This article investigates the predefined time trajectory tracking control problem of upper limb rehabilitation robots in the presence of the model uncertainty and external disturbances. An adaptive neural network-based practical predefined-time fast nonsingular terminal sliding-mode control (PPT-FNTSMC) strategy is proposed, in which a novel sliding mode surface is constructed to achieve predefined-time convergence and avoid the singularity problem as well. Additionally, a neural network is employed to approximate the lumped disturbances, and the estimated values are utilized in the controller for compensation, thereby reducing the required switching gain and mitigating chattering phenomenon. A rigorous theoretical analysis is provided to illustrate that the tracking errors can converge to a vicinity of the origin in a predefined time. Finally, simulations on a two-joint single-arm robot and a 5-DOF upper-limb exoskeleton are conducted and compared with an existing control method to demonstrate the effectiveness and superiority of the proposed control strategy.