Robots Learn Faster with Online Input

Robots like to learn. Yet how they go about it could be enhanced with help from the online community instead of just one person.  

"We are trying to create a method for a robot to seek help from the whole world when it is puzzled by something," said Rajesh Rao, associate professor of computer science and engineering at the University of Washington in Seattle, via a news release. "This is a way to go beyond just one-on-one interaction between a human and a robot by also learning from other humans around the world."

Just like humans, robots often learn by imitating what they've already seen from others. However, many researchers believe that gaining help from crowd-sourcing could assist the efficiency of their learning. 

"Because our robots use machine-learning techniques, they require a lot of data to build accurate models of the task. The more data they have, the better model they can build. Our solution is to get that data from crowdsourcing," added Maya Cakmak, a UW assistant professor of computer science and engineering.

For their study, researchers asked for guidance from the online community to help a robot accomplish a model-building task. Online participants were asked to build a simple model out of colored Lego blocks. Next, they called on the robot to build a similar extension of their creation.

With the help of employees from Amazon Mechanical Turk, the robot searched for the best models to mimic a new one. This type of learning is referred to as "goal-based imitation," as it works with robot's growing ability to determine what their human operators want. For example, a robot could gather the important qualities of a human building a turtle model and try to replicate the same idea, yet it may be simpler for the robot to construct.  

"The end result is still a turtle, but it's something that is manageable for the robot and similar enough to the original model, so it achieves the same goal," Cakmak explained.

In a second part of the experiment, researchers manipulated learning actions on a two-armed robot through physical demonstration. With the help of abstract, interactive visualizations of the action, the robots sought out the crowd's guidance to discover new ways of performing actions. 

As scientists are determining further how crowd-sourcing could help robots learn more efficiently, research teams at Brown University, Worcester Polytechnic Institute and Cornell University look for more assistance through the online world. 

This work will be presented at the Conference on Human Computation and Crowdsourcing in November.