Jul 21 2020
A robot can move more efficiently from one point to another if it comprehends that the first point is the living room couch and the other one is a refrigerator, even if it is positioned in an unknown place.
Devendra Chaplot. Image Credit. Carnegie Mellon University.
This is the sensible concept behind a “semantic” navigation system that has been designed by Carnegie Mellon University and Facebook AI Research (FAIR).
Named SemExp, the new navigation system won the Habitat ObjectNav Challenge last month at the virtual Computer Vision and Pattern Recognition conference, surpassing a research group from Samsung Research China. The CMU team has recorded the second consecutive first-place finish in the annual challenge.
Goal-Oriented Semantic Exploration, or SemExp for short, makes use of machine learning to teach a robot to identify objects—recognizing the difference between an end table and a kitchen table, for example—and to realize at which locations in a home such objects can possibly be found. This allows the system to think strategically on how to look for something.
Common sense says that if you're looking for a refrigerator, you'd better go to the kitchen.
Devendra S. Chaplot, PhD Student, Machine Learning Department, Carnegie Mellon University
On the other hand, classical robotic navigation systems investigate a space by developing a map showing obstacles. Ultimately, the robot can reach the place where it has to go, but the route can be indirect.
Earlier efforts to employ machine learning to train semantic navigation systems have proved futile since they memorize objects and their places in particular surroundings. Apart from the complexity of the environments, the system usually finds it hard to generalize what it has learned from various surroundings.
In collaboration with Dhiraj Gandhi from FAIR and Abhinav Gupta, an associate professor in the Robotics Institute, and Ruslan Salakhutdinov, a professor in the Machine learning Department, Chaplot overcame the issue by developing SemExp into a modular system.
According to Chaplot, the system makes use of its semantic understandings to identify the best places to search for a particular object.
Once you decide where to go, you can just use classical planning to get you there.
Devendra S. Chaplot, PhD Student, Machine Learning Department, Carnegie Mellon University
This modular method is efficient in different ways. The process of learning can focus on relationships between room layouts and objects instead of additionally learning route planning. The semantic reasoning identifies the most efficient search scheme. Eventually, classical navigation planning makes the robot reach its destination as soon as possible.
Ultimately, semantic navigation will make it simpler for people to communicate with robots, allowing them to just instruct the robot to bring an item in a specific place, or give it directions like “go to the second door on the left.”
This study was financially supported by the U.S. Army, the Intelligence Advanced Research Projects Agency, the Office of Naval Research, and the Defense Advanced Research Projects Agency.
Common Sense Guides Robots Around the House
Researchers from Carnegie Mellon University and Facebook AI Research have created a navigation system for robots powered by common sense. The technique uses machine learning to teach robots how to recognize objects and understand where they’re likely to be found in house. The result allows the machines to search more strategically. Video Credit: Carnegie Mellon University.