Geometric priors as decision-interface constraints for knowledge-guided multi-robot exploration

Knowledge-guided decision making for multi-robot exploration remains limited by the weak operational coupling between explicit geometric knowledge and continuous policy learning. In most hybrid frameworks, topological knowledge is used only as auxiliary guidance, which can lead to unstable mode switching, redundant exploration of structurally invalid regions, and reliance on computationally heavy perception modules. To address this problem, we propose PRIME, a prior-guided framework that embeds geometric priors as decision-interface constraints for continuous policy learning. By combining dynamic Voronoi partitioning with multidimensional utility evaluation, PRIME filters inadmissible exploration candidates before action generation and couples macro-level topological knowledge with micro-level adaptive control in a unified decision process. This prior-constrained design also enables a lightweight multilayer perceptron policy with compact multimodal state encoding, supporting responsive obstacle avoidance without convolutional feature extraction. Simulation experiments and controlled physical deployment tests show that PRIME converges in about 500 episodes with roughly 4 min of wall-clock training in the reported simulation setting. In unseen test environments, it reduces exploration time by 16.8% relative to the geometric baseline CURE and improves average map coverage from 94.0% to 97.6%. PRIME further achieves a 94.5% zero-shot transfer success rate while maintaining lightweight onboard inference. These results indicate that explicit geometric knowledge can be transformed into operative decision-interface constraints to improve exploration efficiency, transfer performance, and practical deployability in multi-robot exploration.