[Paper Review] Corner Case Generation and Analysis for Safety Assessment of Autonomous Vehicles

Haowei Sun, Shuo Feng+2 more|arXiv (Cornell University)|Feb 6, 2021

Autonomous Vehicle Technology and Safety38 references4 citations

TL;DR

This paper proposes a unified framework for generating and analyzing high-dimensional corner cases in autonomous vehicle decision-making systems using Markov Decision Processes (MDP) and deep reinforcement learning (DRL). By training background vehicle (BV) policies to behave more aggressively, the method systematically generates safety-critical scenarios, and feature extraction with PCA and clustering identifies representative, valuable corner cases—demonstrated effectively in highway driving simulations with realistic, high-impact failure modes identified.

ABSTRACT

Testing and evaluation is a crucial step in the development and deployment of Connected and Automated Vehicles (CAVs). To comprehensively evaluate the performance of CAVs, it is of necessity to test the CAVs in safety-critical scenarios, which rarely happen in naturalistic driving environment. Therefore, how to purposely and systematically generate these corner cases becomes an important problem. Most existing studies focus on generating adversarial examples for perception systems of CAVs, whereas limited efforts have been put on the decision-making systems, which is the highlight of this paper. As the CAVs need to interact with numerous background vehicles (BVs) for a long duration, variables that define the corner cases are usually high dimensional, which makes the generation a challenging problem. In this paper, a unified framework is proposed to generate corner cases for the decision-making systems. To address the challenge brought by high dimensionality, the driving environment is formulated based on Markov Decision Process, and the deep reinforcement learning techniques are applied to learn the behavior policy of BVs. With the learned policy, BVs will behave and interact with the CAVs more aggressively, resulting in more corner cases. To further analyze the generated corner cases, the techniques of feature extraction and clustering are utilized. By selecting representative cases of each cluster and outliers, the valuable corner cases can be identified from all generated corner cases. Simulation results of a highway driving environment show that the proposed methods can effectively generate and identify the valuable corner cases.

Motivation & Objective

• To address the lack of systematic methods for generating safety-critical corner cases in autonomous vehicle decision-making systems.
• To overcome the high dimensionality of real-world traffic interactions involving multiple background vehicles (BVs) in complex driving environments.
• To generate more realistic and diverse corner cases than naturalistic driving data by learning aggressive BV behavior policies via deep reinforcement learning.
• To identify and extract 'valuable' corner cases—representative clusters and outliers—for effective CAV evaluation and model improvement.
• To provide a scalable, data-driven method for safety assessment of CAVs that focuses on decision-making failures rather than perception-level adversarial examples.

Proposed method

• Formulate the traffic environment as a Markov Decision Process (MDP) to model temporal dependencies and enable policy learning for background vehicles (BVs).
• Use deep reinforcement learning (DRL) to train BV behavior policies that maximize the probability of generating corner cases during CAV interactions.
• Incorporate constraints from naturalistic driving data (NDD) to ensure generated actions remain realistic and feasible, improving simulation fidelity.
• Apply Principal Component Analysis (PCA) to reduce the dimensionality of high-dimensional corner case features for efficient clustering.
• Utilize DBSCAN clustering on PCA-projected features to identify distinct clusters of similar corner cases and outliers representing rare or extreme scenarios.
• Select representative cases from each cluster and outliers as 'valuable' corner cases for in-depth analysis and model refinement.

Experimental results

Research questions

• RQ1 How can high-dimensional, safety-critical corner cases be systematically generated for autonomous vehicle decision-making systems?
• RQ2 To what extent can deep reinforcement learning be used to train background vehicles to generate more aggressive and realistic corner cases?
• RQ3 How can feature extraction and clustering techniques identify the most informative and representative corner cases from a large set of generated scenarios?
• RQ4 What types of decision-making failures (e.g., lane conflict, aggressive cut-in) are most prevalent in generated corner cases?
• RQ5 Can constrained DRL training ensure that generated corner cases remain realistic and aligned with naturalistic driving behavior?

Key findings

• The proposed DRL-based framework successfully generated a higher density of safety-critical corner cases in highway driving simulations compared to naturalistic driving data.
• Using PCA and DBSCAN, the method identified two main clusters—'aggressive cut-in' and 'lane conflict'—and distinct outlier cases with unique failure dynamics.
• The 'aggressive cut-in' cluster was characterized by sudden lane changes by BVs leading to rear-end collisions, indicating a critical flaw in lateral collision avoidance.
• The 'lane conflict' cluster included three subtypes: side-by-side collisions, rear-end crashes during lane change, and delayed rear-end collisions after lane alignment, showing scenario-specific failure patterns.
• Outlier cases (D) revealed rare, extreme scenarios such as delayed crashes after lane change, highlighting edge cases that could be avoided with improved CAV decision-making.
• The analysis revealed that the CAV model had a significant limitation in lateral collision avoidance during lane changes, suggesting a key area for model improvement.