• Giancarlo Sturla and Matthew Gombolay (front) collaborating with the PR2 robot on an assembly task

    Giancarlo Sturla and Matthew Gombolay (front) collaborating with the PR2 robot on an assembly task

    Photo: Jason Dorfman/CSAIL

    Full Screen

Want a happy worker? Let robots take control.

CSAIL study finds that human subjects prefer when robots give the orders. Watch Video

Press Contact

Adam Conner-Simons
Email: aconner@csail.mit.edu
Phone: 617-324-9135
MIT Computer Science & Artificial Intelligence Lab

If you’ve seen a sci-fi flick with autonomous robots in the last 40 years, you may be wary of giving robots any semblance of control.

But new research coming out of MIT’s Computer Science and Artificial Intelligence Lab (CSAIL) suggests that letting robots have control over human tasks in manufacturing is not just more efficient — it’s actually preferred by workers.

While manufacturers have long recognized the benefits of automation in streamlining processes and freeing humans from tedious tasks, such as aisle-running, there’s always a concern that workers may feel devalued or even replaceable.

“In our research we were seeking to find that sweet spot for ensuring that the human workforce is both satisfied and productive,” says project lead Matthew Gombolay, a PhD student at CSAIL. “We discovered that the answer is to actually give machines more autonomy, if it helps people to work together more fluently with robot teammates.”

Specifically, in the study, groups of two humans and one robot worked together in one of three conditions: manual (all tasks allocated by a human); fully autonomous (all tasks allocated by the robot); and semi-autonomous (one human allocates tasks to self, and a robot allocates tasks to other human).

The fully-autonomous condition proved to be not only the most effective for the task, but also the method preferred by human workers. The workers were more likely to say that the robots “better understood them” and “improved the efficiency of the team.”

See how CSAIL researchers experiment with human-robot collaboration in the workspace.

Video courtesy of the researchers

Gombolay emphasizes that giving robots control doesn’t mean a team of cyborgs will be running the show. It means the tasks are delegated, scheduled, and coordinated via a human-generated algorithm.

“Instead of coming up with a plan by hand, it’s about developing tools to help create plans automatically,” he said.

The algorithm can also conduct on-the-fly replanning, instantly developing an alternate “schedule” for a task if, say, a new part arrives or a machine malfunctions — a clear advantage over its human counterparts, who generally require time to call an audible.

The research — developed by Gombolay, MIT undergraduates Reymundo Gutierrez and Giancarlo Sturla, and assistant professor Julie Shah in the Interactive Robotics Group at CSAIL— is part of a long line of recent advances that allow robots to interact in less predictable environments, and to therefore collaborate directly with human workers in factory settings.

Gombolay says that, in the future, similar algorithms could be applied to human-human collaboration (like scheduling hospital resources), search-and-rescue drones, and even one-on-one, human-robot collaboration in which the robot could help someone with discrete building and construction tasks.

Topics: Robotics, Robots, Artificial intelligence, Computer Science and Artificial Intelligence Laboratory (CSAIL), Aeronautical and astronautical engineering, Electrical Engineering & Computer Science (eecs), School of Engineering


A few inconsistencies and weaknesses:
-The comment "The robot worker does not understand what I am trying to accomplish" at 2:03 is seemingly contradictory to the main message of the article or the video comment "the autonomous robot better understood [the workers]" at 2:12.
-The graph at 2:52 is seemingly too good to be true, whatever the undefined "value" parameter is, straight lines make the graph look made up, bias and unscientific.
-It's hard to ascertain what scheduling the robot did and how it did it. From observation, it seems like software alone could've done the trick just as well.

Back to the top