Robots need to be taught by people to perform specific tasks. The traditional thinking has been: Train em’ up and let ‘em work while their human co-workers go on to do other things. Now, new research out of MIT posits that robotic and human “co-workers” can work better together by cross-training each other.
In heavy industrial applications like automotive manufacturing, robots perform tasks that are too big or too dangerous for human workers. In such jobs they are isolated from people for safety reasons. But what if human and robotic workers need to work together efficiently and safely in close proximity? That’s a problem Julie Shah, head of the Interactive Robotics Group at MIT’s Computer Science and Artificial Intelligence Lab (CSAIL) is attacking with new research. According to a statement Shaw made to MIT News:
“People aren’t robots, they don’t do things the same way every single time … so there is a mismatch between the way we program robots to perform tasks in exactly the same way each time and what we need them to do if they are going to work in concert with people.”
Traditionally, human trainers reward robots when they do the task correctly and provide negative feedback when they fail to do so. But military research has shown that a similar approach with people alone is inefficient and does not encourage teamwork. That led Shaw to look into other ways to train people to work together well and to see if those methods could also apply to mixed teams of robots and people.
Cross-training — in which team members switch roles on different days — gives members a better idea of how their individual work affects their co-workers. To bring robots into the fold, Shaw and Ph.D. candidate Stefanos Nikolaidis built a algorithm to teach devices how to learn from their role-swapping experiences.
They ended up tweaking existing reinforcement-learning algorithms to allow the robots to take in not only positive and negative rewards, but also other information pertinent to the job at hand. Their findings? Mixed teams that cross-trained were much more efficient than mixed teams that used the older interactive reward method.
The researchers discovered that the amount of time that robots and people that cross-trained were able to work concurrently rose 71 percent while concurrent work time by the other teams fell 41 percent. This is an important measure: If human team members have to wait for a robot to complete a task (or vice versa) before resuming work, that’s a lot of down time. Cross-trained robotic/human co-workers can work in tandem and get more done much more of the time than teams trained the old way.
Shah and Nikolaidis will present their research at the International Conference on Human-Robot Interaction in Tokyo next month. Check out the videos on their work . If this research pans out, look for more mixed workgroups on the factory floor and some pretty impressive productivity gains.