Lower limb exoskeletons have complex dynamics that mimic human motion. They need to be able to replicate lower limb motion such as walking. The trajectory of the exoskeleton joints and the control signal generated are essential to the system's operation. Current learning from demonstration methods has only been combined with linear quadratic regulators; this limits the applicability of processes since most robotic systems have non-linear dynamics. The Asynchronous Multi-Body Framework simulates the dynamics and allows for real-time control. Eleven gait cycle demonstrations were recorded from volunteers using motion capture and encoded using Task Parameterized Gaussian mixture models. An iterative linear quadratic regulator is used to find an optimal control signal to drive the exoskeleton joints through the desired trajectories. A PD controller is added as a feed-forward control component for unmodeled dynamics and optimized using the Bayesian Information Criterion. We show how the trajectory is learned, and the control signal is optimized by reducing the required bins for learning. The framework presented produces optimal control signals to allow the exoskeleton's legs to follow human motion demonstrations.