A few reports suggest differences in injury outcomes between cadaver tests and real-world accidents under almost similar conditions. This study hypothesized that muscle activity could primarily cause the differences, and then developed a human body finite element (FE) model with individual muscles. Each muscle was modeled as a hybrid model of bar elements with active properties and solid elements with passive properties. The model without muscle activation was firstly validated against five series of cadaver test data on impact responses in the anterior-posterior direction. The model with muscle activation levels estimated based on electromyography (EMG) data was secondly validated against four series of volunteer test data on bracing effects for stiffness and thickness of an upper arm muscle, and braced driver's responses under a static environment and a brake deceleration. A muscle controller using reinforcement learning (RL), which is a mathematical model of learning process in the basal ganglia associated with human postural controls, were newly proposed to estimate muscle activity in various occupant conditions including inattentive and attentive conditions. Control of individual muscles predicted by RL reproduced more human like head-neck motions than conventional control of two groups of agonist and antagonist muscles. The model and the controller demonstrated that head-neck motions of an occupant under an impact deceleration of frontal crash were different in between a bracing condition with maximal braking force and an occupant condition predicted by RL. The model and the controller have the potential to investigate muscular effects in various occupant conditions during frontal crashes.