With　the　rapid　development　of　intelligent　technology,　Automated　Guided　Vehicles　(AGVs)　are　widely　used　in　large　outdoor　scenes　with　complex　terrain　to　serve　various　industrial　activities.　Since　the　collaboration　of　numerous　AGVs　has　become　a　major　trend,　efficient　Multi-Agent　Path　Finding　(MAPF)　algorithms　are　urgently　needed.　In　this　paper,　the　scenario　served　by　multiple　groups　of　AGVs　is　considered,　in　which　each　group　of　AGVs　has　the　same　destination.　The　path　planning　problem　of　each　AGV　is　modeled　as　a　Markov　Decision　Processes　(MDP).　The　state　is　constructed　based　on　the　local　environment　information　of　all　the　grids　around　the　AGV,　including　the　terrain,　occupation　status　and　estimated　cost　of　reaching　the　destination.　The　optimization　objective　is　defined　as　the　cumulative　reward　of　the　whole　path　containing　the　cost　of　passing　through　different　terrains,　the　penalties　for　staying,　the　penalties　for　collisions　and　conflicts,　and　the　rewards　for　reaching　the　destination.　As　all　the　AGVs　plan　their　paths　independently,　the　MAPF　problem　is　a　decentralized　MDP.　Therefore,　a　distributed　collaborative　path　planning　algorithm　based　on　Multi-Agent　Advantage　Actor-Critic　(MA-A2C)　is　designed.　Each　AGV　has　a　path　planning　agent,　and　the　AGVs　in　the　same　group　can　share　the　parameters　of　the　agent.　Numerical　simulation　results　validate　the　convergence　of　the　proposed　algorithm.　Compared　to　the　Waiting-stop　A∗　algorithm,　the　proposed　algorithm　demonstrates　better　performance　with　regard　to　the　cumulative　reward,　the　ratio　of　arrival　at　destination,　and　the　time　of　passage.　©　2024　IEEE.