Robots might one day trace the origin of their consciousness to recent experiments aimed at instilling them with the ability to reflect on their own thinking.

Although granting machines self-awareness might seem more like the stuff of science fiction than science, there are solid practical reasons for doing so, explains roboticist Hod Lipson at Cornell University's Computational Synthesis Laboratory.

"The greatest challenge for robots today is figuring out how to adapt to new situations," he says. "There are millions of robots out there, mostly in factories, and if everything is in the right place at the right time for them, they are superhuman in their precision, in their power, in their speed, in their ability to work repetitively 24/7 in hazardous environments—but if a bolt falls out of place, game over."

This lack of adaptability "is the reason we don't have many robots in the home, which is much more unstructured than the factory," Lipson adds. "The key is for robots to create a model of themselves to figure out what is working and not working in order to adapt."

So, Lipson and his colleagues developed a robot shaped like a four-legged starfish whose brain, or controller, developed a model of what its body was like. The researchers started the droid off with an idea of what motors and other parts it had, but not how they were arranged, and gave it a directive to move. By trial and error, receiving feedback from its sensors with each motion, the machine used repeated simulations to figure out how its body was put together and evolved an ungainly but effective form of movement all on its own. Then "we removed a leg," and over time the robot's self-image changed and learned how to move without it, Lipson says.

Now, instead of having robots modeling their own bodies Lipson and Juan Zagal, now at the University of Chile in Santiago , have developed ones that essentially reflect on their own thoughts. They achieve such thinking about thinking, or metacognition, by placing two minds in one bot. One controller was rewarded for chasing dots of blue light moving in random circular patterns and avoiding red dots as if they were poison, whereas a second controller modeled how the first behaved and whether it was successful or not.

So why might two brains be better than one? The researchers changed the rules so that chasing red dots and avoiding blue dots were rewarded instead. By reflecting on the first controller's actions, the second one could make changes to adapt to failures—for instance, it filtered sensory data to make red dots seem blue and blue dots seem red, Lipson says. In this way the robot could adapt after just four to 10 physical experiments instead of the thousands it would take using traditional evolutionary robotic techniques.

"This could lead to a way to identify dangerous situations, learning from them without having to physically go through them—that's something that's been missing in robotics," says computer scientist Josh Bongard at the University of Vermont, a past collaborator of Lipson's who did not take part in this study.

Beyond robots that think about what they are thinking, Lipson and his colleagues are also exploring if robots can model what others are thinking, a property that psychologists call "theory of mind". For instance, the team had one robot observe another wheeling about in an erratic spiraling manner toward a light. Over time, the observer could predict the other's movements well enough to know where to lay a "trap" for it on the ground. "It's basically mind reading," Lipson says.

"Our holy grail is to give machines the same kind of self-awareness capabilities that humans have," Lipson says. "This research might also shed new light on the very difficult topic of our self-awareness from a new angle—how it works, why and how it developed."

One potential application they have tested for self-aware machines is with a model bridge, with sensors continuously monitoring vibrations across its frame to develop a self-image of its "body". "In simulations we've shown that it could identify weakened joints a lot sooner than via traditional civil engineering methods," Lipson says. "The bridge isn't going to suddenly wake up one day and say hello, but in a primitive sense you can say it has self-image, enough to turn on a red light if something's wrong."

A key question for this research concerns how far it can actually go. "These are very simple robots, maybe eight or a dozen moving parts, so it's relatively easy to construct models of everything. But if you scale it up, will it still be able to make a good model of self?" Bongard asks. "That question also extends to social robots observing a human or something else complex. The question of scalability is what research is examining at the moment."

Intriguingly, the research also revealed what mental illness robots might develop. For instance, the starfishlike robot that developed a body image "spontaneously developed 'phantom limb' syndrome, thinking it had arms and legs where it didn't," Lipson says. "As robots become more complex and evolve themselves, we could see the same kinds of disorders we [humans can] have appear in machines."

Lipson detailed his team's research February 19 at the annual meeting of the American Association for Advancement of Science conference in Washington, D.C.