I started my interest in AI & Super Mario here:
(which was a full 8 years ago so I’m sure techniques/computation has improved dramatically since then).
Here’s a surprisingly effective attempt by someone else 3 years ago:
The issue I had with this architecture was the lack of a time dimension. There is no facility I can see in this architecture where the bot can learn patterns of a linked series of actions, e.g. run and jump, run and double jump, etc.
But it’s surprisingly effective nonetheless.
Another video on the subject:
Very relevant project: