생성형 AI 기반 고위험 시나리오 생성을 통한 자율주행 레이싱 시스템 안전성 평가 프레임워크

Abstract

This paper introduces a generative artificial intelligence (AI)-based safety evaluation framework for autonomous racing systems, focusing on efficiently searching safety-critical racing scenarios utilizing domain knowledge, optimization, and machine learning. The proposed framework consists of three main phases: 1) dataset generation, 2) conditional variational auto-encoder (CVAE) model training, and 3) safety-critical scenario generation and evaluation. In the first phase, the dynamic scenario generation is automatically processed by leveraging ontological domain knowledge and genetic algorithm to efficiently establish a potentially safetycritical driving dataset. In the second phase, we train the CVAE network with the driving dataset generated from the first phase, allowing for diverse and realistic variations in driving scenarios. In the final phase, safety-critical scenarios are generated through the trained CVAE network by adversarially variating the scenarios. Experimental results show that the proposed framework identifies various safety-critical scenarios in different racing conditions, exhibiting its effectiveness for safety evaluation of autonomous racing systems