Adversarial Safety-Critical Scenario Generation using Naturalistic Human Driving Priors

TL;DR

This work tackles the challenge of evaluating autonomous driving decision-making by generating safety-critical test scenarios that are both adversarial and natural. It combines IDM/MOBIL-based calibration with Generative Adversarial Imitation Learning (GAIL) to form a human driving prior, and uses Proximal Policy Optimization (PPO) to train a natural adversarial policy guided by a dual reward: adversariality and naturalness. The framework is validated on real-world datasets (NGSIM and INTERACTION) and shows improved realism and safety-relevant stress testing compared to baselines, with reduced collision rates and more plausible driving behavior. This approach offers a scalable, data-driven way to stress-test AV decision-making in a way that reflects real-world driving while still exposing adversarial situations.

Abstract

Evaluating the decision-making system is indispensable in developing autonomous vehicles, while realistic and challenging safety-critical test scenarios play a crucial role. Obtaining these scenarios is non-trivial, thanks to the long-tailed distribution, sparsity, and rarity in real-world data sets. To tackle this problem, in this paper, we introduce a natural adversarial scenario generation solution using naturalistic human driving priors and reinforcement learning techniques. By doing this, we can obtain large-scale test scenarios that are both diverse and realistic. Specifically, we build a simulation environment that mimics natural traffic interaction scenarios. Informed by this environment, we implement a two-stage procedure. The first stage incorporates conventional rule-based models, e.g., IDM~(Intelligent Driver Model) and MOBIL~(Minimizing Overall Braking Induced by Lane changes) model, to coarsely and discretely capture and calibrate key control parameters from the real-world dataset. Next, we leverage GAIL~(Generative Adversarial Imitation Learning) to represent driver behaviors continuously. The derived GAIL can be further used to design a PPO~(Proximal Policy Optimization)-based actor-critic network framework to fine-tune the reward function, and then optimizes our natural adversarial scenario generation solution. Extensive experiments have been conducted in the NGSIM dataset including the trajectory of 3,000 vehicles. Essential traffic parameters were measured in comparison with the baseline model, e.g., the collision rate, accelerations, steering, and the number of lane changes. Our findings demonstrate that the proposed model can generate realistic safety-critical test scenarios covering both naturalness and adversariality, which can be a cornerstone for the development of autonomous vehicles.

Figures (17)

• Figure 1: Examples of counter-intuitive adversarial behaviors in an urban scenario. @CARLA dosovitskiy2017carla
• Figure 2: The overall framework of our safety-critical scenario generation solution.
• Figure 3: The update process of actor-network and critic-network of PPO algorithms.
• Figure 4: The network structure of PPO-enabled GAIL.
• Figure 5: Constructing road structures from the real world datasets.