GridMind: Multi-Agent Navigation Using Reinforcement Learning in a 3D Environment

Multi-agent navigation is a fundamental problem in robotics, autonomous vehicles, warehouse automation, traffic management, and other systems in which several autonomous entities must reach individual destinations while sharing limited space. Classical Multi-Agent Path Finding (MAPF) focuses on producing collision-free paths, whereas Multi-Agent Reinforcement Learning (MARL) learns policies through interaction and can adapt to changing agent behavior. This literature paper reviews the foundations of reinforcement learning, Q-learning, Deep Q-Networks (DQN), multi-agent reinforcement learning, coordination, collision avoidance, centralized training with decentralized execution (CTDE), and learning-based multiagent pathfinding. The literature identifies non-stationarity, partial observability, joint action complexity, reward design, communication overhead, scalability, and evaluation consistency as recurring challenges. Based on these findings, GridMind is positioned as a controlled research environment in which multiple agents learn navigation policies in a discrete grid while the same environment is rendered as an interactive 3D simula-tion. A staged approach—random baseline, Q-learning, DQN, independent multi-agent learning, and coordinated learning—provides an understandable and experimentally defensible pathway for studying learned coordination.