Algorithm to build 3-D maps requires a low-cost camera, no human input.
Robots could one day navigate through constantly changing surroundings with virtually no input from humans, thanks to a system that allows them to build and continuously update a three-dimensional map of their environment using a low-cost camera such as Microsoft’s Kinect.
The system, being developed by researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), could also allow blind people to make their way unaided through crowded buildings such as hospitals and shopping malls.
To explore unknown environments, robots need to be able to map them as they move around — estimating the distance between themselves and nearby walls, for example — and to plan a route around any obstacles, says Maurice Fallon, a research scientist at CSAIL who is developing these systems alongside John J. Leonard, professor of mechanical and ocean engineering, and graduate student Hordur Johannsson.
But while a large amount of research has been devoted to developing one-off maps that robots can use to navigate around an area, these systems cannot adjust to changes in the surroundings over time, Fallon says: “If you see objects that were not there previously, it is difficult for a robot to incorporate that into its map.”
The new approach, based on a technique called Simultaneous Localization and Mapping (SLAM), will allow robots to constantly update a map as they learn new information over time, he says. The team has previously tested the approach on robots equipped with expensive laser-scanners, but in a paper to be presented this May at the International Conference on Robotics and Automation in St. Paul, Minn., they have now shown how a robot can locate itself in such a map with just a low-cost Kinect-like camera.
As the robot travels through an unexplored area, the Kinect sensor’s visible-light video camera and infrared depth sensor scan the surroundings, building up a 3-D model of the walls of the room and the objects within it. Then, when the robot passes through the same area again, the system compares the features of the new image it has created — including details such as the edges of walls, for example — with all the previous images it has taken until it finds a match.
At the same time, the system constantly estimates the robot’s motion, using on-board sensors that measure the distance its wheels have rotated. By combining the visual information with this motion data, it can determine where within the building the robot is positioned. Combining the two sources of information allows the system to eliminate errors that might creep in if it relied on the robot’s on-board sensors alone, Fallon says.
