In the movies, robots can walk, talk and do everything from making our beds to destroying cities. But in real life, they are officially unintelligent. Sure, modern robots have been designed to put out fires, mow lawns, and cook meals, but even the most advanced robots have specific pre-defined ways of operating and are subsequently less adaptable for work outside of controlled environments such as labs or factories. If they are faced with something unexpected, like a new object or a broken part, robots are pretty much useless. Engineers all over the world are working to achieve artificial intelligence, and this week two research teams announced major breakthroughs in robotics.
Robots Learn by Trial and Error
A UC Berkeley team developed algorithms that enable robots to learn motor tasks the same way that humans do: by trial and error. Instead of programming different commands for thousands of possible scenarios, the Berkeley team took a new approach that empowered the robot to learn.
The work is part of a field of artificial intelligence called deep learning, which is inspired by the brain and involves creating “neural nets” of artificial neurons that process overlapping raw sensory data, such as sound waves or image pixels. This web of information helps a robot recognize patterns and categories among data so that it doesn’t have to be reprogramed each time it faces something new. The robot, called BRETT (Berkeley Robot for the Elimination of Tedious Tasks), was given a series of motor tasks, such as putting blocks into matching openings or stacking Legos. The algorithm controlling BRETT’s learning included a reward function that gave higher scores to movements that brought the robot closer to completing a task. The score then fed back through the neural net, so BRETT could learn which movements were working best.
“The key is that when a robot is faced with something new, we won’t have to reprogram it,” said Pieter Abbeel of UC Berkeley’s Department of Electrical Engineering and Computer Sciences. “The exact same software, which encodes how the robot can learn, was used to allow the robot to learn all the different tasks we gave it.” Their work was presented on Thursday, May 28, in Seattle at the International Conference on Robotics and Automation.
Robots Can Adapt to Injuries
Another exciting advance comes from French and U.S. researchers, who developed algorithms that empower robots to adapt quickly when they are damaged. Using this new technology, a spiderlike robot continued crawling across a floor even after one of its legs was broken. Another demonstration included a robotic arm that adapted to a broken joint and learned a new way to drop a ball into a bin.
With this new method, a computer simulation filters out all the possible solutions in order to find the ones that are the most effective and unique, so that the robot doesn’t waste time testing billions of possibilities. The robot starts by trying the most likely solution, and if that fails it continues to try other likely solutions until it succeeds. Using this technique, robots learned a new strategy to complete a task in under a minute, as compared to the several days that previous robots would have taken. Details on their research and video demonstrations published on Wednesday, May 27 in the prestigious journal Nature.
What’s Next
These new smart, adaptable robots could have some epic battles, but the researchers are more interested in practical applications. Smarter robots could be used as in-home assistants for sick or elderly people, to explore other planets, or to survey dangerous areas in natural disasters or war zones. With sophisticated algorithms for learning new skills and adapting if they are damaged, robots can help solve problems without human intervention. Improvements in artificial intelligence certainly raise ethical concerns, but don’t worry about robots taking over the world just yet. Much more testing is needed in order to prepare the robots for even the most innocent tasks in the real world.
