Online learning of multiple perceptual models for navigation in unknown terrain

Autonomous robots in unknown and unstructured environments must be able to distinguish safe and unsafe terrain in order to navigate effectively. Stereo depth data is effective in the near field, but agents should also be able to observe and learn perceptual models for identifying traversable surfaces and obstacles in the far field. As the robot passes through the environment however, the appearance of ground plane and obstacles may vary, for example in open fields versus tree cover or paved versus gravel or dirt tracks. In this paper we describe a working robot navigation system based primarily on colour imaging, which learns sets of fast, efficient density-based models online. As the robot moves through the environment the system chooses whether to apply current models, discard inappropriate models or acquire new ones. These models operate on complex natural images and are acquired and used in real time as the robot navigates.