This is what I've said plenty of times, actually. I think there was a paper, but it basically described how a series of relatively simple learned behaviors could be put together to execute complex tasks.
Possibly outing myself if someone recalls it, but I employed the same principle when coding a maze-runner in College Robotics. Move forward, watch for openings. Measure the distance of any detected opening to determine if it is likely a valid turn. If a valid turn is detected, make it by slowing the inside wheel in order to minimize speed loss. And of course, a straightening routine consisting of a jag away and then slightly less back the other way if a wall is bumped. That's all it took to execute a maze run smooth and fast enough to post a time that feels like it may still be the fastest after... well, more than a decade.
Biting the hand that feeds IT
