Upgrading Multi-Agent Pathfinding for the Real World

This talk aims to invite you to the forefront of MAPF research directly 🚀 This is a re-recording of my invited talk at EurMAPF-25, held in Bologna, Italy. Abstract: Multi-agent pathfinding (MAPF) has been actively studied since the 2010s. While early work focused on developing efficient optimal algorithms, recent research has increasingly shifted toward more practical directions — including suboptimal yet highly scalable MAPF methods aimed at real-time multi-robot fleet control in logistics. So, where do we stand today? This talk will explore these modern foundational algorithms and their extensions, which are almost ready to tackle real-world challenges. The talk also considers some promising, albeit premature, techniques for pushing MAPF research into new frontiers. Contents: 1:36 Motivation to MAPF 7:19 Feedback control loop 11:24 Quick summary of existing algorithms: A*, (E)CBS, PP, LNS2 19:50 Ultra-lightweight methods: PIBT & LaCAM 31:08 Improving solution quality, LNS, congestion mitigation 38:58 Learning and search for MAPF: GNN, LaGAT, HMAGAT 52:50 Multi-robot planning and control: LF, db-LaCAM, D4orm 1:02:46 Wrap-up Slides: https://speakerdeck.com/kei18/eurmapf... Disclaimer: Unfortunately, my cursor movements are not being recorded, so some parts may appear slightly awkward. Also, there are sound timing errors for the robot movies caused by the PowerPoint export function 😞 Feel free to contact me if you have any questions 📨