Can a Robot Help with the Dishes?

Robot learns to work side-by-side with humans at our pace, thanks to algorithms and Kinect sensors.

Researchers at the University of Wisconsin-Madison (UW-M) and the University of Washington (UW) have taught a robot how to help clean the kitchen.

Researchers using a Kinova Mico robot arm to do the dishes were not just being lazy. They were investigating the efficiency of interactions between robots and humans in a simple working scenario that may have implications in the world of automated robots and manufacturing.

What was special about the experiment was how the robot learned to work with its human partners.

Using a Kinect sensor, the robot could track the motion of a human arm as it moved plates from a drying rack to a shelf. The researchers reported that humans use a combination of two methods for coping with a sluggish colleague. Some wait for their partners to be ready for the next dish while others slow down. In some pairs, the individual receiving the plate was forced to work more slowly.

The robot watched these interactions and collected data on the humans. Then the robot was inserted into the experiment. The robot was tasked with handing a plate off to a human partner and used the Kinect sensor to monitor the human’s performance.

First studied were human-to-human handovers (top). Human-to-robot handovers followed, using a Kinova Mico robot arm.

First studied were human-to-human handovers (top). Human-to-robot handovers followed, using a Kinova Mico robot arm.

The algorithm behind the robot’s intelligence was able to predict its partner’s readiness with an accuracy of over 90 percent.

To up the challenge, researchers programmed the robot to employ three different strategies. With some partners, the robot would work as fast as it could, proactively holding out the dish whether or not its partner was ready to take it. Sometimes it would wait until the receiver finished stacking the previous dish, employing a reactive strategy. With a few others, it would try to adapt to the human pace.

A second level of depth involved low- and high-task demand scenarios. In the low demand setting, humans received four plates and two cups in a stack to be placed on the shelf — not that much of a challenge.

For the high-demand scenarios, the human participants had to solve mathematical problems involving nine single-digit numbers with four arithmetic operations. The problems were randomly assigned to objects that had to be placed in specific areas of the shelf, depicting the problem’s solution.

With the experiment’s conclusion, human partners were asked to rate each strategy for awareness, fluency, intelligence and patience.

Overall, it was found that the human participants preferred to work with the robot when employing reactive and adaptive strategies, even though the proactive strategy (holding the plate out whether or not the human was ready) had the highest team throughput.

“We want robots to follow our lead, or at least plan their actions with an awareness of ours,” said Bilge Mutlu, UW-M associate professor of industrial engineering and computer science. “Moving forward, we’ll want to sample a variety of tasks so that not only can we understand the common elements but also how each task varies.”

Mutlu, one of the paper’s authors, thinks robots can eventually help hand human workers parts for assembly, among other applications. I’d like to point out, however, that automated robots and the possible dangers of super-AI (artificial intelligence) scares a lot of people, especially after recent accidents. Maybe we’ll see a real-world setting of a robot like Baxter and UW-M’s algorithm later, rather than sooner.

The paper detailing the researchers’ findings was presented at Robotics Science and Systems in Rome, July 2015. To view the paper yourself, visit