We propose a novel online model-based motion planning algorithm for a family of rehabilitation exoskeletons to improve transparency in user-guided operation. In this study, we assume that the short-term human movement intention can be embedded in the time-delay dimensions of motion signals. The model-based estimation is employed to obtain the interaction load between the dynamical subsystems respectively controlled by the human and exoskeleton. The objective of the proposed motion planning algorithm is to reduce the interaction load, which leads to the establishment of a least-square optimization problem. A Support Vector Regression (SVR) model, driven by the time-delayed motion data, is implemented to solve the optimization problem by generating the acceleration of tracking reference. The motion planning algorithm based on SVR can be combined with a variety of trajectory tracking controllers. To ensure the efficiency of the algorithm for online applications, we also design the SVR model so that its properties can be calculated recursively based on latest data sets. The performance and characteristics of the motion planning algorithm are then observed and discussed through the control simulations of a wearable wrist exoskeleton designed for pathological tremor alleviation. The results show that while the planned tracking reference can approximate the synthetic human movement intention, the motion planning accuracy can be limited by system disturbances, and the delay of signals caused by digital filters.