"All robots will become better as a result of this work." That's the bold claim of University of Wyoming researcher Jeff Clune whom, together with colleagues in France, has developed a revolutionary programming algorithm to make robots "learn like we do".
The motivation for the work, published this week in Nature, was to realise the goal of using robots to do jobs that are too dangerous for humans. "But that can only work if we have a way for robots to continue to operate even when they become damaged," explains Clune.
"Right now, as soon as something goes wrong, the programming no longer works." You could develop a system so robots can teach themselves by searching through all the possibilities left open to them and adapting accordingly, "but current approaches to doing that mean that there would be more possible solutions than there are molecules on Earth!"
Instead, what was needed was a new way of working intelligently towards a solution, based on prior knowledge. "If you hurt your heel, you'd automatically walk on tiptoe," says Clune, "because you know you can do that, and that it would be a pain-free way to walk. Basically you experiment until you find an effective solution."
The team achieve the robotic equivalent by endowing their robots with a "simulated childhood".
In the same way that, as youngsters, we learn how our bodies work and what we can and cannot do, the team let their robots programming "play" in a virtual space, testing out various strategies and teaching itself how it works.
Then, when it is deployed into the field, if a part is suddenly disabled, rather than laboriously searching through all the solutions, the algorithm adopts and experiments with a series of solutions - based on its simulated childhoods - that are refined and used to inform the likely success of other solutions.
This enables the system to identify rapidly - in under two minutes - a creative approach to overcome the problem.
The algorithm can be built in to any robotic system, Clune says, and the "simulated childhood" and subsequent learning can be cloned between similar robots, making the system very efficient to employ.
"It can also be quite surprising," says Clune. "When we were testing it we told it to walk without letting its feet touch the floor. We thought this would stump it! Amazingly, it flipped itself onto its back and walked using its elbows..."