They should minimize damage to their own occupants, but using some kind of superrational decision theory so they won't defect in prisoner's dilemmas against each other. I suspect that in sufficiently symmetric situations the result is the same as minimizing damage to everybody using causal decision theory.
When a collision is unavoidable, should a self-driving car try to maximize the survival chances of its occupants, or of all people involved?
http://www.wired.com/2014/05/the-robot-car-of-tomorrow-might-just-be-programmed-to-hit-you/