ASC-MAPPO: attention-based communication and safety-constrained multi-agent proximal policy optimization for cooperative path planning and dynamic obstacle avoidance of connected autonomous vehicles

Fleets of connected autonomous vehicles have to solve two problems at once: they must agree on who goes first, and each of them must stay clear of traffic that nobody controls. This paper presents ASC-MAPPO, an attention- and safety-constrained variant of multi-agent proximal policy optimization designed for exactly that coupling. The planning task is cast as a decentralized partially observable Markov decision process and solved with centralized training and communication-enabled distributed execution. Each vehicle encodes its own observation into a 128-dimensional embedding, broadcasts it inside a 50 m radius, and aggregates the messages it receives through a masked four-head attention module that weights neighbours by situational relevance rather than treating them alike. A shared-parameter actor emits continuous steering increments and longitudinal accelerations; agent-wise critics and a monotonic mixing network resolve credit assignment during training only. Safety is enforced on two levels. A differentiable Gaussian risk surrogate shapes the reward and supplies smooth gradients, and a projection layer confines every executed action to an admissible set defined by obstacle clearance, road boundaries, lane keeping and friction-limited vehicle dynamics. Collisions themselves are always decided by a binary contact test, never by the surrogate. In CARLA experiments covering highway merging, a four-way intersection and a curved rural road, with four cooperative vehicles and three obstacle densities, ASC-MAPPO reaches a 99.2% avoidance success rate in the dedicated avoidance suite, shortens the mean fleet arrival time in the high-density intersection by 18.2% relative to MAPPO (11.6% when averaged over all scenarios and densities) and attains a cooperative efficiency of 0.93. Ablations separate the contributions of communication, attention, value decomposition and the projection layer; transfer tests, a complexity analysis and latency measurements for up to twelve vehicles delimit the operating envelope of the method.