The Lower Limb Rehabilitation Robot (LLRR) can provide timely and precise rehabilitation training for patients, but each patient has different body parameters and is in a different rehabilitation period, which means that different training methods have to be used. Therefore, how to develop individualized training methods based on patients’ physical and gait parameters has become a major research topic at this stage. In this paper, we choose the Long Short-Term Memory network (LSTM) as a model for gait generation and collect body data and gait data of 23 healthy individuals, with a total of 726 sets of data, to train the LSTM network, and establish the gait generation from input (body parameters. gait parameters) to the output (gait trajectory), and the experimental results were analyzed by mean absolute deviation (MAD), which showed that the trajectories generated by the LSTM network were closer to the actual joint trajectories than those generated by previous methods.