It depends what level of fail-safe we can develop. Not much you can do about the behaviour machine-learning robots which are autonomous, though! Our only defence then is a physical one.

I think largely, robots will be mind-controlled so they'll need abstract stimulus from humans? Not sure