A Review of Indoor Navigation Methods for Autonomous Mobile Robots: From Path Planning to Large Model-Based Decision-Making

Indoor navigation is fundamental to autonomous mobile robots (AMRs), yet the relationships between explicit planning, learned policies, and large model-based decision-making remain difficult to assess across heterogeneous experiments. This review establishes a unified taxonomy of path planning-based, learning-based, and large model-based navigation according to their dominant decision-making mechanisms, while recognizing hybrid systems. We summarize the foundations of simultaneous localization and mapping (SLAM), deep and reinforcement learning, and large models, then examine representative methods through their navigation tasks, decision roles, and validation settings. To move beyond qualitative comparisons, we synthesize reported quantitative evidence on navigation success, computational cost, online latency, dynamic adaptability, physical validation, and safety, explicitly distinguishing task protocols and measurement boundaries. We also compare common simulation platforms, datasets, and physical-robot platforms, examining how sensor noise, moving obstacles, and execution uncertainty affect sim-to-real transfer. The analysis shows that explicit planning provides interpretable, constraint-aware motion generation, learned policies support experience-driven adaptation, and large models enable semantic instruction understanding and high-level task reasoning. However, no paradigm alone resolves the challenges of generalization, embodied grounding, inference delay, and reliable physical deployment. Rather than ranking methods across incomparable evaluations, we identify complementary capabilities and the value of hybrid architectures that couple semantic reasoning with constraint-checked low-level execution and feedback. Future work should address lifelong adaptation, embodied memory and world models, efficient inference, and measurable safety and failure recovery.