Enhancing Simulation-Based Testing for Autonomous Driving Systems

The field of autonomous driving systems (ADS) is rapidly evolving, with a strong focus on enhancing simulation-based testing methodologies to ensure safety and reliability. Recent developments highlight a shift towards more dynamic and interactive testing environments, where non-player character (NPC) vehicles can adversarially interact with the EGO vehicle to uncover critical scenarios more efficiently. This approach not only accelerates the detection of violations but also increases the proportion of violations attributed to the EGO vehicle, thereby providing more targeted insights into system vulnerabilities. Additionally, there is a growing emphasis on generating diverse and realistic road scenarios, leveraging advanced machine learning models such as multimodal Large Language Models (LLMs) to create challenging corner cases that mirror real-world conditions. These innovations aim to bridge the gap between simulated and real-world testing, ensuring that ADSs are robust against a wide range of unpredictable scenarios. Furthermore, optimization techniques like Black-Box Optimization (BBO) are being refined to cover more critical subspaces in logical scenario spaces, enhancing the comprehensive safety evaluation of ADSs. The integration of reinforcement learning (RL) for generating critical scenarios and the exploration of data scaling laws in imitation learning are also notable advancements, contributing to the broader goal of achieving safe and scalable autonomous driving systems.

Noteworthy papers include:

  • AdvFuzz: Introduces adversarial NPC vehicles to dynamically interact with the EGO vehicle, significantly increasing the efficiency of violation detection.
  • AutoScenario: Utilizes multimodal LLMs to generate realistic corner cases, enhancing the diversity and realism of test scenarios.
  • LAMBDA: Proposes a novel approach to cover critical subspaces in logical scenario spaces, accelerating the safety evaluation process.

Sources

AdvFuzz: Finding More Violations Caused by the EGO Vehicle in Simulation Testing by Adversarial NPC Vehicles

RoadGen: Generating Road Scenarios for Autonomous Vehicle Testing

Realistic Corner Case Generation for Autonomous Vehicles with Multimodal Large Language Model

LAMBDA: Covering the Multimodal Critical Scenarios for Automated Driving Systems by Search Space Quantization

Robot Learning with Super-Linear Scaling

Generating Critical Scenarios for Testing Automated Driving Systems

Preliminary Investigation into Data Scaling Laws for Imitation Learning-Based End-to-End Autonomous Driving

Testing CPS with Design Assumptions-Based Metamorphic Relations and Genetic Programming

Automated Test-Case Generation for REST APIs Using Model Inference Search Heuristic

Towards an Autonomous Test Driver: High-Performance Driver Modeling via Reinforcement Learning

Using Cooperative Co-evolutionary Search to Generate Metamorphic Test Cases for Autonomous Driving Systems

Exploring Behaviors of Hybrid Systems via the Voronoi Bias over Output Signals

Built with on top of