Teaching Robots to Move Like Humans

Researchers at the Georgia Tech found that when robots move in a more human-like fashion - with one movement leading into the next - people can not only better recognize what the robot is doing, but they can also better mimic it themselves. Andrea Thomaz, assistant professor in the School of Interactive Computing at Georgia Tech's College of Computing, and Ph.D. student Michael Gielniak discuss their latest study, which was presented at the Human-Robot Interaction conference in Lausanne, Switzerland on March 7, 2011.

Transcript

00:00:00 this project in particular was thinking about you robots that have natural humanlike behaviors and so you know one reason to do that is that we think that robots that have uh that meet people's social expectations in terms of their natural behavior would be easier for people to understand and understand how to approach and and understand how to interact with okay so for this

00:00:27 particular motion you're going to see the robot pick up a box and there's no context there's no box obviously um but one of the things that people responded more favorably to the human-like motion was that you can fill in some of the details that are missing in the original motion um with the humanlike motion um simply because that's one of the things that our algorithm buys you when the

00:00:49 motion was more humanlike um human beings were able to watch the motion and perceive what the robot was doing more easily were able to recognize and label the motion more accurately for the field of human robot interaction what's really interesting about this is that um typically we so we don't have a lot of metrics or quantitative things to say that one particular robot motion is more

00:01:20 or less humanlike than another um it's it's typically been just judgment calls so you can get a group of people to watch your robot uh behave in different ways and say oh yeah we liked that one better than that one and so what this does is it gives you a quantitative metric rather than relying on people giving you um their their their judgment and and so then you can use that

00:01:46 additionally now that you have a quantitative metric you can use that for optimizing Behavior so you can use that for generating lifelike Behavior rather than just uh evaluating life-like Behavior