Achieving Socially-Aware Car-Following in Mixed Traffic: A Safe Multi-Task Deep Reinforcement Learning Approach

Abstract

Current research on autonomous vehicle (AV) car-following control primarily emphasizes optimizing AV performance without considering interactions between AVs and human-driven vehicles (HDVs). This oversight can lead to socially non-compliant AV behaviors that confuse following HDVs, destabilize traffic flow, and increase congestion and crash risks. Moreover, while most studies employ safe deep reinforcement learning (DRL) methods focused solely on AV safety, we argue that AVs, which are equipped with superior sensing and decision-making capabilities, should proactively contribute to the safety of following HDVs. To address these limitations, we propose a socially aware AV car-following controller based on social value orientation (SVO). Given the broad spectrum of SVO, we design a safe multi-task DRL (SMTDRL) framework that includes a single-task DRL backbone, a curriculum-based multi-task learning strategy for efficient cross-SVO knowledge transfer, and a differentiable safety module using control barrier functions (CBFs) to ensure safety for both the AV and its HDV followers. Comprehensive simulations demonstrate that SMTDRL improves traffic flow stability, ride comfort, and fuel economy while enforcing strict safety constraints at the platoon level, outperforming both single-task DRL and baseline multi-task DRL approaches. For example, in safety-critical scenarios, SMTDRL achieves a 100% collision-free rate, significantly outperforming baseline multi-task DRL approaches. In terms of traffic performance, it reduces the average cumulative damping ratio by approximately 7% compared to single-task DRL models. Furthermore, SMTDRL exhibits strong generalizability to unseen SVO values, highlighting its robust adaptability in mixed traffic environments.