Robots Can Learn To Hold Knives — and Not Stab Humans

aurtherdent2000 writes "We humans enjoy not having knives inside of us. Robots don't know this (Three Laws be damned). Therefore, it's important for humans to explain this information to robots using careful training. Researchers at Cornell University are developing a co-active learning method, where humans can correct a robot's motions, showing it how to properly use objects such as knives. They use it for a robot performing grocery checkout tasks."

Comments

It shouldn't have to be pointed out

[Anonymous Coward]

If they can be taught to not stab a human...They can also be taught to stab a human. All it takes is one psychopath or curious idiot.

Re:It shouldn't have to be pointed out

[Anonymous Coward]

They don't need to be taught to stab humans, stabbing is its natural state of being. Being taught NOT to, well, that's the big news here.

Re:It shouldn't have to be pointed out

[93 Escort Wagon]

Well, sheesh - you named the robot "Stabby"... what did you THINK was going to happen?

Re:It shouldn't have to be pointed out

[dmomo]

It very much has to be pointed out. Because, to be fair.. that's how stabbing works. You point it out.

Re:It shouldn't have to be pointed out

[LifesABeach]

Sorry I'm slow this morning, I'm still working on the phrase, "... They use it for a robot performing grocery checkout tasks." I can't tell you how many times I've encountered a grocery store clerk holding a knife. It's not clear to me now how this task can be shifted over to Robots, but the day is early still.

Re:It shouldn't have to be pointed out

[lxs]

I think this is what's called emergent behavior.
The good news is that the robot gets it right after several tries so each unit is expected to operate flawlessly after disemboweling at most five grad students.

Re:It shouldn't have to be pointed out

[TheCarp]

sure but after you program it to stab and slice slabs of meat, or cut open boxes, how do you make sure it doesn't decide you must be the box it needs to open? Its not just about the action but the context; and recognizing the dirty bag of mostly water they are supposed to cut vs the one that they are not supposed to cut.

Re:It shouldn't have to be pointed out

[somersault]

I'd think it would be more like teaching it collision avoidance, and to be especially careful with certain classes of objects. Programming a car to follow a road is relatively simple. Programming it to avoid crashing into other road users and pedestrians is more complicated.

Re:It shouldn't have to be pointed out

[femtobyte]

Presumably, the knife is meant to be a "worst-case" stand-in for any object. If a robot can safely handle knives in close quarters to humans, then everything else is safe. In the grocery checkout situation, you don't want the robot to accidentally swing a can of beans through a customer's head when the absentminded customer leans over the counter to pick up the coupon they dropped.

Oblig. Futurama

[dcollins]

http://www.youtube.com/watch?v=uj2dmQruJXs

Delusional much?

[s.petry]

Robots will do what ever they are programmed to do. Programming them to recognize that stabbing someone is wrong is no different than programming them to claim stabbing is right. Simply change a 0 to a 1.

The same can be said for any act of harm mind you, not just using a knife. Smarter people than me have warned about things you should never try and teach in artificial intelligence (hinted at in TFA). The Military pretty much said "fuck them" when DARPA started developing AI to shoot and blow people up autonomously. Trying to pacify people now does what exactly? Are they going to try and convince us that nobody could ever change the bit in memory? Puhleaze!

Re:Delusional much?

[MrEricSir]

Programming them to recognize that stabbing someone is wrong is no different than programming them to claim stabbing is right. Simply change a 0 to a 1.

Given some of the atrocities in the news recently, I'm pretty sure that concern applies to us wet goo bag robots as well. But it's much easier to address systemic problems with a metal machine than an organic one.

Robots and knives

[girlintraining]

We humans enjoy not having knives inside of us. Robots don't know this (Three Laws be damned).

No, but we do enjoy programming them to put knives in humans we don't like. That's actually been a reason for much of the development of robotics: Programming them to kill for us. Scifi authors of the 50s and 60s imagined robots helping us in our daily lives -- cooking, cleaning, and today even driving us around. But whereas many have viewed the development of robotics as beneficial for mankind, the truth is much of the investment in robotics has been because of its military applications. It's just a happy accident that we've been able to declassify and repurpose much of this for private use. The google car for example, is based on technology first developed for DARPA as a way of creating vehicle that could deliver cargo to soldiers in the field.

Re:Lo, how the mighty have fallen

[glavenoid]

Nobody, for one, seems to welcome our new not-stabbing robot overlords, you insensitive clod!

Bishop's Knife Trick

[dido]

On seeing the headline I suddenly remembered this scene.

Grocery checkout tasks?

[Theaetetus]

Researchers at Cornell University are developing a co-active learning method, where humans can correct a robot's motions, showing it how to properly use objects such as knives. They use it for a robot performing grocery checkout tasks.

I believe using a knife at the grocery checkout is called armed robbery.

Robot Safety Lessons

[XMark3]

For our safety, we should teach robots what types of actions would cause the most amount of bodily harm to a human, and where all our vital organs are located, so they'll have a better idea how to behave safely around us and prevent injury. I see no possible way this could backfire.

Re:Robots Can **be programmed** To Hold Knives

[Fjandr]

Will never work. There are too many stupid humans, and they out-breed the smart humans by an enormous ratio.

So... what's new?

[jfengel]

God forbid, I actually read TFA, and I still don't get it.

As far as I can tell, it's some sort of planning exercise, an important if well-worn area of robotics. They're adding feedback, in the form of "No, this trajectory sucks". It's got nothing to do with either knives or humans, but just a "Go back and re-plan with this additional constraint".

But I can't figure out just how far it's generalizing. The trivial lesson would be "avoid this point", which is just another obstacle. I gather that it's more than that, since it took multiple trials to learn, but I can't figure out what. The human was in the same place in every trial, so it wasn't learning anything about "avoid humans". It didn't seem to be told that it couldn't go through that space with a knife but could have with, say, a dust mop.

I think I may just be misunderstanding the context of the problem. The machine has a lot of joints and there are many different plans it could use; there's an optimization problem in an enormous space. They wanted to show some kind of algorithm that could be adapted over time with user feedback, but honestly I would have assumed that was a solved problem.

So does somebody with a better understanding of actual robotics problems (as opposed to fictional ones) know what's going on here?