Image-based path planning for outdoor mobile robots

Mobile robots operating in natural terrain need some sort of long-range perception in order to navigate efficiently. Whereas LADAR is a commonly used sensor on such systems, providing range data out to 25 m and beyond, we have instead focused on what information can be extracted from vision. Our robot has only two stereo camera pairs for terrain sensing; they provide reliable stereo data up to 5 m away, but this is not enough to prevent myopic behavior. To overcome this problem, we have developed a novel approach to navigation using monocular imagery by planning a path in the image space. We take a monocular image and apply a learned color-to-cost mapping to transform the raw image into a cost image. Then, after a pseudo-configuration-space transform, we search for a pixel-to-pixel path from a point in front of the robot to the projected goal point in the cost image. Our implementation has been shown to react to obstacles at a range of 93 m, far beyond the range of our stereo perception. We describe the details of our method and results from testing under the DARPA Learning Applied to Ground Robots (LAGR) program and discuss the characteristics and trade-offs of our approach. © 2009 Wiley Periodicals, Inc.