The main thing I want to try, however, is attaching some PC components as add-on's for image processing.
From a cognitive stand-point, a thinking, autonomous robot must have some of the same capabilities as a human does. Currently, robots are stupid and dumb because they lack these capabilities for thinking, as well as the ability to sense accurately. DARPA challengers failed because they lacked the processing power to see and analyze in real-time, something a human has the power to do. If we could harness an actual CPU (and maybe score with a CUDA GPU, who knows?), an autonomous program is definitely feasible as long as the problem is approached from the correct angle, by trying to emulate a human's problem-solving skills. Otherwise, it'll eventually just encounter an anomaly and be unable to correct itself.
/rant, it's late at night
