Aiming at the problem of high precision trajectory tracking control of lower limb exoskeleton rehabilitation robot used by stroke patients during passive training, a PD control method based on deep deterministic policy gradient was proposed. The joint angles of normal human walking were collected by Vicon 3D motion capture system as the desired joint angles trajectories and the dynamic model of lower limb exoskeleton robot was established. According to each error input and the reward value obtained by interaction with the dynamic model, the own network parameters were dynamically updated to adaptively output the best PD parameters. The simulation results show that compared with the traditional PD control, using this method, the tracking errors of hip, knee and ankle are reduced by 94% on average, and the errors tend to converge and tend to zero, indicating that this method can effectively track the desired joint angles trajectories and has good tracking control accuracy to ensure the rehabilitation effect of patients.