Author(s)

Zhongyuan Jiang

Corresponding Author

Zhongyuan Jiang

Abstract

Ethical decision-making in autonomous driving remains a fundamental challenge, particularly in dilemma scenarios where vehicles must weigh passenger safety against potential harm to pedestrians or property. In this work, we propose a reinforcement learning framework enhanced with a Social Variational Autoencoder (Social-VAE) to capture interactive behaviors among traffic participants. Ethical dilemmas are modeled through parametric two-lane scenarios, incorporating sensitive variables such as pedestrian identity, passenger composition, and traffic signal states. A nonlinear loss function balances vehicle self-damage and pedestrian harm, enabling continuous control over steering, acceleration, and braking. Training and evaluation are conducted in the CARLA simulator using the INTERACTION dataset to ensure realistic multi-agent dynamics. For benchmarking, we adopt the PCLA leaderboard evaluation protocol, which provides standardized comparison across safety, efficiency, and ethical trade-offs. Our results demonstrate that the proposed framework achieves improved decision consistency and robustness in ethically challenging scenarios, bridging moral reasoning with practical control policies in autonomous vehicles.

Keywords

Ethical decision-making in autonomous driving remains a fundamental challenge, particularly in dilemma scenarios where vehicles must weigh passenger safety against potential harm to pedestrians or property. In this work, we propose a reinforcement learning framework enhanced with a Social Variational Autoencoder (Social-VAE) to capture interactive behaviors among traffic participants. Ethical dilemmas are modeled through parametric two-lane scenarios, incorporating sensitive variables such as pe