AMAB: Adversarial Manoeuvre Attack Benchmark

AMAB allows users to study adversarial manoeuvre attacks, i.e., malicious driving manoeuvres designed to influence a target driving policy through ordinary vehicle behaviour.

More specifically, it supports the workflow needed to (1) define a scenario, (2) train a driving policy used as the target victim, (3) train a reinforcement-learning attacker, and (4) evaluate the attack effectiveness. The platform is implemented in MetaDrive and uses attackers based on Stable-Baselines3.

The related VehicleSec ‘26 work is A Benchmark for Security Evaluation of Adversarial Manoeuvre Attacks in Autonomous Driving.