And no human ever was able to play montezuma's revenge, given only a let's play video. Usually it requires 5-10 years of accumulating prior information. It is still remarkable achievement.
This is highly significant because it was one of the games that the Atari games that the original Deep Q learning model could not beat.
DQN is solving a model-free RL problem. This method is not. You're not allowed to reset states in model-free RL. If you have access to the model/simulator which allows you to reset, you might as well use a model-based method like MCTS.