Autonomous Cross-Country Navigation

Safe navigation over unstructured terrain is critical for autonomous scouting vehicles and work machines operating in mines, construction sites, or unknown areas. Numerous obstacle avoidance systems have been built for indoor mobile robots, where the obstacles are objects such as furniture, walls, and people, and the floors are flat and level. The hazards are considerably more varied and harder to avoid for outdoor robots operating in unstructured terrain. In addition to collisions with objects such as trees and rocks, the robot must recognize and avoid steep or pockmarked terrain resulting in tipover or high centering. Hazardous objects such as tree stumps must be distinguished from harmless objects such as weeds and small bushes. Greater speeds may be necessary in outdoor environments, factoring vehicle dynamics into fray along with kinematics and statics.

Early systems focussed on slow navigation in barren terrain. A number of systems were developed that used range sensing to detect obstacles and drive wheeled vehicles around them at walking speed or slower. At CMU, the Environmental Institute of Michigan (ERIM) laser rangefinder was used to guide the NAVigational LABoratory II (NAVLAB II) robot around obstacles. The NAVLAB II is a Highly Mobile Multi-Wheeled Vehicle (HMMWV), the Army's current jeep. As speeds greater than 5 m.p.h. were attempted, dynamics became more important and the performance of the existing systems suffered. Furthermore, densely vegetated areas the robot could not always opt for the conservative approach of avoiding every plant. Subsequent research by two of my former students (now graduated) is addressing these two problems.

The RANGER system (Real-Time Autonomous Navigator with Geometric Engine), developed by Alonzo Kelly, primarily addresses the issue of higher-speed navigation. As the robot drives, RANGER takes range data from a laser rangefinder or stereo vision system and builds an elevation map of the terrain in front of the robot. Using a vehicle model, RANGER simulates motion of the vehicle over the terrain and measures the likely hazard for each steering angle in a candidate set. The hazards include tipover, collision, and rough response in roll and pitch. RANGER then selects the steering angle with the lowest hazard and transmits this command to the vehicle controller and the process repeats. In order to support higher-speed navigation, RANGER dynamically subselects from all available sensor data to process only that data essential for computing the next steering command, thus minimizing computational requirements. Furthermore, RANGER incorporates a dynamic model for vehicle steering essential for accurately predicting motion at higher speeds. RANGER has successfully driven the NAVLAB II at a maximum speed of 10 m.p.h. for a distance of 10 miles.

The MAMMOTH system (Modular Architecture Multi-Modality Theory), developed by Ian Davis, primarily addresses the issue of handling vegetation. MAMMOTH starts with the successful ALVINN work (Autonomous Land Vehicle in a Neural Network) for road following and extends it for use in cross-country applications. The motivation behind MAMMOTH is twofold: 1) vegetation is difficult to characterize using simple physical models, and 2) multiple features are needed to distinguish vegetation from other natural objects. Like ALVINN, MAMMOTH uses a neural network to learn a model for vegetation by associating video data with a human-driven classification of terrain. Once this neural module has been trained to recognize vegetation features, it is combined with a module trained to recognize large geometric shapes from range data, and the resultant system is trained to avoid any geometric object that is not a small plant. The advantage of this modular approach is that the user of the system can focus the learning on key features in the data, thus increasing confidence that the system will act properly in critical situations, and reducing the overall training time required to achieve this performance. In preliminary tests, MAMMOTH has succeeded in driving the NAVLAB II through dense vegetation, avoiding only those obstacles that pose a hazard to the vehicle.

Back to home page.