Researchers let delivery robots find front door without prior knowledge of the area
Scientists at MIT and Ford have developed a navigation method for delivery robots that doesn’t rely on pre-mapping the area with all the houses. The method relies on clues like ‘garage’ or ‘front door’ to find the right address.
According to the researchers, this technique could result in the robots taking a much shorter time to explore a terrain and find their final destination. Moreover, this does not require images of specific houses. MIT researcher Michael Everett says the technique makes it possible to essentially place the robot at the end of someone’s driveway, after which it finds the door with its own hands.
The basis of this method is the introduction of natural, semantic language for robot systems, so that the robots are trained on the basis of this. In practice, this means that when they see a door, they also process this object as a door and not simply as a rectangular obstacle. So the point is that robots get an ‘idea’ of what exactly the things in their immediate environment are.
The researchers use these kinds of semantic techniques as a kind of springboard for their new navigation method, which uses existing algorithms. These algorithms extract properties from visual data to create a new map of the same environment, displaying the context and the semantic cues.
According to the researchers, robots with these algorithms were already able to recognize and map objects in their immediate environment. Until now, however, they have not been able to make immediate decisions while navigating a new environment, such as the most efficient route to a semantic destination, such as a front door. Previously, the robot did get there, but it took a relatively long time, according to the researchers.
The improvement was made possible by translating the semantic map into a second map, in which different shades of gray indicate how close certain objects on the map are to the final destination. Dark regions mean further away from the target and lighter regions mean the robot is close. For example, the sidewalk will be darker than the lighter driveway. The new algorithm needed for this was trained on the basis of satellite images from Bing Maps, showing 77 houses from three suburban neighborhoods. The system translated a semantic map to the grayscale map, so that the robot can follow the lighter parts to its end goal.
During a test at a completely new house, in which the accumulated data set was not used, it turned out that the robot was able to find the front door 189 percent faster than with the use of classic navigation algorithms. According to the researchers, this difference can be explained by the fact that no context or meaning is used in the classical method. According to the scientists, the robots also manage to find the end goal in areas that have not yet been mapped; probably thanks to clues that match earlier areas.
The research will be presented this week at the International Conference on Intelligent Robots and Systems and is funded in part by the Ford Motor Company.