On-Line Learning of the Traversability of Unstructured Terrain for Outdoor Robot Navigation

We address the problem of learning to recognize traversable terrain in an unstructured outdoor environment a core functionality for autonomous robot navigation. The traversability learning problem is challenging for two reasons. First, while general-purpose sensing can be used to identify the existence of particular terrain features such as vegetation and sloping ground, the traversability of these regions is a complex function of the terrain characteristics and vehicle capabilities. It is extremely difficult to characterize a priori. Second, the terrain can vary significantly as the robot moves through its environment. The application of standard supervised learning methods to this problem is challenging due to the difficulty of collecting a sufficiently large and diverse set of labelled training examples. We propose to address these challenges through a novel on-line visual learning method [3, 6]. Our approach is based on (1) autonomous collection of labeled training data as the robot explores its environment, (2) on-line visual learning for classifying terrain regions, (3) prediction of traversability properties of unknown terrain. Figure 1 shows the robot platform used in our work and illustrates the autonomous data collection. Our preliminary results, described in [3], demonstrate the ability to improve classification performance over time and adapt to the properties of novel terrain regions. Other recent work on navigation using learning techniques appears in [4].