A Survey of Learning-based Multi-agent Navigation

The Hong Kong University of Science and Technology (Guangzhou)

数据科学与分析学域

PhD Qualifying Examination

By Miss HUANG Yunjie

摘要

Multi-Agent Path Finding (MAPF) asks a team of robots to reach individual goals without collisions—an operation that underpins warehouse fulfilment, traffic coordination, and swarm robotics. Classical search planners such as A*, Conflict-Based Search, and their bounded-optimal extensions guarantee short paths, yet their running time and memory grow steeply with agent density, map size, and partial observability. As fleets scale into the hundreds, these costs hinder real-time deployment and motivate alternative approaches.

Reinforcement learning (RL) offers a complementary route: by training decentralised policies from experience, agents can react on-the-fly instead of exhaustively enumerating the joint state space. Recent work enriches RL with imitation, heuristic guidance, graph neural reasoning, and hierarchical abstractions, extending learned navigation from small grids to continuous, dynamics-aware domains. At the same time, GPU-based simulators such as Isaac Gym, ROS-integrated Gazebo, and lightweight benchmarks like VMAS and POGEMA provide the large-batch roll-outs needed to refine and test these methods under realistic sensing and actuation constraints.

This survey organises the field along four themes—classical planners, RL founda tions, learning-centric MAPF algorithms, and evaluation tool-chains—highlighting key advances, open safety and energy challenges, and the prospects for hybrid solutions that blend RL adaptability with search guarantees.

PQE Committee

Chair of Committee: Prof. LUO Qiong

Prime Supervisor: Prof. LI Lei

Co-Supervisor: Prof. LI Haoang

Examiner: Prof. LIANG Yuxuan