It's also a great demonstration of how fast the state of the art is advancing in machine learning. The original DQN work published by DeepMind in 2014 could competitively play simple Atari games from circa 1980. Three years later, a student project can perform adequately on MK64, released in 1996/7.
Disclaimer - didn't go through most of the content (plan to later). But possible feedback for growth would be to take the dataset from the N64 low resolution images, normalize them into coordinate pixel ratio percicions so network input dimensions can interact with more recent versions of mario Kart. You can then work with the DS mario emulator for the newer versions of mario cart (faster, more agressive bots, road detail). The point to doing this would help provide insight of how Tesla/etc can generate more training data for their auto poilets. Think about it, if the transition or generalization of data likes this is possible, you could switch game plaform and provide actual data their auto poilet systems need.
