Researchers from Carnegie Mellon University and Facebook Artificial Intelligence (AI) Research created a semantic navigation system for robots that uses common sense. The system allows robots to travel from point A to point B efficiently by knowing what the points are and any obstacles that stand in the way.
The new system is called SemExp, or Goal Oriented Semantic Exploration. SemExp uses machine learning to train a robot to recognize objects and understand where those objects are most likely to be. This allows the system to think strategically about how to search for something and decide the best route.
Classic systems explore a space by building a map and showing where obstacles are. With this method, robots eventually get where they are supposed to go, but will often take an unnecessary and cumbersome route.
Machine learning has been previously used to train semantic navigation systems, but it did not work efficiently because the robot must memorize objects and locations in specific environments. This does not always translate well to new environments.
Environments are complex and robot systems have a hard time generalizing what it has learned to new environments. To overcome this problem, the team made the SemExp modular system. The system uses semantic insights to determine the best place to look for an object. The robot decides where to go and use classical planning to get there.
The modular approach is efficient. The learning process can concentrate on relationships between objects and room layouts, rather than learning route planning. Semantic reasoning determines an efficient search strategy and classical navigation planning gets the robot where it needs to go.
This development will make interactions between people and robots easier. Humans will be able to tell a robot to get an item from a place and give directions and the robot will choose the most efficient way to complete the task.
This research was presented at the Computer Vision and Pattern Recognition Conference where it won the Habitat ObjectNav Challenge.