acm-header
Sign In

Communications of the ACM

ACM TechNews

Helping Robots Learn What They Can and Can't Do in New Situations


View as: Print Mobile App Share: Send by email Share on reddit Share on StumbleUpon Share on Hacker News Share on Tweeter Share on Facebook
Illustration of a quizzical robot.

What should a robot do when it cannot trust the model it was trained on?

Credit: pinimg.com

University of Michigan researchers have developed a method of helping robots to predict when the model on which they were trained is unreliable, and to learn from interacting with the environment.

Their approach involved creating a simple model of a rope's dynamics while moving it around an open space, adding obstacles, creating a classifier that learned when the model was reliable without learning how the rope interacted with the objects, and including recovery steps for when the classifier determined the model was unreliable.

The researchers found their approach was successful 84% of the time, versus 18% for a full dynamics model, which aims to incorporate all possible scenarios.

The approach also was successful in two real-world settings that involved grabbing a phone charging cable, and manipulating hoses and straps under a car's hood.

Michigan's Dmitry Berenson said, "This method can allow robots to generalize their knowledge to new situations that they have never encountered before."

From The Michigan Engineer News Center
View Full Article

 

Abstracts Copyright © 2021 SmithBucklin, Washington, DC, USA


 

No entries found