A learning and model-based hybrid-driven approach for safety control of autonomous vehicles under mixed traffic merging scenarios

Highway merging in mixed traffic environments, involving autonomous vehicles (AVs) and human-driven vehicles (HDVs), poses significant challenges due to the stochastic behaviors of HDVs. While reinforcement learning is adaptable, excessive exploration can pose safety risks, and existing safety mechanisms often rely on unrealistic assumptions about full environmental knowledge. To address this gap, this study proposes a multi-agent reinforcement learning framework integrated with safety potential field theory. The approach employs a crash risk-driven mechanism that assesses interaction risks using a novel relative safety potential field metric and triggers protective interventions when thresholds are exceeded. The framework utilizes the multi-agent proximal policy optimization algorithm with an attention-based neural network to ensure efficient and safe coordination of ramp and mainline AVs. Experimental results show a significant reduction in crash rates (from above 0.164 to 0.006) while maintaining high efficiency, thereby demonstrating its robustness in complex merging scenarios in mixed traffic.