Actually the whole purpose of modern military training is precisely to psychologically condition people to kill. Back in the "old" days (ww2), in a typical battle only 15% might actively participate. By the time of the Mai Li massacre the use of psychological conditioning had already brought kill rates up to 85%.
Besides the ease of which it enables civilian massacres and war crimes, from Mail Li to Falluja, the other problem with psychological conditioning is that soldiers are discharged cheaply, but there is no immediate "off" switch to such training. But governments care not, for it is not part of the military budget to also return people sanely to civilian life. Some end up in violent incidents and hence the result is damage in the civilian population, but many choose to kill themselves. Just as one example, the suicide death rate for U.S. soldiers who had served in Iraq is actually higher than the battlefield casualty rate! That is more soldiers kill themselves than die in battle today.
The problem is often not that soldiers will not kill civilians, such as to protect an unpopular leader. The problem rather is that most nations cannot afford the psychological conditioning and training needed to maintain a force that will, certainly on a large scale. This was the dilemma faced for example by the Chinese government at Tiananmen Square, who back then did not have the resources to condition a military that completely, though eventually they found units from the countryside who had no connection to the region that would kill.
War robotics can however do more than simply remove people (who may still control them) from combat. It can be used to remove people from knowledge of who is being killed and why, particularly useful when using such troops in local suppression. Imagine if they are told they are fighting a terrorist group in the midst of a city in Afghanistan, with all the audio falsely altered so the language people are speaking no longer sounds English, and the video feeds scrubbed of other identifying features, when in reality they are controlling robots suppressing a domestic protest in Detroit? Of course, if people are no longer needed to control them, then even this issue is eliminated. In this, I agree an AI that follows orders without conscience would be the very best friend of a modern police state.