One, "bringing back" sufficient situational awareness to the pilot/operator. Maybe a complete bubble/mock-up of a cockpit would do it. I still think dogfighting is going to be problematic. Strike, sure, we have RPAs that do that now. BVR air-to-air missile engagements, not much different. But mixing it up, that is going to be a problem.
Two, the data link. For an RPA, by definition there has to be a data link between aircraft and pilot. Something bringing back sensor data and sending out command and control. That link is a vulnerability. It isn't stealthy, it is an emitter. It can be jammed/taken-down. It can maybe even be hacked and you could find your own weapons turned against you. (remote possibility, we'll assume their encryption is good enough) If you put in a local autonomous mode in case the link is lost that is fine, but it is going to drastically reduce mission effectiveness while increasing risk of loss of the RPA. (An RPA that has gone "lost-link" and is executing an autonomous RTB is going to be a sitting duck for an enemy to engage air-to-air.)
So I don't know. Interesting technical and operational challenges.
Morons. Not gonna work in our lifetime, maybe never. Would anybody fly in a drone airliner? Air Combat is 10000 times more complex...
French ingenuity: the pilot can surrender before the flight takes off.