GA2AD: a generalized adaptive adversarial training framework considering surrounding hybrid risk field for autonomous driving

Ensuring autonomous vehicle safety in mixed traffic remains difficult because human-driven vehicles can create rare and unpredictable interactions that are underrepresented in conventional training. Here we present GA2AD, a generalized adaptive adversarial training framework that generates challenging but controllable driving scenarios for autonomous vehicle decision-making. The framework treats the target autonomous vehicle model as a black box and trains a background-vehicle model to create adversarial maneuvers, including cutting in, hard braking and speeding. A hybrid risk field estimates the interaction risk imposed by surrounding vehicles, while an activation function selects when adversarial maneuvers should be applied. An adaptive switching module alternates training between the target vehicle and the background-vehicle model according to collision intervals, avoiding overly weak or overly severe adversarial conditions. In simulation, GA2AD reduces collision rates by 55-70% and lowers the 95th percentile absolute jerk by 21-42%, with only small decreases in average speed. Tests on an autonomous vehicle track using synchronized virtual traffic further show improved safety and comfort under the tested adversarial conditions. These results indicate that adaptive adversarial scenario generation can improve autonomous vehicle robustness in complex mixed traffic.