Performance Evaluation of a Terrain Traversability Learning Algorithm in The DARPA LAGR Program

The Defense Applied Research Projects Agency (DARPA) Learning Applied to Ground Vehicles (LAGR) program aims to develop algorithms for autonomous vehicle navigation that learn how to operate in complex terrain. For the LAGR program, The National Institute of Standards and Technology (NIST) has embedded learning into a control system architecture called 4D/RCS to enable the small robot used in the program to learn to navigate through a range of terrain types. This paper describes performance evaluation experiments on one of the algorithms developed under the program to learn terrain traversability. The algorithm uses color and texture to build models describing regions of terrain seen by the vehicle’s stereo cameras. Range measurements from stereo are used to assign traversability measures to the regions. The assumption is made that regions that look alike have similar traversability. Thus, regions that match one of the models inherit the traversability stored in the model. This allows all areas of images seen by the vehicle to be classified, and enables a path planner to determine a traversable path to the goal. The algorithm is evaluated by comparison with ground truth generated by a human observer. A graphical user interface (GUI) was developed that displays an image and randomly generates a point to be classified. The human assigns a traversability label to the point, and the learning algorithm associates its own label with the point. When a large number of such points have been labeled across a sequence of images, the performance of the learning algorithm is determined in terms of error rates. The learning algorithm is outlined in the paper, and results of performance evaluation are described.