I skimmed through the video. Doesn't sound like AI learning really at all if this what AI is all about.
Seems like the AI learning is just doing like 100,000 trial and error attempts and eventually it gets it right by piecing together what it did in the past as if every new 90 degree turn is a totally new thing in the world.
Whatever future learning it does, youd think the AI would be programmed so it takes corners and knows how to accel and brake based on what it perceives (like a human playing a racing game where someone can complete this track in one shot if they got decent enough reflexes). A human in a racing game instinctively will try to slowdown a hairpin turn and then gas it coming out of it. This AI looks like it'd just smash into a brick wall, then on attempt #2 it may take the turn correct or smash into the brick wall again. If it does take the turn correctly, it might smash head on into the next brick wall at the next turn until it retries etc...
I thought AI learning was more about the program being coded so it can interpret right away what to do well, instead of endless hours and attempts of failure before finally getting it right.