Robots Learn To Deceive Other Robots (And You) At Hide And Seek
Writing by Evan Ackerman on Friday, 10 of September , 2010 at 12:13 am
Last August, we posted about some swarm robots at EPFL that evolved deceptive behaviors to fool other robots into starving to death (or something). These behaviors were passive, however, in that the robots weren’t making intelligent decisions to deceive other robots, but simply executing evolved behaviors where deception improved overall fitness.
Researchers at Georgia Tech have taught a robot to be able to identify situations where deception might be a good idea, and then use deceptive tactics to help it accomplish its goals. While the immediately obvious application is militaryish (and the research is sponsored by the Office of Naval Research), deception is potentially important in lots of other situations. One example is a search and rescue scenario, where a robot might be better off deceiving victims than informing them that (say) there’s a 93.77% chance that everything is about to go terribly horribly wrong.
Their first step was to teach the deceiving robot how to recognize a situation that warranted the use of deception. A situation had to satisfy two key conditions to warrant deception — there must be conflict between the deceiving robot and the seeker, and the deceiver must benefit from the deception. Once a situation was deemed to warrant deception, the robot carried out a deceptive act by providing a false communication to benefit itself. The technique developed by the Georgia Tech researchers based a robot’s deceptive action selection on its understanding of the individual robot it was attempting to deceive.
To test out the deception programming, the researchers had their robots play hide and seek. The hiding robot had a choice of three hiding spots, each with a set of markers in front of it. Recognizing that the situation could benefit from a deceptive tactic, the hiding robot would knock down markers in front of one hiding spot, and then go find a different spot, avoiding the markers. This technique fooled the seeking robot 100% of the time, as long as the hiding robot was able to knock down the correct markers.
One of the researchers involved in this project is Ron Arkin, whom you may remember from a few previous posts on robot ethics. And it’s no coincidence that he’s involved in sneaky robots:
While there may be advantages to creating robots with the capacity for deception, there are also ethical implications that need to be considered to ensure that these creations are consistent with the overall expectations and well-being of society, according to the researchers.
“We have been concerned from the very beginning with the ethical implications related to the creation of robots capable of deception and we understand that there are beneficial and deleterious aspects,” explained Arkin. “We strongly encourage discussion about the appropriateness of deceptive robots to determine what, if any, regulations or guidelines should constrain the development of these systems.”
Robots capable of deception is a pretty serious thing. Of course, it has to be understood that when we say robots being deceptive, we mean robots that we’ve programmed to execute behaviors that appear deceptive… It’s not that inherently, some robots are somehow deceptive. I suppose you could argue that there’s no functional difference, but the important thing to remember is that robots do what we tell them to do. If we tell them to be sneaky, they’ll be sneaky. And if there are potential ethical issues with any of this, look to the humans, not the robots.
[ Project Website ] VIA [ Physorg ] and [ GT News ]
Comments (5)
- Add this post to
- Del.icio.us -
- Digg
Comment by David
Made Friday, 10 of September , 2010 at 2:59 pm
Run that by me again.
They have programmed a robot to knock down some indicator markers and then go some where else.
And they have programmed another robot to look in places indicated by the knocked down markers.
Who is deceiving whom? (not the same as “what is deceiving what”)
Comment by Evan Ackerman
Made Friday, 10 of September , 2010 at 3:36 pm
@David The hiding robot knows that the seeking robot uses markers to determine where to search, so the hiding robot intentionally knocks down markers leading to one hiding place, and then goes and hides somewhere else. The seeking robot has no special programming and is not expecting to be deceived, it just looks for indicators as to where the hiding robot might be (in this case, the markers). Does that make sense?
Comment by David
Made Saturday, 11 of September , 2010 at 4:37 am
So the robot ‘knows’.
- Did it sort of look about and just realise that there was a connection between knocking down some markers and being found? Or was it in the code?
- Was it programmed to learn to arrange the world so it wouldn’t be found? Markers being the only variable.
- If the seeking robot is looking for indicators of where the hiding robot might be then markers still standing would be a good pointer. Or was it programmed to look for knocked over markers?
As I wrote before, “Who is deceiving whom?” it still looks as if it is humans who are being deceived.
Comment by Evan Ackerman
Made Saturday, 11 of September , 2010 at 1:28 pm
I’ll have to go read the actual paper, but I believe that the hiding robot “knows” that the seeking robot uses the markers to determine where to search… As in, it was programmed with that information. I think the hiding robot itself determined that it could use the technique of knocking over markers to spoof the seeking robot, though. From the project site:
“We use an interdependence theory framework and game theoretic notation to develop algorithms which allow a robot or artificial agent to recognize situations that warrant deception and to select the best deceptive strategy given knowledge of the mark (the individual being deceived).”
The seeking robot is “dumb,” it isn’t programmed to look for deception; it’s simply programmed to look for indicators as to where the hiding robot may have gone. But again, the hiding robot knows this, which is why it knocks over markers leading somewhere else in the first place.
Comment by quantum_flux
Made Saturday, 2 of October , 2010 at 9:51 pm
Primitive, yet funny nonetheless.
