杏吧原创

A robot that tricks its opponent in a game of hide and seek is a step towards machines that can intuit our thoughts, intentions and feelings

ROVIO the robotic car is creating a decoy. It trundles forward and knocks over a marker pen stood on its end. The pen is positioned along the path to a hiding place 鈥� but Rovio doesn鈥檛 hide there. It sneaks away and conceals itself elsewhere.

When a second Rovio arrives, it sees the felled pen and assumes that its prey must have passed this way. It rolls onwards, but is soon disappointed.

The behaviour of the deceptive Rovio represents something much more significant than a crude game of robot hide-and-seek. It is a demonstration of an aspect of social intelligence known as theory of mind, which humans only develop around the age of 4 or 5. If robots can be made to display theory of mind in other situations, it could endow them with a sophisticated intelligence. They might then be able to reason the thoughts, intentions and even feelings of people and other robots.

鈥淚t was definitely exciting to see it work,鈥� says of the Georgia Institute of Technology in Atlanta, who programmed Rovio with colleague Ron Arkin. 鈥淲e have expanded the boundary of understanding deception and how deception relates to artificial systems.鈥�

The defining feature of theory of mind is the ability to model the beliefs and intentions of others as distinct to one鈥檚 own. Robots have previously hinted at this ability by performing a variety of mental tricks (see 鈥淏ecoming aware鈥�).

Deception, though, is definite progress. It requires not only the modelling of a distinct mind but also the ability to anticipate and manipulate the actions of others. 鈥淎 deceiving agent knows what the other agent knows and intends to change what the other agent knows,鈥� says , a cognitive scientist at the Massachusetts Institute of Technology.

To demonstrate artificial deception, Wagner and Arkin recruited two , made by WowWee in Hong Kong, for a game of hide-and-seek. Before the game, the robots were released to learn about the game environment and the effect of their own actions on it.

The environment featured three adjacent hiding places. On the path leading towards each of these hidey-holes, the researchers placed a marker pen stood on its end (see diagram).

Programmed to learn, the first thing seeker Rovio did was move into one of the caches, knocking the pen over on its way. The pen was reset and the robot repeated the process, 10 times in all. Using a combination of its camera and probabilistic software, it learned to associate fallen pen and hiding place.

Hider Rovio came to the same conclusions as it explored the environment, but crucially, it had been given the ability to learn how to send a false signal.

The game then began. Hider Rovio鈥檚 learned knowledge allowed it to predict what seeker Rovio would do in the same situation. It calculated that knocking over a pen and sneaking elsewhere would fool its seeker. 鈥淚t uses its own model of itself to determine how best to deceive the other individual,鈥� says Wagner.

In 15 of the 20 times the game was played, the seeker chose the wrong corridor (International Journal of Social Robotics, ).

However, many researchers point out that it is unsurprising that the deceiving robot should succeed, given the extent of its pre-programming. 鈥淚t seems to me the theory of mind is in the experimenter, not the robot,鈥� says , a roboticist and evolutionary biologist at Harvard University. 鈥淚t鈥檚 like it鈥檚 staged; [the robot] is like a puppet rather than a child.鈥�

In 2009, Mitri was part of a team at the Swiss Federal Institute of Technology in Lausanne (EPFL) which created basic robots that evolved the ability to discourage other robots from accessing a common, finite 鈥渇ood鈥� source, without being programmed to do so ().

Wagner acknowledges that the robots are pre-programmed, but emphasises that each one learns on its own. 鈥淚t would learn how its movement affected the markers, which told it basically how to deceive,鈥� he says.

, who researches machine consciousness at the Carlos III University of Madrid in Spain, agrees. 鈥淚t is actually an implementation of theory of mind, because it is using the learning mechanism to update the model of the [opponent] or its own model.鈥� He notes, however, that the robot can鈥檛 transfer the knowledge autonomously to another situation. 鈥淚t is not like a human, it is something in between,鈥� he says.

Humans have a generalised concept of deception, which wasn鈥檛 demonstrated by these robots. 鈥淚t was implemented on a very specific task, for a very particular interaction,鈥� says , an AI researcher at the Rochester Institute of Technology in New York state. 鈥淚t鈥檚 a far cry from human theory of mind because it is so specific to the task.鈥�

Yet Gold acknowledges that it is a step in the right direction. 鈥淚t provides a kind of outline for things that would be fleshed out in a more general system,鈥� he says. 鈥淭here is real stuff under the hood going on here.鈥�

Already, Wagner and Arkin have implemented a more complicated version of the hide-and-seek game in a software-only simulation. In this version, the software deceiver can tailor how it deceives its opponent depending on what particular sensors it discovers its opponent has 鈥� some have audio, others vision, others infrared. This extends the degree to which the deceiver reasons about its opponent鈥檚 abilities.

By introducing games where cooperation brings benefits or penalties, the deceiver can also be programmed to decide whether or not it is worthwhile to fool its opponent, by judging what kind of task it is.

Wagner hopes to develop the deception software to help robots become more human-like in their interactions with us, for example, by enhancing a machine鈥檚 ability to reason about a person鈥檚 likes, dislikes or other emotions. 鈥淚t鈥檚 a springboard to reasoning about all kinds of other things,鈥� he says. 鈥淲e want a robot that could interact with people in any kind of situation.鈥�

鈥淯sing deception is a springboard to robots reasoning about all kinds of other things鈥�

Could such a robot ever make moral decisions? 鈥淚 think moral judgement is reserved for conscious agents,鈥� says Young. 鈥淗owever, if robots are able to compute intentions, actions and outcomes, then in a sense robots may be able to deliver the moral judgements of others.鈥�

Exploring robot theory of mind in these ways could even tell us something about ourselves, Young says. 鈥淎ny computational code that supports the external behavioural properties of theory of mind and deception can help us understand what鈥檚 going on in the human mind.鈥�