One thing robots are notoriously bad at is learning by doing. You can pack plenty of information into a robotic brain, but ask a bot to teach itself a new motor task — even one as simple as stacking blocks or unscrewing a water bottle — and you're probably out of luck.
That, however, might be changing very soon. Researchers at UC Berkeley are now developing algorithms that robots can use to learn all sorts of tasks through trial and error, just like humans do. In practical terms, this could eventually lead to home service robots capable of handling any number of tedious tasks we'd rather not do — screwing in lightbulbs, plunging toilets, folding laundry.
Traditionally, robots make their way through the world with a vast amount of pre-programming that equips them to handle a range of scenarios. While this works reasonably well in controlled environments — laboratories or medical facilities, for instance — learning to adapt to the unknown is a critical step our robots will need to take if they're ever going to become more integrated into our daily lives.
To that end, researchers involved in Berkeley's "People and Robotics Initiative" are turning to a new branch of artificial intelligence known as deep learning, which draws inspiration from how the human brain's neural circuitry perceives and interacts with the world.
"For all our versatility, humans are not born with a repertoire of behaviours that can be deployed like a Swiss army knife, and we do not need to be programmed," said robotics researcher Sergey Levine in a press release. "Instead, we learn new skills over the course of our life from experience and from other humans. This learning process is so deeply rooted in our nervous system, that we cannot even communicate to another person precisely how the resulting skill should be executed. We can at best hope to offer pointers and guidance as they learn it on their own."
If you've ever used Siri, Google's speech-to-text program, or Google Street View, you've already benefited from recent advances in this field. But applying deep learning to motor skills has proven much more challenging. In terms of complexity, physical tasks go far beyond passive recognition of sights or sounds.
In recent experiments, researchers have been working with a small personal robot they call the "Berkeley Robot for the Elimination of Tedious Tasks", or BRETT. For his training, BRETT is presented with a series of simple motor tasks, such as placing pegs into holes or stacking LEGO bricks. The algorithm controlling BRETT's learning includes a reward function that scores BRETT based on how well he learns a new task. That reward system is key: Movements that bring BRETT closer to completing his task score higher than those that don't, and this information is relayed across thousands of parameters in his neural net.
So far, the results of BRETT's training have been astounding. Given the location of objects in a scene, BRETT is typically able to master a new assignment within ten minutes. If BRETT doesn't have the location of objects and instead needs to learn vision and motor control together, the process can take several hours.
"We still have a long way to go before our robots can learn to clean a house or sort laundry, but our initial results indicate that these kinds of deep learning techniques can have a transformative effect in terms of enabling robots to learn complex tasks entirely from scratch," said Pieter Abbeel of UC Berkeley's Department of Electrical Engineering and Computer Sciences. "In the next five to 10 years, we may see significant advances in robot learning capabilities through this line of work."
As someone who's about to commit three solid days to spring cleaning, this is very heartening news.