I think the issue with subsumption is that 3layers aren't enough. It would be interesting to combine recent deep neuralnet architectures with the Dr Brooks overall subsumption design (so multiple layers of individual deepnets layered hierarchically)
No, the problem with subsumption is that it is a poor composition mechanism. Even the name reveals this to be true: a higher layer has to subsume a lower layer, because (according to Brooks) when a higher layer is active the lower layer's outputs are suppressed. As long as the higher layer is active, the lower layer contributes nothing to the final result and hence may as well not exist. In fact, it's really not a "layered" architecture at all, it's just a big cascading series of if-then-else clauses. Subsumption doesn't really do anything other than obscure this fact.
I think that is overly harsh, the beauty of it for me continues to be that it is default stable. No input and things are stable. That makes it very resilient in low compute environments. That is the same reason I would recommend it for a smart house architecture, when the upper (planning) layer fails (or is broken) you can always get the lower layers to do what you want (lights turn on, or thermostat adjusts) because the local command the top of the 'if-then' tree in your lexicon.
But it isn't. In order for two layers to interact properly you have to specifically engineer them that way. If a higher layer starts or stops subsuming at the wrong time you can get catastrophic failures.
Also, smart houses are a very different problem than autonomous mobile robots.
That was the rhetoric. But the reality was that it was just a big kludge. That is why subsumption-based robots can't do anything more today than they could 30 years ago. Subsumption pretty much tops out at wandering randomly without bumping in to things. But that turns out to just not be a particularly hard problem.
