Safety-oriented decision-making for connected and automated vehicles in highway mixed-autonomy traffic via dynamic attention-enhanced graph reinforcement learning

Ensuring safe and efficient driving in highway weaving sections remains a formidable challenge for connected and automated vehicles (CAVs) operating alongside human-driven vehicles (HDVs). The complex interplay of merging, diverging, and through traffic requires decision-making systems capable of accurately perceiving spatiotemporal interactions and deriving reliable driving policies. This paper proposes a dynamic attention-enhanced graph reinforcement learning (DAEGRL) framework that advances CAV decision-making through three synergistic innovations. First, a dynamic dual-gated multi-head graph attention network (D2G-MGAT) is developed to model spatial vehicle interactions with adaptive head weighting and selective activation, producing refined spatial feature embeddings that capture safety-critical inter-vehicle dependencies. Second, a decision-token-augmented spatiotemporal transformer (DASTformer) extends these spatial representations into the temporal domain through a learnable decision token that aggregates a sequence of historical spatial embeddings. Third, a dynamic attention-based multi-critic proximal policy optimisation (DAMC-PPO) algorithm is introduced to stabilise policy training by aggregating multiple independently parameterised value estimates through state-conditioned attention. Extensive experiments conducted in a representative three-lane highway weaving scenario demonstrate that DAEGRL achieves the highest average speed and cumulative reward and the lowest return variability among the evaluated graph-RL methods. No collisions are observed for DAEGRL across the test episodes under the standard condition; it also records lower surrogate conflict measures and smoother longitudinal behaviour than selected graph-RL baselines. A comparison with a conventional model-based benchmark shows more favourable surrogate-safety and comfort outcomes on the majority of evaluated metrics and higher efficiency for the proposed method, while fixed-policy tests assess robustness to traffic-density and CAV-penetration shifts and transfer to unseen merge-only and diverge-only route-demand scenarios within the same corridor geometry.