A couple of months ago, Jean-Baptiste Mouret, a roboticist at the Université Pierre et Marie Curie in Paris, was asked to demo the resilience of his insect-like robots for a group of famous scientists. His robots were programmed to figure out how to keep walking even if one of their six legs was broken. In the days before the demonstration, Mouret and his grad students tested the robots again and again, and were certain they’d show off their ability to quickly relearn how to walk after being maimed for science. That would help prove the technology could be used to power robots for search-and-rescue missions or healthcare, scenarios where things tend to go wrong and robots would need to adapt on the fly.
The morning of the demonstration though, Mouret and his team arrived at the university and discovered to their dismay that the floors had been waxed. The administration wanted to make a good impression, but Mouret worried it would result in a terrible one: The robots hadn’t been trained to walk on slippery surfaces.
It could have been a disaster. Robots these days are trained to do a specific job under very specific circumstances. Take them out of their “comfort zone,” and they almost always poop out or work in ways they shouldn’t. That’s one of the key reasons we don’t have more robots in our homes. This limitation makes them unreliable and, in certain circumstances, downright dangerous.
But Mouret’s robots were different. The “brains” controlling them were designed to deal with off-script situations like this one. Within minutes of turning the robots on, they’d adapted to the slippery floors, learning to walk with their legs in a more vertical position, which helped them avoid pushing too much on their hind legs and slipping.
“This makes the algorithm even more interesting. Subtle, almost invisible differences, like the exact angle between the feet and the floor, can have a huge impact in performance,” Mouret told me.
Programming how every minute detail might affect performance would require developers to predict and hand-code every situation a robot might find itself in. Because that’s just not feasible, researchers are trying to make robots more adaptive, so they can help do things like assess nuclear power plant meltdowns or search for survivors of natural disasters. That means upgrading robots from dumb devices to creative machines with a little bit of intuition so they can get themselves out of tricky situations. In essence, they’re trying to make robots think more like humans.
To make their bots adaptable when it came to difficulty walking, Mouret’s team created a simulated environment in which a robobrain could “experience” some 13,000 different ways to walk. It then ranked these different gaits and clustered them according to how useful they might be in certain situations, like if a robot’s right leg suddenly got chopped off. If that were to happen, the robot could look to that no-right-leg cluster and quickly experiment with those pared-down alternatives instead of sifting through all 13,000 options one by one. For the researchers, this was just a proof-of-concept, but in an emergency situation, that capability could prove to be enormously important.
Once the software had been trained in a virtual environment, the researchers uploaded it to commercial six-legged robots, did a little “robot torture,” and set them loose in their Parisian lab. One robot had half its leg yanked off. Others had one or more legs missing. The researchers cut off power to one robot’s leg.
“We basically abused these robots in a lot of different ways,” said Jeff Clune, one of the researchers collaborating with Mouret on this project, which is described in a paper published today in the journal Nature. “Current robots — the second that anything goes wrong, that they’re damaged in any way — they become basically inoperable….This [algorithm] allows any robot that’s out there…to basically keep performing its task until you can take it in for maintenance.”
Within about two minutes, the robots had figured out an alternative way to walk, and their gait was between three and seven times faster compared to injured robots loaded with an algorithm that didn’t cluster behaviors. Even uninjured bots performed better, scouring across the floor at speeds 30 percent faster. These abilities could make their algorithm handy for search-and-rescue applications, once its beefed up to deal with more complex situations and environments.
In theory, the researchers say, the same algorithm should work for any robot. Mouret’s team used the same algorithm for the six-legged robots and a robotic jointed arm. It allows for robotic creativity even absent damage. Another experiment, which the researchers dubbed the ‘Ministry of Silly Walks’ after the famous Monty Python skit, challenged one of their hexapods to find the fastest way to walk without setting a robofoot on the ground. After some trial and error, the robot figured out it could flip itself upside down and walk on its knees.
“It’s extraordinarily creative,” says Clune. “That’s one of the great things about working with evolutionary techniques. They often do things that you would never imagine.”
Evolutionary algorithms are decades old, but recently a handful of artificial intelligence experts — Mouret’s team included — have turned their attention back to evolutionary computing in an attempt to develop intelligence that mimics biological brains.
Chris Adami, an artificial-intelligence expert who’s working on evolutionary algorithms at Michigan State University, says Mouret’s study “provides an important guidepost to what’s possible.” “It’s not AI the way it’s always been done,” Adami told me. “This is a very different approach.”
The project is still in its early stages and Mouret and his team haven’t started talking to robotics companies — such as the legion now owned by Google — which have more resources at their disposal, to really scale these algorithms up. Another key thing to bear in mind is that even though the software is incredibly good at honing in on optimal behaviors, it doesn’t really understand what’s going on.
“It’s as different as finding a way to limp with a painful knee and knowing exactly what’s wrong by visiting a doctor,” Mouret told me. “We’re not trying to understand what’s wrong. We’re just trying to find something that works in spite of the damage.” But, he says, it’s not unlike what animals do in the wild to survive, and that’s the point.