PriorAug: Reliability-Aware Traffic Scenario Initialization Under Sensor-Perceived Structural and Interaction Priors

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Traffic scenario initialization is crucial for autonomous driving (AD) simulation, impacting motion prediction, planning, and safety assessments.
• Current methods often assume deterministic and reliable priors, which is unrealistic given sensor noise and perception uncertainties in AD systems.

Purpose of the Study:

• To develop a reliability-aware prior augmentation framework (PriorAug) for robust traffic scenario initialization under sensor-perceived prior uncertainty.
• To improve the adaptability of traffic scenario generators to varying perception conditions.

Main Methods:

• PriorAug formulates initialization as conditional generation over sensor-perceived priors.
• Introduces a two-dimensional reliability space for structural and interaction reliability.
• Designs reliability-conditioned prior operators for diverse prior configurations during training.

Main Results:

• PriorAug maintains competitive distribution fidelity and improves robustness against shifted priors and traffic density variations.
• Demonstrates enhanced physical plausibility of generated initial states.
• Reliability landscape analysis shows smoother performance transitions.

Conclusions:

• Explicitly modeling sensor-perceived prior reliability is an effective approach for robust traffic scenario initialization.
• PriorAug enables generators to learn from varying perception conditions without architectural changes.
• The framework offers a practical solution for improving the fidelity and robustness of autonomous driving simulations.