while delivering huge social and economic benefits. Yet, these recovery programmes demand high patient engagement which is often challenged by low patient motivation and stress on healthcare resource. Socially Assistive Robots (SARs) have emerged as promising tools for physical recovery, surpassing virtual agents in user engagement, with the ability to adopt instructor-like roles. In real-world scenarios, human experts learn over their long-term relationship with the patient their physical abilities and social preferences, allowing them to optimise recovery sessions to boost engagement and drive higher therapeutic outcomes. For SARs, real-time user state understanding paired with intelligent decision-making approaches such as reinforcement learning may bring us closer to such human-like interactions over long-term interaction. This PhD project employs participatory design techniques to develop a precise model of physiotherapy behaviours. From this, a behavioural learning system for a SAR is developed using reinforcement learning and real-time user state perception. The aim is to develop and evaluate an adaptive SAR capable of facilitating recovery programs for individual patients over the long term.