<p>Traditional Q-learning algorithms often exhibit slow convergence, inefficient exploration, and excessive turning points when applied to global path planning in grid-based environments. To address these limitations, this paper proposes a fuzzy logic-guided Q-learning algorithm (FL-GQL), which integrates heuristic Q-table initialization, hybrid reward shaping, and fuzzy-guided dynamic exploration into a unified learning framework. Unlike existing studies that mainly improve only one component of Q-learning, the proposed method coordinates initialization, exploration guidance, and reward feedback within a single learning loop, thereby improving convergence efficiency, path quality, and planning reliability in a collaborative manner. Comparative experiments are conducted across six environments of varying scale and obstacle structure, including both randomly generated maps and manually designed structured environments featuring corridor constraints, concave obstacles, and clustered distributions. FL-GQL demonstrates consistently better average performance than five improved Q-learning algorithms and a DQN baseline, achieving an average success rate of 66.43% across all environments versus 37.16% for BQL and 43.87% for DQN, while generating paths with fewer turns and requiring fewer convergence episodes. Ablation experiments further show that the performance gains arise from the coordinated contribution of the three components, as removing any single module leads to degradation in convergence speed, path smoothness, or planning reliability. Robustness and sensitivity analyses indicate that FL-GQL maintains stable performance across varied parameter settings, with favorable results obtained under a relatively high learning rate and exploration rate together with a moderate discount factor. These results suggest that FL-GQL provides an effective solution for global path planning of autonomous mobile robots in structured grid environments.</p>