> A number of the current hardware random number generators use either resistor noise or balanced diodes.
Yes, by sampling avalanche noise etc. But these devices have to be carefully timed and balanced to eliminate sampling biases. I doubt that such finicky devices would be used in remotely deployed systems. Indeed, the pseudo-random generators tend to be far more useful, in a systems engineering sense, because test sequences can be easily generated by repeating a seed number, for regression testing etc.
In any case, I don't think human free-will (if it indeed exists) is the same as perfect randomness, because I think we'll all agree that human behavior is somewhat predictable on a broad scale. But at the smallest scale, human actions can seem to be very "surprising", yet simultaneously "intelligent". Is that free will? I don't really know.
But so far, no functioning robot, AFAIK, has made any decision that it was not programmed to make, even if it was just flipping random switches.