Advancing Machine Learning Safety in Autonomous Vehicles: Strategies, Innovations, & Applications

Abstract

Autonomous vehicles, driven by cutting-edge machine learning (ML) technologies, have the potential to revolutionize transportation. However, ensuring their safety and reliability remains a significant challenge. This presentation examines the gaps in current automotive safety standards, such as ISO 26262 and SOTIF, when applied to ML-driven systems, and introduces innovative solutions to address these shortcomings. Key strategies discussed include the integration of robust error detection mechanisms to ensure safe failure modes and the strengthening of ML algorithms’ resilience across diverse operational environments.

The session showcases real-world implementations, such as a student model for predicting steering control failures in NVIDIA’s PilotNet, which has reduced errors in real-time operations. Additionally, a self-supervised out-of-distribution (OOD) detector, capable of identifying hazardous conditions, has improved safety margins. The presentation also covers a cross-domain object detection model for UAVs, demonstrating enhanced robustness against weather, altitude, and viewpoint variations.

Future research directions, including adversarial attack defense mechanisms, are also explored, highlighting their potential to reduce attack success rates. The importance of interdisciplinary collaboration between software, hardware, and human-computer interaction experts is emphasized, providing new pathways for addressing ML safety challenges in autonomous vehicles.

By bridging the gap between theoretical safety frameworks and practical solutions, this presentation offers actionable insights aimed at enhancing the reliability and security of autonomous driving systems, contributing to the evolution of next-generation automotive safety standards.