Integrating Justice Theory into Moral Decision-Making for Autonomous Vehicles

As autonomous vehicles (AVs) become integral to future intelligent transportation systems, ensuring the morality of their decisions in complex, real-world scenarios remains a critical challenge. This paper addresses the limitations of current AV decision-making frameworks, which are constrained by restricted moral environments and ethical theories. We propose a novel reinforcement learning (RL) framework that integrates a broader range of moral theories, including justice, deontology, and utilitarianism, to navigate moral uncertainty. By simulating diverse, ethically challenging scenarios inspired by the Moral Machine framework, we evaluate the effectiveness of two voting mechanisms-Nash and Variance voting-in balancing competing ethical principles. Our findings show that Variance voting outperforms Nash voting in both sequential and non-sequential environments, offering a more nuanced and adaptable approach to ethical decision-making in AVs. Additionally, we introduce new reward structures and dimensions, such as action and dimensional harms, to provide a deeper understanding of the consequences of AV actions. This work contributes to the advancement of ethical AV systems by aligning their behavior more closely with human values and legal norms.

I. Introduction

Autonomous vehicles (AVs) are poised to revolutionize future transportation systems [1], largely due to advancements in reinforcement learning (RL) that enable these systems to make intelligent decisions in complex environments [2]. While modern RL algorithms can effectively support current AV adoption, new ethical challenges emerge as AVs are increasingly integrated into public roadways. A particularly pressing issue is the behavior of AVs in morally uncertain situations, where predefined rules may not suffice, and the consequences of their actions can have significant ethical implications [3], [4].