History-Aware C-FAME: Context-Aligned Frontier Assignment for Multi-Robot Exploration

Effective coordination in multi-robot exploration demands decision-making that dynamically adapts to historical context and team behaviour, yet existing methods suffer from fragmented reasoning that neglects temporal continuity or spatial coordination. We present C-FAME (Context-Aligned Frontier Assignment in Multi-Robot Exploration), a hierarchical framework that reformulates exploration as a history-aware spatio-temporal alignment problem. At its core, C-FAME integrates three innovations: 1) a Bi-LSTM encoder that preserves temporal continuity by modelling agents' state histories, 2) an attention-based mechanism that evaluates frontiers through joint spatial coordination and historical alignment, and 3) a context-aligned regularisation term that enforces spatially dispersed target assignment. The framework operates hierarchically: a high-level policy performs context-aligned frontier assignment over the shared frontier set, while low-level agents execute goal-conditioned navigation using decentralised SAC controllers. Extensive simulations demonstrate that C-FAME achieves up to 37.5% faster exploration and up to 63% fewer collisions compared to classical and learning-based baselines. Ablation studies confirm the necessity of both temporal encoding and spatial attention, while trajectory visualisations reveal coherent, collision-free paths unattainable by reactive or partitioned strategies. By treating history-aware context as a structured decision signal rather than passive input, C-FAME improves temporal consistency and spatial coordination in hierarchical multi-robot exploration.