Autonomous characterization of unknown environments

The key to the autonomous exploration of an unknown area by a scientific robotic rover is the ability of the vehicle to autonomously recognize objects of interest and generalize about the region. This paper presents a Bayesian framework under which a mobile robot can learn how different classes of objects are distributed over a geographical region, using imperfect observations and non-random sampling. This yields dramatic improvements in classification accuracy by exploiting the interdependencies between objects in an area and allows the robot to autonomously characterize the region. This is demonstrated with data from Carnegie Mellon University´s Nomad robot in Antarctica, where it traversed the ice sheet, classifying rocks in its path.