“Reinforcement Learning (RL) should be abandoned in favor of model predictive control, and only used to adjust the world model or critic when planning fails.”