Intelligent assistance systems play an increasingly important role in today's physiotherapy treatment. Computer-aided systems have proven themselves as a supporting element in the medical staff and are being used more and more frequently. In physiotherapy, the correct identification of rehabilitation exercises is particularly important. This is often based on the heuristic evaluation of image data recorded by a stereoscopic camera. For unambiguous exercises and gestures, the recognition through previously defined rules (e.g. at which angle the right leg has to be raised when standing on one leg) is satisfactory, since the associated image data can be easily evaluated and compared with the rules. However, heuristic methods do not provide satisfactory solutions for gesture recognition for more complex movement sequences or exercises in which individual parts of the body are often covered (e.g. during exercises in the squat or lying position). AI-based systems should be considered here as a possible problem solution. This issue is addressed by this project. Assisting exercise programs are developed in cooperation with patients and therapists. The movements are analyzed by means of (depth) sensors and multimodal instructions (e.g. via speech and gestures) are given by agents / robots. If movements are inadequately recognized, human experts instruct and continuously improve the CNN to be developed using supervised learning.