CMU Researchers Working To Build Robots That Know When They've Done A Good Job

Jun 5, 2018

Robots are able to perform a wide variety of tasks, from providing companionship to senior citizens to searching for survivors in the rubble of an earthquake, but they can't always reflect on how well they performed. 

A team of Carnegie Mellon University researchers is working with roboticists across the country on a self-assessing robot named Baxter. He's about 6 feet tall with two arms and hands that can grasp. Researchers have Baxter do things that test his dexterity, such as finding a specific box of cereal out of a bag of groceries. It's the kind of task that should be pretty easy for people, said CMU roboticist Henny Admoni, but for robots, it's much more difficult.

"Because it’s hard, it gives us a nice platform for testing sophisticated strategies for telling when the robot is doing well and when it’s failing,” Admoni said.

The goal is for the robot to reflect on how well it has done on a task and then communicate that assessment to a person.

Aaron Steinfeld, associate research professor with the CMU Robotics Institute, said the robot could use its powers of self-assessment to predict how well it could do on a new task by troubleshooting potential problems and explaining why it came to a certain confusion.

For a robot, Steinfeld said it could sound like, "I'm going to have trouble doing this task, because I don't have the [correct] hand on my arm."

The researchers involved in this five year project -- including those at Brigham Young University, Tufts University and the University of Massachusetts Lowell -- are just getting started. For now, each affiliate tests strategies independently, but eventually, they'll come together and find the best solution.

The project is part of a broader effort through the U.S. Office of Naval Research to tackle "basic science" programs, according to Steinfeld. He said robots like Baxter could one day be sent to perform tasks in locations unsafe for humans.