Leveraging AI to Personalize Physical Assistance Policies in Upper-Limb Robot-Aided Rehabilitation

Robot-aided rehabilitation has emerged as an effective approach to support recovery of upper limb motor functions in patients with musculoskeletal and neuromuscular disorders. Integrating Artificial Intelligence (AI) into robotic platforms allows therapies to be personalized and adaptively tailored to individual patient needs, enhancing both patient engagement and the efficiency of clinical interventions. The work focuses on analyzing different levels of robotic autonomy in rehabilitation by leveraging AI as an enabling tool. In this context, AI is not treated as a substitute for clinical expertise nor as the primary decision-maker, but as a mechanism to implement and support distinct forms of physiotherapist-guided adaptive behaviors. Through AI-driven policies, the robot can modulate low-level control parameters, such as assistance level, within a clinically defined therapeutic framework, while the physiotherapist remains responsible for defining, supervising, and evaluating the overall rehabilitation strategy. The focus is therefore on defining innovative methodologies to analyze and produce varying degrees of autonomy grounded in physiotherapists' expert knowledge, with the aim of identifying which level of autonomy most effectively supports personalized rehabilitation in a structured and safe manner. To support this objective, the work begins with a systematic literature review to identify AI-enabled functional modules integrated into upper limb rehabilitation systems, providing a foundation for the subsequent development of intelligent control strategies. It then examines physiotherapists’ decision-making processes, analyzing how key parameters such as assistance levels are adapted in response to patient performance and engagement. This analysis establishes a clinically grounded reference model of expert behavior, which informs the design of AI-driven approaches aimed at replicating and progressively approximating expert adaptation strategies across different levels of autonomy. Importantly, the study explicitly compares these levels rather than assuming full autonomy as the optimal target, evaluating how different degrees of AI support influence the ability to replicate clinically meaningful adaptation within the constrained scope of assistance modulation. Building upon this foundation, a supervised framework is proposed using behavioral cloning, enabling the system to reproduce physiotherapist-inspired adaptation strategies in real time. This intermediate level of autonomy reflects a first step toward embedding expert knowledge into the robot, providing decision support that remains aligned with clinical reasoning while assisting the physiotherapist in adjusting therapy parameters. Lastly, higher autonomy is investigated through online reinforcement learning, where the robot dynamically adapts control parameters based on multimodal patient data. This paradigm enables autonomous optimization of specific components of motor performance, such as kinematic outcomes, within clinically defined constraints. However, it also highlights a key limitation: designing reward functions that capture the multidimensional nature of rehabilitation, including motor performance, engagement, and assistance balance, is inherently complex, making reinforcement learning particularly suitable for targeted optimization. This progression from physiotherapist-guided control, to supervised behavioral cloning, to higher levels of autonomous reinforcement learning, illustrates a continuum of increasing autonomy in low-level control policies, emphasizing patient-centric rehabilitation while preserving safety, interpretability, and continuous clinical oversight. Taken together, the contributions of this thesis demonstrate that expert physiotherapist knowledge can be embedded into AI-based robotic systems through structured and complementary levels of autonomy, enabling scalable and clinically grounded adaptive rehabilitation strategies.