Breaking through safety performance stagnation in autonomous vehicles with dense learning.

Area of Science:

• Autonomous Driving Systems
• Artificial Intelligence
• Machine Learning

Background:

• Autonomous vehicles' commercial adoption is hindered by safety performance stagnation.
• Current deep learning methods struggle with rare safety-critical events, leading to the seesaw effect where improvements in one area cause regressions in others.

Purpose of the Study:

• To introduce a novel dense learning approach for autonomous vehicles.
• To overcome the limitations of existing deep learning methods by prioritizing informative failures and successes.

Main Methods:

• Developed a dense learning approach that samples data proportionally to its contribution to policy gradient and exposure frequency, excluding non-informative samples.
• Trained a safety-critical driving agent for a highly automated vehicle using mixed reality on an urban test track.

Main Results:

• The dense learning approach successfully broke performance stagnation in autonomous driving agents.
• Enhanced the model's overall safety performance by one to two orders of magnitude.
• Enabled training for tasks previously intractable for existing deep learning methods.

Conclusions:

• The proposed dense learning method significantly reduces learning variance without introducing bias.
• This approach represents a significant advancement towards achieving human-level safety in autonomous vehicles.
• Paves the way for the widespread adoption of autonomous vehicles.