Robotic devices represent an undeniable alternative method to traditional therapy. However, the clinical use of robot-assisted rehabilitation for patients with hemiplegia after stroke has been restricted for decades by i) limited functional characterisation of robotic devices; ii) non-efficient and widely accepted robot-measured kinematic parameters to assess performances of robot-assisted therapies; and iii) limited identification of rehabilitative strategies based on recovery mechanisms. This led to a misuse of robotic systems and poorly effective clinical trials. For example, a recent work in a large cohort of patients showed that the use of an upper limb exoskeleton, compared to conventional therapy, promoted a larger arm mobility recovery but also a lower improvement in arm strength. These ambiguous results strengthen the idea that, while robotic devices provide a mean for intensive practice, hand-on would always be crucial for turning benefits into functional gains. I argue that the limits of robot-assisted rehabilitation stem from a lack of fundamental understanding of its interactions with the human sensori-motor system and not by an intrinsic limited value. To this aim I developed a multimodal framework to investigate and characterise the full spectrum of interactions between a robot and the human sensori-motor system, from muscles to brain signals. This framework could boost robot-assisted rehabilitation and, thus, its efficacy for daily clinical practice. This framework includes i) a new developed exoskeleton â i.e., Arm Light Exoskeleton (ALEx); ii) a computational model for the selection of the optimal robot-measured kinematic parameters; iii) a novel statistical tool for the analysis of the brain activity; iv) an advanced signal processing technique to analyse the brain activity in relation to muscle activity during arm movements. Overall, results show that i) a deep characterisation of robotic systems allows understanding the robot-human sensori-motor system interactions and selecting optimal assistive modalities for the rehabilitative training; ii) robot-measured kinematic parameters can be informative of central and/or peripheral mechanisms of motor recovery; iii) the analysis of the cortico-muscular organisation supported by advanced statistical methods could be useful to monitor the neuro-biomechanical state of post-stroke patients. These results might have important implications in the understanding of the neural correlates of motor recovery to drive successful robot-assisted rehabilitation. A clinical trial that applies the conceptual framework, scientific rationale, and technical details explained in this Thesis has recently started at the University Hospital of Geneva, Switzerland, and at the University Hospital of Pisa, Italy. The study aims, among other goals, at defining the neuro-biomechanical state of the patient and its evolution during the therapy by using the multimodal framework here presented.