Arita and Okumura's LLLG achieves state-of-the-art lifelong multi-agent pathfinding

Arita and Okumura introduce LLLG (Lifelong LaCAM with Local Guidance), a new algorithm for lifelong multi-agent pathfinding (LMAPF) where tasks arrive continuously. Building on the scalable, suboptimal LaCAM solver, they add local guidance — spatiotemporal hints around each agent — that was previously proven effective in one-shot MAPF. LLLG employs a receding-horizon windowed planning framework and warm-starts guidance from the prior time step’s solution, allowing it to maintain high throughput even in compact, dense environments. The method scales effectively with tight time budgets and surpasses existing planners, marking a new frontier for real-time LMAPF.

The paper, accepted to SoCS 2026, includes 10 pages and 11 figures demonstrating performance gains. By adapting local guidance to the lifelong setting, the researchers show that agents can reduce waiting and congestion without sacrificing scalability. This has practical implications for warehouse robotics, autonomous vehicle coordination, and any domain requiring persistent, collision-free navigation in confined spaces. LLLG offers a clear path toward deploying robust multi-agent systems in dynamic, real-world scenarios where planning must happen continuously under time constraints.