|NeuroSolutions Home | Products | Downloads | Resources | Support | Order||NeuroDimension Home|
NeuroSolutions for Excel
Custom Solution Wizard
NS for MATLAB
Source Code License
Neural Network Course
Free Evaluation Copy
Intro to Neural Networks
News and Reviews
Licensed User Center
Order Online Now
ND Corporate Website
Neural Network Technology used in the Development of Autonomous Ground Vehicles
Gainesville, FL -- August 15th, 2003 -- NeuroDimension, Inc. (www.nd.com) announces the decision to sponsor American Industrial Magic, LLC (http://aimagic.org) for the DARPA Grand Challenge Race (www.darpa.mil/grandchallenge/). The Defense Advanced Research Projects Agency (DARPA) will give $1 million to the team whose robotic car drives itself the fastest from Los Angeles to Las Vegas, on an off-road course. The race, which must be won within 10 hours, will take place on March 13 next year.
American Industrial Magic, LLC is an all volunteer organization of scientists and engineers formed to produce racing vehicles for the Grand Challenge. At the moment the team consists of about 53 people working from three continents. They have reserved three race slots under the vehicle names AV Andrea Morgan, AV Sydney Bristow, and AV Wendy Darling. The AV Sydney Bristow is going to be an H1 Hummer, the AV Wendy Darling is a 6x6 Army cargo truck and the identity of AIM's mystery vehicle - AV Andrea Morgan - remains a secret.
The team is constructing six robot vehicles. They are the three primary race vehicles and their duplicate back up vehicles. At the last moment they will select the best working of each to run in the DARPA Grand Challenge Race. It is planned that all six will act as support vehicles for the team and drive themselves from Traverse City, MI to the Los Angeles start of the race, a distance of about 2,240 miles and the best three will run the 300 mile race then all six will drive themselves back to Traverse City, MI, an additional 2,240 miles. NeuroSolutions (www.neurosolutions.com) has been chosen as the primary artificial intelligence engine that will be necessary to accomplish these tasks.
The NeuroSolutions' neural network engine will be used to accomplish processing of the sonar and radar signals to produce the visual imagery the robot will use for navigation. In addition, networks will be created for motor control to calculate the maximum amount of force that can be applied to the wheels at any given time.
The simpler of the two tasks will consist of a total of five networks connected in star fashion. The entry point into the network layer will be given desired speed, needed acceleration/deceleration rate, amount of turn to apply (left/right, +/-) and needed turn rate. The output from this network will be fed into the four other networks in the form of needed torque and needed speed. This network will also receive feedback from the motor control hardware of current RPM and current power usage. This information, with data from the main sonar and radar systems of current speed, will be used to detect wheel slippage that the network will try to minimize in real time. The network will be responsible for keeping the robot moving in the needed direction and speed by synchronizing each of the drive wheels.
The Motor control networks that receive the needed speed and torque information will be trained to maximize the power of the motor. To accomplish this, the network will receive feedback of the current RPM and power consumption, for the each wheel.
The sonar and radar systems are broken down into several smaller networks consisting of:
The sonar data will be obtained from three transducers spaced equal distance apart on a plane. This data, along with the actual sonar pulse, will be supplied to temporal axons connected to the next layer using SOFM for pre-processing. This will allow the network to generalize the incoming signals and to simplify the input stream without losing too much accuracy. The output of this layer will be fed into each of the five different networks for specific processing.
The terrain scanning sonar net will be used to scan the terrain similar to a flatbed scanner as the robot moves across the ground. When there is an obstacle between the robot and the place on the ground (like a hill or ridge) where the sample is to be taken this network will be trained to return -1 to signify this event. The output will consist of the elevation of the terrain at a specified distance.
The LOS (Line of Sight) Obstacle Tracking Net will return the distance to the closest obstacle that is at least a minimum height off the ground (beyond vehicle ground clearance) and note it as an object to be avoided.
The Grade Detection Net will be used to measure the elevation of the terrain in front of the robot out to a specified distance in a straight line. This will allow the robot to stop and rotate left and right to get a better picture of the terrain in the event there is too much unknown or greater detail is needed. This will also allow it to sense the general grade change of the terrain in front of the robot.
The Terrain Detection Network will be used to categorize the current type of terrain that the robot is on into five different categories that it has been trained to recognize. The Speed Detection Network will be used to gauge the speed of the robot both forward and sideways. This information will then be fed into the other areas of the robot to determine current position and also for use with the motor control networks.
Following the example of recent successes with adaptive flight control, the team is planning to produce an adaptive driving system what will relearn the vehicles' dynamic characteristics in real time as driving conditions change.
About NeuroDimension, Inc.
Headquartered in Gainesville, Florida, NeuroDimension (www.nd.com) is the world's leading provider of neural network development tools. Neural networks are a form of artificial intelligence that enables the computer to "learn" in a way similar to the human mind. Modeled after neurons in the brain, neural networks analyze data to detect patterns or trends. Since NeuroSolutions was introduced in 1994, it has attracted thousands of users from around the globe, making it the standard in neural network development tools. The company has also used NeuroSolutions as the basis for its TradingSolutions product, in addition to research grants, neural network courses, books, and consulting.
[an error occurred while processing this directive]