Ive put a lot of thought into this area, and ive got this beginning self programming system going, but when it comes to making the sensors, you realize once youve got those, you dont really need to make them self program, you just code the animation yourself manually.
a) stereo cameras make the most awesome sensors for hardwired automonous death bots, and you can make them now if you are really that bloody crazy. but then its a program in control.
b) telepathic control of robots. this makes decent full body virtual reality possible (in a chair), and alot of other things. its still in its infancy alot of places, maybe not everywhere, and is a big part of the future. will be big for movies, lots of things. any invasive method i am completely against.
I always wondered how well you could spacially match a kinect up to an eeg, with a net. 
Imagine the "robo bank robber" and thats what he used to do. hehehe
c) self programming robots have to most potential (only in the end) for all terrain, and carrying out an actual possibly somewhat interesting conversation, that other robots would be too predictable, they are slightly more unpredictable. they are the smartest ones, out of them all. and thats the one i want to make. but the only interactivity it has is just treating you like a foe at this stage, no real communication happening, and it goes really slow.
