Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment.

Area of Science:

• Autonomous vehicle safety
• Artificial intelligence in transportation
• Simulation-based testing

Background:

• Autonomous vehicle (AV) development relies heavily on testing in realistic driving simulations.
• Current methods require extensive mileage (hundreds of millions of miles) due to environmental complexity and rare critical events, proving inefficient.
• Demonstrating AV safety performance necessitates overcoming these mileage limitations.

Purpose of the Study:

• To introduce a novel testing environment that drastically reduces the mileage needed for AV safety validation.
• To maintain the unbiasedness of evaluation while improving testing efficiency.
• To create an intelligent testing environment that learns to challenge AV systems effectively.

Main Methods:

• Development of a 'naturalistic and adversarial driving environment' by introducing sparse, adversarial adjustments to standard simulations.
• Training background vehicles within the simulation to intelligently execute adversarial maneuvers at opportune moments.
• Evaluation of the proposed environment using highway driving simulations.

Main Results:

• The naturalistic and adversarial driving environment significantly reduces the number of miles required for testing compared to purely naturalistic environments.
• The proposed method accelerates the evaluation process by several orders of magnitude.
• Evaluation unbiasedness is preserved despite the adversarial nature of the testing.

Conclusions:

• The naturalistic and adversarial driving environment offers a highly efficient solution for testing autonomous vehicle intelligence.
• This intelligent testing approach accelerates the development and deployment of safer autonomous vehicles.
• Adversarial adjustments in simulations are a viable strategy to overcome the limitations of traditional AV testing methods.