IE 11 is not supported. For an optimal experience visit our site on another browser.

Teaching robots to follow the leader

Researchers at the University of California at Davis hope to incorporate subtle behavioral cues that humans use to indicate where they’re going to help robots respond to directional signals.

Don’t look now, but that robot might be tailing you.

Following someone for any length of time has been a huge challenge for robots, notwithstanding WALL-E, R2-D2 and other clever movie versions. But by incorporating subtle behavioral cues that humans use to indicate where they’re going, a team of researchers at the University of California at Davis hopes to close the artificial intelligence gap with a system that helps robots easily respond to directional signals.

“The fundamental problem we were trying to look into is how to get a robot to follow people around, in buildings or in work environments, or possibly trying to get robots to follow other robots around,” said Sanjay Joshi, an associate professor of mechanical and aeronautical engineering at the University of California at Davis. With an accurate system in hand, robots could be programmed to trail doctors on their rounds, construction workers to a building site or senior citizens requiring special assistance.

Traditional following programs have relied on two techniques to guide the internal controllers of robots: one based on physics and another on images.

The physics model, Joshi said, uses the leader’s speed and direction to calculate where it is headed and then directs the follower to move to that spot. “You want to take the robot to where the leader will be in the future, time and time again, and that allows it to track the leader,” he said.

For the second technique, the follower periodically looks to see where the leader is and then heads to the position depicted in that image.

The physics model often cannot account for sudden blind turns, however, and the imaging model struggles once the leader leaves the field of view.

Developing a 'sixth sense'
Humans, on the other hand, have learned to pick up on telling body language, like the well-established observation that people tend to turn their heads slightly to the left or right just before heading off in that direction.

The trick for Joshi’s team has been working out a system that allows robots to identify and use as many of those conscious and unconscious behavioral cues as possible. “We really were trying to understand how much of the behavior cues add to the accuracy and reliability of the following,” he said.

Eventually, Joshi said, physiologists and psychologists could collaborate with computer scientists and engineers to teach robots how to automatically identify the facial or body movements or physiological reactions that come naturally to us.

“It’s building up a sixth sense about movement,” he said.

If someone is driving down a highway and sees a car moving aggressively behind her, for example, she can often sense from its movements whether it will attempt to pass or cut in front of her. That “sixth sense” is shaped by a wealth of experience observing how other drivers behave.

“The question is, how do we get robots to use that?” Joshi said.

Before answering that question, though, researchers have had to address the underlying problem of how to get robots to reliably identify and respond to basic cues.

“You have to make the computer identify the pattern of images to know, ‘This is a head turn,’ and then use the identification of that head turn in your controller,” Joshi said. And that problem can be further divided into the question of identifying a cue versus properly responding to it.

Identifying cues
For the new study, published in the August issue of the journal IEEE Transactions on Industrial Electronics, Joshi’s group used two commercially available Scorpion robots, manufactured by Pasadena, Calif.-based Evolution Robotics, to focus on the question of cue response.

The team, including graduate student Michael Chueh, and undergraduates William Au Yeung and Calvin Lei, programmed the lead robot to always follow a specific trajectory, moving in a wavy pattern down the hallway or taking a sudden, sharp turn. The second Scorpion trailed at a distance of about three yards, with a camera trained on a target attached to the lead robot’s back.

In one series of trials, the follower robot relied solely on an image-based controller to keep up with the leader. In a second, the physics-based system supplemented the imaging program. And in a third, the cue-based model joined the other two.

The study bypassed the problem of how to identify cues for turning left or right by using a human observer as a messenger in the third series of trials. Whenever the lead Scorpion approached a corner, the human observer wirelessly sent the follower a message that essentially said, “The leader just made a cue. Now it’s up to you, follower, to use that information.”

As Joshi explained, “Once you give that cue, the complication is not in accepting that cue, it’s in using the cue to extrapolate where the leader will be in the future. The cue is just the beginning of the whole process.”

The idea, he said, was for the recognition of a cue to trigger the cue-modeling program and combine its prediction of the leader’s trajectory with the other two following programs to chart the best course.

“The results were that basically if you are following a leader who always stays in your field of view, it’s good enough to just use image-based following,” Joshi said. “Once that leader leaves your field of view, that’s when our behavioral cue model really shows its value.”

Robots that used the behavioral information in their decision-making, in fact, did much better at following the leader around corners.

But what if the three techniques contradict each other and instruct the follower to go in different directions? Joshi said the information can be combined through an informational filter that weighs each of the techniques in real time and decides how much to trust each of them based on the current conditions and movements of the leader.

“As you gain experience to how believable that behavioral cue really is, you’ll update your filter,” he said. “If you had a specific human who was always turning his head but didn’t really turn, the filter will be updated so you will not trust that cue anymore. On the other hand you may find other cues that are always more reliable.”

When robots rule the roads
As for specific cues, Joshi said he’s been looking mainly at the slight movements of the head that people make just before turning to the left or right, though other cues might include a person’s state of alertness or stress. Someone who is frantically moving about may be more likely to make sudden turns. On the other hand, a person wheeling a giant bed around a hallway is more likely to make larger, gradual turns, he said.

Some cues, of course, are intentional, like the turn signals used by a driver before switching lanes or turning onto a side street. Joshi noted that some drivers prolong their signals to alert a caravan behind them to upcoming turns. One robot following another could be programmed to pick up on similar signaling cues.

“The bottom line, take-home message from our work is that it looks like behavioral cues do have potential with the robot-following problem,” Joshi said. “We think that if we can embed these cues in control systems, we can make following more reliable.”

Fernando De la Torre, co-director of the Human Sensing Lab at Carnegie Mellon University in Pittsburgh, agreed that the research question is an intriguing one. But he cautioned that a follow-the-leader system based on modest behavioral cues like head tilts, or a change in where the leader is looking, would require a high level of sophistication to cut through all of the background noise. “Can you pick up something subtle like a change in the gaze direction of only 15 degrees?” he asked.

Perhaps, he said, it would suffice if a programmer used an accurate image-recognition program to pick up on image-based cues, whether those cues are from the direction of the gaze or position of the feet. Following someone from behind obviously complicates any system based on tracking eye movement, but biometric elements, such as picking out the leader’s distinctive gait or clothes, could also help a follower keep tabs on a specific target.

De la Torre also pointed out that with both a human and robot in motion, the robot’s camera may be unstable and negatively impact its vision. Alternatively, cameras mounted at regular intervals along a hallway may be a better tracking aid for robots.

For robots following other robots, though, De la Torre said he believed the cue-based system might work well, especially if installed in autonomous vehicles.

Such a system would need to control for unexpected behavior: If a lead vehicle suddenly leaves the follower’s field of vision, the follower would need to figure out whether it should similarly strike out on a new path or continue along its normal route. A wrong decision, after all, could lead an autonomous vehicle into becoming little more than a high-tech lemming.