Efficient Training of Artificial Neural Networks for Autonomous Navigation

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Autonomous vehicle navigation presents complex perception challenges.
• Real-time training of artificial neural networks is crucial for dynamic environments.

Purpose of the Study:

• To develop and demonstrate efficient training techniques for autonomous driving using neural networks.
• To enable a vehicle to learn complex driving behaviors through observation.

Main Methods:

• Utilized a backpropagation neural network, ALVINN (Autonomous Land Vehicle In a Neural Network).
• Trained the network by observing a human driver's control inputs in real-time.
• Implemented training on the CMU Navlab, a modified Chevy van.

Main Results:

• ALVINN achieved autonomous driving capabilities in under 5 minutes of training.
• The system successfully navigated various road types (paved, unpaved, lined, unlined).
• Autonomous control was demonstrated at speeds up to 20 miles per hour.

Conclusions:

• Real-time learning through observation is an effective method for training autonomous driving systems.
• ALVINN demonstrates the potential for rapid development of adaptable autonomous vehicles.
• The approach shows promise for handling diverse driving conditions with minimal explicit programming.