New PUSH planner coordinates 10k agents in under a second

Lifelong Multi-Agent Path Finding (LMAPF) is critical for warehouse robotics, autonomous vehicles, and other large-scale fleets. Existing approaches face a stark trade-off: reactive rule-based methods like PIBT and Enhanced PIBT (EPIBT) scale gracefully to thousands of agents but only look one step ahead, causing short-sighted decisions in long-horizon tasks. Windowed planners like RHCR offer foresight but incur heavy computational overhead that hinders scalability. The new PUSH (Path Updates over Staggered Horizons) algorithm, proposed by researchers Vaibhav Sanjay and Jiaoyang Li in arXiv:2608.06702, sidesteps this by planning windowed paths for only a subset of agents at each timestep—dramatically reducing complexity while preserving multi-step reasoning.

PUSH draws on the strengths of three prior methods: it uses TP's staggered subset planning to cut compute, RHCR's windowed path plans for long-horizon foresight, and EPIBT's priority inheritance, backtracking, and anytime improvements to maintain throughput in dense traffic. Crucially, unlike TP, PUSH does not rely on restrictive map assumptions, making it applicable to general environments. In empirical tests across two realistic scenarios, PUSH matched EPIBT's massive scalability—handling 10,000 agents—while achieving significantly higher system throughput than all baselines. This combination of scale and foresight could unlock more efficient real-time coordination for robotic fleets in warehouses, logistics, and other domains where long-horizon planning under tight real-time constraints is essential.