Carnegie Mellon researchers are developing a software control system for cross country autonomous vehicles called RANGER, for Real-time Autonomous Navigator with a Geometric Engine. The goal of the project is to increase speed and enhance the reliability of robotic vehicles in rugged outdoor settings.

RANGER has navigated over distances of 15 autonomous kilometers, moving continuously, and has at times reached speeds of 15 km/hr. The system has been used successfully on a converted U.S. Army jeep called the NAVLAB II and on a specialized Lunar Rover vehicle that may, one day, explore the moon.

The key to the success of the system is its adaptability. It explicitly computes the vehicle reaction time and required sensory throughput and adapts its perception and planning systems to meet the demands of the moment.

At sufficiently high speeds, it becomes necessary to explicitly account for the difference between the ideal response of the vehicle to its commands and its actual response. RANGER models the vehicle as a dynamic system in the sense of modern control theory. The linear system model is expressed in the following generic block diagram.

FIFO queues and time tags are used to model the delays associated with physical i/o and to register contemporary events in time. The command vector u includes the steering, brake, and throttle commands. The disturbances ud model the terrain contact constraint. The state vector x includes the 3D position and 3 axis orientation of the vehicle body as well as its linear and angular velocity. The system dynamics matrix A propagates the state of the vehicle forward in time. The output vector y is a time continuous expression of predicted hazards where each element of the vector is a different hazard.

The process of predicting hazardous conditions involves the numerical solution of the equations of motion while enforcing the constraint that the vehicle remain in contact with the terrain. This process is a feedforward process where the current vehicle state furnishes the initial conditions for numerical integration. The feedforward approach to hazard assessment imparts high-speed stability to both goal-seeking and hazard avoidance behaviors.

System components above the state space model in the software hierarchy translate the hazard signals y(t) into a vote vector. This is accomplished by integrating out the time dimension to generate a vote for each steering direction based on a normalization of the worst case of all of the considered hazards.

A new range image perception algorithm has been developed for RANGER. It selectively extracts a very small portion of each range image in order to reduce the perceptual throughput to a bare minimum. In this way, vehicle speed is less limited by the computer speed.

The algorithm searches each image for a band of geometry that is between two range extremes, called the range window as shown in the figure below of a range image of a hill in front of the vehicle. Only the data between the white lines is processed by RANGER. The algorithm also accounts for vehicle speed by moving the range window out as speeds increase.

The approach also stabilizes the sensor in software because the search for the data of interest adapts automatically to both the shape of the terrain and the attitude of the vehicle. It is up to 6000 times faster than traditional approaches and it achieves the throughput necessary for 20 m.p.h. motion on an ordinary computer workstation.