I will present my research towards predictive simulations of human movement for assistive devices and rehabilitation treatment. First, I will talk about a neuromechanical control model based on simple reflexes. The model can generate diverse locomotion behaviors, react to perturbations similarly to humans, and explain why walking performance declines with age. However, as the model was focused on low-level motor control primarily for steady locomotion behaviors, extending and verifying the model for more complex movements and reactions is necessary for producing reliable predictions for novel scenarios. In the later part, I will present recent projects on conducting a human experiment with gait assistive exoskeletons and using deep reinforcement learning to developing complex control models. In the experimental study, we found using human-in-the-loop optimization that ankle exoskeletons can substantially increase walking speed and energy efficiency in healthy younger and older adults. Regarding deep reinforcement learning, we organized the Learn to Move competition, where participants developed controllers for a human musculoskeletal simulation model. The competition has been organized at the NeurIPS conference from 2017 to 2019 and has attracted over 1300 teams from around the world. At last, I will discuss my plan of incorporating rigorous experimental validations and advanced computational techniques toward neuromechanical models that could change the way we design rehabilitation treatment and study human movement.