False positives would be scandalous. We would also need an unspoofable way for the AI to distinguish intent of the weapon drawing, because law-enforcement et cetera. The bot can't always already be at the scene when something happens.
Anyway I know at least one company working towards this goal as we speak, but I bet they aren't anywhere near that yet: http://gimltd.fi/ (Generic Intelligent Machines)
The only valid reason to draw a firearm is defense of life. If there are security robots around, one doesn't need to defend oneself or one's companions. (It's true that life may be threatened by tools of violence other than firearms, but robots might be trained to look for those as well.) Besides, by "incapacitate" I don't mean double-tap, rather something reversible like an enveloping foam etc.
...intent...
This is unnecessary scope creep. Firearms have one legitimate use, killing. They are not props to use to bolster one's argument, even though the current state of jurisprudence allows police to use them as such. Keep it in your pants until it's time. Cops brandishing firearms are barely less of a danger to the public than anyone else doing so.
The bot can't always already be at the scene...
It's true that many locations, at least at first, will not have these robots installed. However, those places that have them, will have enough of them to control extreme numbers of armed humans. Capital costs might be higher than human security, but robots don't need doughnut breaks. So "arriving at the scene and figuring it out" is also scope creep, and incidentally not something humans are good at either. The robots will always be on the scene already, so they'll always be ready to take action.
By false positives I meant non-firearm objects which the robot recognises to be firearms, and incapaciting brought tasers to mind.
I am probably looking at this problem through smaller-and-safer-country specs, Here police shootings are really rare, but so are other kinds of shootings. I thought of bots with high mobility due to the recent advancements in that sector, which would in my opinion make much more sense here than installing dozens or hundreds of stationary devices in every public space.
Congratulations on not living in a nation that has caught the mass-murder-mania. b^)
It might not make economic sense yet, but the trend only goes one direction. Electronic devices get cheaper and more capable over time. Robocop might not ever make sense, because that is a very complicated use case. However, a constellation of devices charged with enforcing a single rule, "no human hands may hold a firearm in this set of rooms", is already nearly within reach of current technology.
Anyway I know at least one company working towards this goal as we speak, but I bet they aren't anywhere near that yet: http://gimltd.fi/ (Generic Intelligent Machines)