CaST-MASAC: Integrating Causal Inference and Spatio-Temporal Attention for Multi-UAV Cooperative Task Planning

The efficient coordination of multi-Unmanned Aerial Vehicle (UAV) systems in the increasingly complex domain of aerial tasks is hampered by significant challenges, including partial observability, low sample efficiency, and difficulties in inter-agent coordination. To address these issues, this paper introduces a novel Causal Spatio-Temporal Multi-Agent Soft Actor–Critic (CaST-MASAC) framework. At its core, CaST-MASAC integrates two key innovations: (1) a spatio-temporal attention (STa) module that extracts features from historical observations to enable accurate target trajectory prediction and dynamic task assignment, thereby enhancing situational awareness and collaborative decision-making in highly dynamic and partially observable environments; and (2) a Causal Inference Experience Replay (CIER) mechanism that significantly improves sample efficiency and convergence speed by identifying and prioritizing experiences with a high causal impact on the task success. Evaluated in 4v4 and 2v2 multi-UAV aerial coordination simulation environments, CaST-MASAC demonstrates superior performance over state-of-the-art baselines such as MAPPO and QMIX in terms of task success rate, cumulative reward, and decision efficiency. Furthermore, extensive ablation studies validate the critical contributions of both the STa and CIER modules to the framework’s overall performance. Consequently, CaST-MASAC offers a novel and effective approach for developing robust and efficient multi-agent coordination strategies in complex dynamic environments.