Techniques for utilizing a Monte Carlo model to perform pre-training of a ground vehicle controller. A sampled distribution of actions and corresponding states can be utilized to train a reinforcement learning controller policy, learn an action-value function, or select a set of control parameters with a predetermined loss.