The mixed reality upper limb rehabilitation training evaluation system based on FMA and deep learning is an intelligent rehabilitation system designed based on FMA and using Microsoft Kinect, HoloLens2.0, HTC positioning base station, tracker, DJI robot, deep learning, A * algorithm, etc. The purpose is to meet the social needs for intelligent, monitoring, evaluation, convenience and high efficiency of limb rehabilitation, at the same time, it solves the problems of the traditional VR rehabilitation system, such as the lack of monitoring module, the immature scene construction algorithm and the weak scalability of the system. It uses laser positioning, somatosensory recognition and other technologies, and has functions such as full-automatic evaluation, automatic obstacle avoidance, gesture interaction, voice interaction, and intelligent decision-making. When using the system, the user makes the specified action according to the specific monitoring task, and the system obtains the evaluation result according to the user's body information and intelligently decides the appropriate training content. During the whole evaluation training process, the user can rely on the robot to complete the auxiliary function, and the data generated by the user will be recorded in the rehabilitation cloud in real time, so that the doctor can backtrack and view.