Generative Artificial Intelligence has recently gained significant attention across various domains, including healthcare, offering benefits beyond data augmentation, enabling clinical scenario simulation to enhance precision and efficiency of medical interventions. This is particularly useful in medical fields such as rehabilitation, where therapists should prescribe tailored exercises based on individual patient characteristics. In this study, a generative conditional diffusion model is employed to simulate movements of upper limb during rehabilitation sessions, generating positional trajectories for the human joints, i.e., shoulder, elbow and wrist, conditioned by multiple factors, including the type of movement being performed, the subject’s posture (sitting or standing), and the body side affected by the impairment. The dataset used to train and test the model comprises positional signals extracted from various movements associated with physical rehabilitation, including multiple repetitions of rehabilitation exercises performed by both patients and healthy individuals. The approach has been analyzed qualitatively, employing a dimensionality reduction technique to evaluate the simulated movements and compare them with the original data. The results indicate that the simulated movements closely resemble the original ones, suggesting that diffusion models demonstrate a strong ability to learn and generate data of this nature.