Maybe that was just to show the recreation. An engine is pure code and a lot of times closed source. Just creating animation and similitude in that ways does not recreate a engine but could create those aspects. Engines have many aspects of game mechanics many you do not really see directly but as a byproduct.

But the impressive sell of this technique was the AI discerned enough patterns and physics from the reference video to auto generate a game engine that replicated/simulated the original game. This alludes that it understood the player character and now can control the sprite representing the player character via unique paths through the game map.

It's waaaayy less impressive if it's just programmatically processing video frames, tracking pixels to generate coalescence, generating a library of sprites based on pixel coalescence, and then playing back the same sequence of sprites programmatically...