Rod Brooks was on my thesis committee. His work was inspirational, but the subsumption architecture was badly flawed and ultimately proved to be a dead end. It's simply not true that "the world is its own best model" because wherever you are you can only see a small part of the world. It works great if all you want to do is avoid obstacles, but if you want to do anything complex and goal-directed, a robot really does have to be able to remember things. The reason he was so influential is because the dominant paradigm at the time, the so-called sense-plan-act model, was just as much of a dead end, and on top of that, produced really boring demos so it was doubly doomed. The right answer is to combine the two approaches with a layer of "glue" that connects them together in a so-called "three-layer" architecture [1]. That's actually the way that autonomous robots work today (AFAIK -- I've been out of the business for a very long time now).