[Paper Review] A Real-time Critical-scenario-generation Framework for Testing Autonomous Driving System

Yizhou Xie, Kunpeng Dai+1 more|arXiv (Cornell University)|Jun 2, 2022

Reinforcement Learning in Robotics4 citations

TL;DR

This paper proposes a real-time critical-scenario-generation (RTCSG) framework for testing autonomous driving systems, using an aggressive-driving algorithm with a self-adaptive cost function to efficiently generate failure-inducing scenarios. The method achieves 16× higher efficiency than reinforcement learning-based approaches while maintaining superior stability and success rates in generating critical scenarios for black-box controllers.

ABSTRACT

In order to find the most likely failure scenarios which may occur under certain given operation domain, critical-scenario-based test is supposed as an effective and widely used method, which gives suggestions for designers to improve the developing algorithm. However, for the state of art, critical-scenario generation approaches commonly utilize random-search or reinforcement learning methods to generate series of scenarios for a specific algorithm, which takes amounts of computing resource for testing a developing target that is always changing, and inapplicable for testing a real-time system. In this paper, we proposed a real-time critical-scenario-generation (RTCSG) framework to address the above challenges. In our framework, an aggressive-driving algorithm is proposed in controlling the virtual agent vehicles, a specially designed cost function is presented to guide scenarios to evolve towards critical conditions, and a self-adaptive coefficient iteration is designed that enable the approach to operate successfully in different conditions. With our proposed method, the critical-scenarios can be directly generated for the target under test which is a black-box system, and the real-time critical-scenario test can be brought into reality. The simulation results show that our approach is able to obtain more critical scenarios in most conditions than current methods, with a higher stability of success. For a real-time testing, our approach improves the efficiency around 16 times.

Motivation & Objective

• To address the inefficiency and inapplicability of existing scenario generation methods for real-time, black-box autonomous driving systems.
• To develop a framework that generates critical scenarios on-the-fly without relying on expensive learning or rewindable simulation.
• To improve stability and success rate in generating failure-inducing scenarios under diverse initial conditions.
• To enable direct, real-time testing of autonomous driving controllers by generating adversarial scenarios tailored to the target system.

Proposed method

• The framework employs an aggressive-driving algorithm to simulate adversarial behavior in virtual agents, guiding them toward critical interactions with the ego vehicle.
• A custom cost function is designed to quantify danger levels based on proximity, relative velocity, and trajectory deviation, promoting near-miss conditions without extreme maneuvers.
• Self-adaptive coefficient iteration dynamically adjusts the cost function parameters to maintain performance across varying initial conditions and improve convergence.
• The method operates online, generating scenarios in real time without requiring simulation rollback or long training phases.
• The approach is applied directly to black-box controllers, enabling scenario generation without access to internal logic or control policies.
• The framework is evaluated through simulation using a variety of initial states, with performance measured via scenario score and step efficiency.

Figure 1 : Diagram of a vehicle-vehicle cut-in scenario.

Experimental results

Research questions

• RQ1 Can a real-time scenario generation framework be developed that efficiently targets black-box autonomous driving controllers without relying on reinforcement learning or simulation rollback?
• RQ2 How does the proposed cost function with self-adaptive coefficients improve scenario generation stability and success rate across diverse initial conditions?
• RQ3 To what extent does the RTCSG framework outperform RL-based and RRT-based methods in terms of generation efficiency and scenario quality?
• RQ4 Can the framework consistently generate critical scenarios—especially near-misses—under adverse initial conditions where other methods fail?
• RQ5 How does the framework’s performance scale across a wide range of initial relative positions and velocities between ego and agent vehicles?

Key findings

• The RTCSG framework achieved a median scenario score of 0.9433, significantly higher than RL-based (0.8728) and RRT-based (0.8661) methods, indicating better scenario quality and stability.
• In 85% of tested initial conditions, RTCSG outperformed both RL-based and RRT-based methods in scenario score, demonstrating superior reliability.
• The framework improved generation efficiency by a median factor of 16.0025 times compared to the RL-based method, as measured by step-spend-ratio.
• RTCSG showed minimal mild and no extreme outliers in performance, indicating consistent and robust behavior across diverse scenarios.
• The method maintained high performance even under adverse initial conditions (e.g., large initial distance or velocity difference), where RL and RRT methods often failed or required excessive steps.
• The self-adaptive coefficient iteration significantly enhanced the framework’s ability to converge under varying initial states, reducing sensitivity to initialization.

Figure 2 : Structure of RTCSG framework.