Can a robot be programmed to detect a trap?


One problem in robotics is to allow the robot to detect danger. It’s harder than it looks; it is a question of evaluating the uncertainties. In their open access article, a Harvey Mudd College research group describes the situation in dramatic terms by asking human readers to imagine a dilemma:

Georges montañez

Imagine that a wealthy individual announces that he has hidden a large sum of money in an abandoned mine. You feel particularly adventurous and visit the mine in search of treasure. As you approach one of the mine’s many entrances, your excitement collapses when you notice the dangerous conditions. The precarious wooden floor boards that separate you from a 50 foot drop are worn and rotten. Streaks of crumbling rock intermittently fall from the roof and walls, indicating potential collapse at any time. You slowly realize that it may not be accidental; perhaps the mine owner intended to make the situation hopelessly dangerous. As you examine the space with growing skepticism, you notice strange beam and rope structures attached to some of the platforms – their trap-like appearance triggers additional red flags. By weighing your safety against the possibility of a reward, you decide that the perilous quest is not worth the risk.

C. Hom et al., “The Gopher’s Gambit: Survival Benefits of Artefact-Based Intention Perception.” 13th International Conference on Agents and Artificial Intelligence (ICAART 2021), online, February 4-6, 2021.

The problem of moving forward may seem “obvious” to humans. But transmitting this meaning of “red flag” to a robot is another matter. The robot has no independent life experience to rely on.

Can we realize the potential danger? The Harvey Mudd team, led by Assistant Professor George Montañez, decided to use digital waffles instead of humans (waffles are simpler) to test the possibilities of programming robots to detect the intentions behind the circumstances ( as in “it looks like a trap …”):

The computer simulation created virtual waffles and “tested whether intention could be perceived through artifacts, namely structures in an environment that could potentially be traps,” Montanez added. In addition, the study investigated whether the “virtual ground squirrel’s knowledge of whether the structures were intentional or accidental could improve survival rates of artificial agents (in our case, artificial ground squirrels)”.

Marjorie Hecht, “Harvey Mudd’s Virtual Gopher Study Examines How Detecting Deliberate Traps Improves Survival” at Daily science of today (November 15, 2021) The article is open access.

Digital waffles face risks. The one in the middle made an unlucky choice.

It’s not a simple decision: the way the program works is if the “waffle” avoids too many possible food sources. he will starve, but if he throws himself into too many structures in search of food, he will be trapped.

What about the result of coding gophers to have – or not to have – the perception of intention?

The researchers found that “those ground squirrels who were able to accurately determine whether the structures were designed with intent to harm allowed intention-perception ground squirrels to avoid many deadly traps and, therefore, live longer lives. long time.

Marjorie Hecht, “Harvey Mudd’s Virtual Gopher Study Examines How Detecting Deliberate Traps Improves Survival” at Daily science of today (November 15, 2021) The article is open access.

The benefits were greater than expected. From the article: “We find that ground squirrels with the ability to perceive intention have significantly better survival outcomes than those without perceived intention in most of the cases evaluated. “

Perception of intent, if it can be generalized, could prevent complex robotic devices from being recovered or destroyed by relatively straightforward hacks.

The team is planning two more papers, focusing on the advantages of direct intention detection over indirect detection.

Natural gopher:


Comments are closed.