Model-based reinforcement learning can leverage physical priors to conduct imagination rollouts, reducing the number of interactions with the real environment and avoiding dangerous states.