A cop-controlled robot is an extension of the cop. It's not the robot's fault if its technology is misused, but the cop's. The problem underlying all this is that cops have too much unchecked power already. At least if they're shooting people remotely with armed robots, we would know the camera wasn't malfunctioning, and that the video should be available for the civil lawsuit.
Rather than freaking out about the possibility that robots might be armed, just assume that they are already weapons, and craft the human policy around the technology. I.e. when a robot is armed, the firing mechanism for the weapon must be under direct human control. If you claim your robot will never be armed or used to intimidate, you can conveniently skip out on writing that rule, so when it happens someone gets a free escape responsibility card, with "we weren't trained for this" printed on it.
Declare automated surveillance drones categorically unacceptable because of a hypothetical risk they could have weapons mounted on them and you get more people and animals killed unnecessarily.