When 1,000 robots share a warehouse floor, planning their routes together falls apart past 200 agents. A Zaragoza team built something different: decouple planning from execution, add FIFO queues, and waiting time drops 70 percent. The catch: it only works if the warehouse geometry cooperates.
image from grok
Researchers at the University of Zaragoza developed GCP+DLC, a hybrid multi-agent pathfinding architecture that separates sequential A* route planning from decentralized FIFO execution, achieving near-linear scalability to 1,000 agents with 70% less waiting time compared to baseline approaches. The key insight is that centralized planning and decentralized execution don't need the same information—planning benefits from global priority awareness while execution needs only local FIFO queuing—enabling near-centralized solution quality at near-decentralized computational cost. However, the approach guarantees 100% success only under a geometric feasibility assumption that may not hold in continuously-arriving dense fulfillment scenarios.
When 1,000 robots share the same floor, the obvious solution—plan everyone's route together, all at once—breaks down around 200 agents. A team at the University of Zaragoza has published a paper showing a different architectural bet works better: split the problem in two, hand each half to a different algorithm, and let a simple FIFO queue handle the rest.
Their system, called GCP+DLC (Geometric Conflict Preemption plus Decentralized Local Controller), plans routes sequentially using A* search—one agent at a time, each aware of where higher-priority agents will be. At runtime, a lightweight decentralized protocol handles execution through per-vertex first-in-first-out queues, inserting wait actions only when geometrically unavoidable. On standard benchmark maps, the approach scales near-linearly to 1,000 agents with waiting time cut roughly 70 percent, even as spatial detours increase about 16 percent, according to results posted to arXiv on March 11, 2026.
The paper, a six-page WODES workshop submission by Fernando Salanova, Cristian Mahulea, and Eduardo Montijano, is careful about what it claims. The 100 percent success rate holds only on instances satisfying the geometric feasibility assumption—roughly, that there exists a sequential execution order where each agent can move while the rest wait. In a sparse, well-designed warehouse this is reasonable. In a dense fulfillment center where new orders arrive continuously, it may not be. The paper proves completeness for the feasible case; it does not claim the assumption always holds in practice.
Multi-agent path finding is the core scheduling problem underneath warehouse robotics, airport tarmac logistics, and factory floor automation. The existing landscape splits roughly into two camps: centralized methods like ECBS that plan jointly for the whole team but struggle to scale beyond a few hundred agents, and decentralized methods like PIBT that execute fast in warehouses but sacrifice solution quality and make no guarantees about success. GCP+DLC's architectural move is to give each layer exactly the information it needs—geometric feasibility for planning, FIFO ordering at execution time—and get near-centralized solution quality with near-decentralized scalability in return.
The numbers bear it out. At 1,000 agents on the standard room benchmark, GCP+DLC reduces sum-of-costs by 23.1 percent compared to running the same algorithm without geometric conflict preemption. The waiting proportion of total path cost drops from 68.4 percent to 52.4 percent—agents spend less time queued and more time moving toward their goals. The priority heuristic also turns out to be a first-order design choice rather than a tuning detail: CL (Conflicting Path Last), which defers routes that conflict with higher-priority paths, outperforms LPF by 18 percent in sum-of-costs at the 1,000-agent scale.
The paper benchmarks against ECBS, Cooperative A*, and PIBT. PIBT—Priority Inheritance Backtracking—is worth dwelling on because it is what Symbotic, a warehouse automation company, was already using in automated warehouses. Symbotic's sponsorship of the WoMAPF workshop at AAAI 2025 signals the industrial MAPF community is not waiting for academic consensus before deploying. The Zaragoza contribution occupies a well-defined academic niche: proving that explicit geometric preemption with decentralized FIFO execution achieves bounded-suboptimal results at scale, and does so cleanly.
Cristian Mahulea, a professor at the University of Zaragoza with roughly 2,390 citations and co-chair of the IEEE Robotics and Automation Society's technical committee on automation in logistics, is the senior author. Fernando Salanova, a master's student, published a related MAPF paper in October 2025—two first-author papers in six months. Eduardo Montijano rounds out the team, which is funded by ONR Global and the Spanish Ministry of Science.
The honest gap is real: no code published, all experiments run on a single Intel Core i7 workstation, no distributed execution, no physical hardware. A concurrent paper from MIT and Harvard posted to arXiv the same week takes a reinforcement learning approach to lifelong MAPF in warehouse settings. Both threads, filed within days of each other, suggest the academic race to solve fleet-scale coordination is accelerating faster than the press coverage tracking it.
For teams evaluating MAPF stacks for fleet sizes between 200 and 1,000 agents, GCP+DLC is a credible architectural signal: the decoupled planning/execution split is now a proven design pattern, not just a theoretical proposal. The practical takeaway is that priority ordering deserves the same engineering attention as path cost heuristics. CL-style ordering that penalizes conflicting paths outperforms naive priority schemes by a meaningful margin at scale. Whether that advantage survives contact with non-stationary arrivals and real robot kinematics is the next open question—and the one that matters most for anyone actually running a warehouse.
Story entered the newsroom
Research completed — 0 sources registered. GCP+DLC decouples geometric planning from temporal execution in MAPF. GCP (A* with cost inflation) reduces downstream congestion; DLC (FIFO queues) ha
Draft (758 words)
Reporter revised draft based on fact-check feedback (1314 words)
Reporter revised draft based on fact-check feedback (752 words)
Reporter revised draft based on fact-check feedback (739 words)
Approved for publication
Headline selected: Your Warehouse Doesn't Have to Stop at 200 Robots
Published (737 words)
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Agentics · 2h 51m ago · 4 min read
Agentics · 4h 29m ago · 5 min read
