Multi-Agent Warehouse Pathfinding

• MAWPF is a warehouse-specific extension of MAPF that incorporates operational constraints like charging stations, slow corridors, and continuous task streams.
• It combines discrete grid search with realistic AGV dynamics and resource modeling to ensure collision-free navigation and efficient task completion.
• Algorithmic paradigms range from ASP-based flexible planning to scalable online methods, highlighting trade-offs between optimality, robustness, and throughput.

Searching arXiv for recent and directly relevant MAWPF papers and the cited works in the provided corpus. Multi-Agent Warehouse Pathfinding (MAWPF) denotes warehouse-grounded variants of multi-agent path finding in which multiple robots must move through a shared warehouse environment without collision while satisfying operational constraints that are weakly represented or absent in classical MAPF, such as charging, speed-restricted areas, online task streams, interdependent pickup-and-delivery structure, and realistic AGV motion constraints (Bogatarkan, 2021, Nagai et al., 15 May 2026, Lehoux-Lebacque et al., 2024). In current research practice, the term spans both one-shot formulations that preserve discrete combinatorial search and lifelong formulations in which robots continuously receive new destinations and system throughput becomes a primary objective (Okumura et al., 2019, Zhang et al., 2023, Zheng et al., 25 Mar 2026).

1. Formal scope and relation to classical MAPF

Classical MAPF is commonly specified either as a tuple $\langle G,s,t\rangle$ with an undirected graph $G=(V,E)$, source map $s$, and target map $t$, or as $(n,\mathcal{G}=(V,E),S,G)$, with synchronous discrete timesteps and per-agent actions consisting of moving to an adjacent vertex or waiting (Stern et al., 2019, Andreychuk et al., 2024). The standard conflict model includes at least vertex conflicts and edge-swap conflicts, and standard optimization objectives are sum of costs, $SoC(Pl)=\sum_i cost(pl_i)$, and makespan, $MS(Pl)=\max_i cost(pl_i)$ (Andreychuk et al., 2024).

MAWPF inherits this core abstraction but changes its interpretation. In warehouse settings, the graph represents aisle cells, intersections, stations, shelf-adjacent endpoints, or a routing graph derived from the physical workspace rather than an abstract obstacle field. The stay-at-target versus disappear-at-target distinction becomes operationally significant because a robot at a workstation, charger, or shelf face may continue to block traffic (Stern et al., 2019, Andreychuk et al., 2024). Lifelong warehouse formulations also replace terminal completion by continual reassignment: once an agent reaches its current goal, it receives a new one, and system quality is naturally measured by throughput rather than only one-shot path cost (Zhang et al., 2023, Zheng et al., 25 Mar 2026).

This broader scope is reflected in warehouse-level metrics. In layout-centric lifelong studies, throughput is defined as

$$\text{throughput} = \text{average number of finished tasks per timestep},$$

with a task completed when an agent reaches its currently assigned goal (Zhang et al., 2023). This marks an important conceptual shift: MAWPF is not merely “MAPF on a warehouse map,” but pathfinding embedded in a continuously operating logistics system.

2. Warehouse-specific modeling extensions

A central feature of MAWPF is the enrichment of the state and constraint model. In ASP-based multi-modal MAPF with resources, warehouse instances include blocked cells for shelf units, charging stations, slow-travel corridors, waypoints, battery levels, and dynamic replanning under agent and obstacle changes (Bogatarkan, 2021). In the illustrative warehouse example of that framework, charging stations appear at cells 24 and 27, a slow corridor covers cells 3–8, movement through that corridor takes 2 timesteps per neighbor transition, and battery decreases at each time step unless a robot is at a charging station (Bogatarkan, 2021). This turns pathfinding into a joint routing-and-resource-feasibility problem.

A different line of work formalizes MAWPF specifically for differential-drive AGVs. There the per-agent state is

$$\mathcal{Q}[i]=(x_i,y_i,\theta_i,v_i),\quad v_i\in \{0,1,\ldots,V_{max}\},$$

and the defining constraints are: straight motion and in-place rotation only; multi-step rotation costs; acceleration and deceleration; and prohibition of follower collisions to prevent rear-end crashes (Nagai et al., 15 May 2026). In that model, rotation by $90^\circ$ may require $G=(V,E)$0 timesteps, speed is part of the state, and an agent moving with $G=(V,E)$1 occupies all swept grid vertices during that timestep (Nagai et al., 15 May 2026). The resulting state graph is directed rather than effectively reversible, which materially changes the search problem.

Warehouse task structure can further expand the model beyond fixed-goal MAPF. In online order-delivery settings with shelves and workstations, robots are represented by continuous-time configurations $G=(V,E)$2, where $G=(V,E)$3 denotes load state, and tasks are sequences of via points requiring specific arrival orientations, zero speed for pickup or drop-off, and workstation exclusivity across all tasks of the same order (Lehoux-Lebacque et al., 2024). This makes MAWPF a combination of trajectory planning, task sequencing, and resource coordination rather than only collision-free waypoint routing.

These formulations suggest that MAWPF is best regarded as a family of warehouse-specific MAPF models. Some remain close to discrete grid search; others incorporate batteries, workstations, directionality, or load-dependent timing. The common element is that warehouse realism is introduced through explicit operational constraints rather than by treating the warehouse as a generic obstacle map.

3. Algorithmic paradigms

The algorithmic landscape of MAWPF is correspondingly heterogeneous. One line prioritizes modeling flexibility. The ASP-based framework for flexible and explainable MAPF uses declarative encodings to support multi-objective optimization, waypoints, resource constraints, multi-modal transportation, and dynamic environment changes, with Clingo as the backend solver (Bogatarkan, 2021). Its strength lies less in specialized graph-search efficiency than in elaboration tolerance: charging, slow regions, and online modifications can be encoded without redesigning the planner (Bogatarkan, 2021).

A second line prioritizes scalable online coordination. Priority Inheritance with Backtracking (PIBT) is explicitly motivated by automated warehouses with hundreds or more robots and solves iterative MAPF by planning one timestep at a time with adaptive priorities and local backtracking (Okumura et al., 2019). Under the condition that all pairs of adjacent nodes belong to a simple cycle, PIBT guarantees that every agent reaches its destination within finite time, and in MAPD-style warehouse experiments it outperformed Token Passing in both runtime and solution quality while avoiding the requirement for large sets of non-task endpoints (Okumura et al., 2019). This makes PIBT representative of methods that sacrifice global optimality for predictable online behavior.

A third line adapts classical optimal or near-optimal MAPF search to warehouse structure. BPM* extends M* by introducing bypass into subdimensional expansion: when a collision occurs, the algorithm first tries to reroute an agent along another path of equal individual optimal cost before enlarging the collision set and coupling the agents into higher-dimensional search (2207.14657). This is particularly relevant in structured aisle systems, where multiple equally short routes around shelving blocks are common. The paper proves BPM* complete and optimal under its MAPF assumptions (2207.14657).

Learning-based methods occupy several distinct positions. MAPF-GPT is a transformer-based imitation-learning policy trained on one billion observation-action pairs from expert MAPF solutions and shown to generalize zero-shot to warehouse maps, including a $G=(V,E)$4 warehouse benchmark with up to 192 agents (Andreychuk et al., 2024). However, in lifelong MAPF on warehouse maps, the reported throughput of MAPF-GPT-6M remained below RHCR and Follower, so its strongest evidence concerns decentralized one-shot coordination rather than sustained warehouse circulation (Andreychuk et al., 2024). RDE instead augments existing RL MAPF policies with a Distance Heat Map policy and a random Escape policy, improving success rates on dense warehouse-like $G=(V,E)$5 structured maps by up to 15% when 70 agents are present on dense layouts (Gao et al., 2023).

The AGV-specific MAWPF benchmark shows that classical lightweight solvers can be adapted, but not trivially. PP, LNS2, PIBT, and LaCAM were all lifted to the $G=(V,E)$6 warehouse state model; PP and LNS2 produced better solution costs but struggled as the number of agents grew, while PIBT-based approaches—especially rolling-horizon LaCAM plus multi-step PIBT—showed preferable scalability at increased cost (Nagai et al., 15 May 2026). A key result is that macro-successor commitment failed badly under the directed MAWPF state graph, whereas rolling-horizon replanning succeeded consistently (Nagai et al., 15 May 2026).

Another alternative replaces explicit space-time search by online control. Multiagent rollout with reshuffling models warehouse routing as an infinite-horizon discounted control problem, penalizes collisions via a large constant $G=(V,E)$7, rewards target achievement with $G=(V,E)$8, and chooses joint actions by sequential rollout plus random order reshuffling (Emanuelsson et al., 2022). In a $G=(V,E)$9 warehouse with 1183 goods and 100–200 robots, it achieved 100% success across tested robot counts, while Cooperative A* success declined from 89% at 100 robots to 58% at 200 robots (Emanuelsson et al., 2022). This suggests that online simulation-based control can be competitive in lifelong settings even when it departs from strict hard-constraint MAPF formulations.

4. Lifelong operation, task coupling, and throughput optimization

In warehouse systems, pathfinding is tightly coupled to persistent task streams and to the geometry of the facility. A decisive empirical result is that layout itself can dominate solver quality. In a lifelong warehouse study using RHCR and DPP, optimized layouts reduced traffic congestion, improved throughput, and in a large workstation setup increased sustainable robot count from fewer than 150 under a human-designed layout to more than 300 under optimized layouts (Zhang et al., 2023). In the same study, human-designed layouts had 0% success rate under RHCR in all four reported benchmark settings, showing that better MAPF alone may not overcome structurally poor warehouse topology (Zhang et al., 2023).

This system-level perspective extends beyond map design. WareRover argues that order scheduling and MAPF must be evaluated in a step-level closed loop rather than as isolated subproblems, and integrates dynamic order streams, structured warehouse layouts, continuous-time collision checking, turning constraints, velocity limits, and AGV failure recovery into one RMFS simulator (Xu et al., 15 Feb 2026). In homogeneous settings, all tested combinations had near-perfect success, but throughput differed strongly between random dispatching and the stronger scheduling strategy TA; in heterogeneous and fault-tolerant settings, algorithm rankings changed substantially, with CBS excelling in throughput under heterogeneity and A* benefiting from lower replanning cost under failures (Xu et al., 15 Feb 2026). This shows that MAWPF performance is shaped jointly by routing, scheduling, heterogeneity, and non-nominal events.

Task coupling can also arise from storage policy. CAL-MAPF augments lifelong MAPF with cache cells near unloading ports, a read/write lock mechanism, and a task assigner that decides whether a demand should be served from shelf or cache (Tang et al., 2024). On a $s$0 warehouse with 1600 shelf grids, up to 80 cache grids, 4 unloading ports, and task queues of length 1000, the method often outperformed the no-cache baseline when demand exhibited locality, but its gains diminished at high robot density because congestion near cache and port regions reduced cache hit rate (Tang et al., 2024). This identifies a specifically warehouse-level design lever: where tasks are sourced from can matter as much as how paths are found.

A more explicit coupling between assignment and routing appears in the intelligent-warehouse TAPF work based on MADDPG, which jointly learns target assignment and path planning under simplified continuous dynamics (Liu et al., 2024). Because collision avoidance there is handled by reward penalties rather than exact MAPF constraints, and because tasks are target points rather than full warehouse pickup-delivery workflows, it is better viewed as a warehouse-inspired continuous TAPF formulation than as a canonical MAWPF solver (Liu et al., 2024). Its inclusion nevertheless underscores a broader trend: warehouse routing increasingly overlaps with assignment, staging, and execution policy.

5. Execution robustness, monitoring, and explainability

MAWPF research increasingly distinguishes planning from execution. Robust execution with the Action Dependency Graph constructs a DAG

$s$1

over actions, with Type 1 edges for intra-agent precedence and Type 2 edges for inter-agent precedence at shared resources (Zahrádka et al., 12 Sep 2025). Estimated action start and completion times are propagated over this graph, then updated online as actual action completion events arrive. The architecture defines edge slack $s$2, action slack $s$3, and a fleet-level slack statistic $s$4, and triggers replanning when

$s$5

using $s$6 ms in the reported experiments (Zahrádka et al., 12 Sep 2025). The key warehouse implication is that replanning is justified not by any delay in isolation, but by delay that propagates through precedence constraints and raises expected execution duration enough to outweigh replanning cost.

Robustness can also be handled by hybrid navigation layers. CHS combines D* Lite with multi-agent reinforcement learning, switching to local RL when more than four nearby agents are present or when loop behavior such as $s$7 is detected (Liu et al., 28 Mar 2025). In partially observable $s$8 grids with 30% obstacles, CHS reached $s$9 and $t$0 for 128 agents, compared with $t$1 for DHC and $t$2 for SCRIMP under the same setting (Liu et al., 28 Mar 2025). Although not warehouse-specific in map design, this work is directly relevant to disturbed warehouse execution where local blockage discovery and dense interactions are more important than static optimality.

Explainability is another warehouse requirement. The ASP-based explainable MAPF framework answers structured questions about feasibility, infeasibility, optimality, nonoptimality, and observed waiting behavior by solving modified planning instances and generating counterfactual justifications (Bogatarkan, 2021). Example outputs include statements such as “There is no solution because Robot~1 and Robot~2 collide at Cell~8 at time step~3” and “Actually, Robot~2 does not have to wait at Cell~8 from time step 0 to 2. Here is an alternative optimal plan” (Bogatarkan, 2021). In warehouse operation, such explanations support debugging chokepoints, validating safety rules, and distinguishing necessary waiting from avoidable inefficiency.

The realistic online warehouse work with Interleaved Prioritized Planning also belongs to this execution-oriented layer. By assigning each robot a designated waiting place and reserving waiting paths, IPP achieves completeness under its assumptions, while VP* plans dynamics-compliant trajectories through via points with moving obstacles using a reservation table (Lehoux-Lebacque et al., 2024). The real-warehouse experiments show that action-duration uncertainty, especially in pickup and drop-off, materially affects trajectory fidelity, and that time margins in reservations improve robustness at the cost of runtime (Lehoux-Lebacque et al., 2024).

6. Empirical picture, misconceptions, and open problems

A recurring misconception is that MAWPF is simply classical MAPF deployed on a warehouse floor. The literature instead shows three distinct broadening moves: richer motion and resource models, lifelong throughput-centric replanning, and system-level coupling to layout, scheduling, or execution monitoring (Bogatarkan, 2021, Zhang et al., 2023, Xu et al., 15 Feb 2026). Another misconception is that better path planners alone determine warehouse performance. Layout optimization demonstrates that map topology can dominate solver behavior, with human-designed layouts leading to 0% RHCR success in all four studied setups and optimized layouts sometimes doubling sustainable robot population (Zhang et al., 2023).

The empirical record is also uneven across paradigms. Learned decentralized policies such as MAPF-GPT show strong zero-shot one-shot generalization to warehouse maps, but specialized lifelong methods still dominate on warehouse throughput benchmarks (Andreychuk et al., 2024). RL-guided Rolling Horizon Prioritized Planning closes part of this gap by learning dynamic priority orders rather than end-to-end paths and achieves the highest total throughput among the reported baselines in the main Amazon experiments and in dense Symbotic settings where congestion dominates (Zheng et al., 25 Mar 2026). This suggests that the most promising learning role in current MAWPF may be search augmentation rather than replacement.

Several operational concerns remain only partially addressed. One-way lanes, deadlines, priorities, lane-direction rules, payload-dependent speeds, charging congestion, integrated task allocation, and 3D geometry are repeatedly absent, implicit, or left for future work across flexible ASP, layout design, simulator, and AGV-adaptation papers (Bogatarkan, 2021, Zhang et al., 2023, Xu et al., 15 Feb 2026, Nagai et al., 15 May 2026). The AGV-specific MAWPF benchmark is a major step toward realism, but even its strongest approach scales to a few hundred agents rather than the thousands sometimes reported for classical grid MAPF, and width-one corridors remain particularly difficult (Nagai et al., 15 May 2026).

The field therefore occupies an intermediate stage. Warehouse research has moved beyond pure graph search on static grids: it now studies batteries, slow regions, workstations, cache zones, order streams, failures, explanation, and execution monitoring. Yet no single framework shown here simultaneously provides large-scale real-time performance, realistic AGV dynamics, integrated scheduling and assignment, robustness to stochastic execution, and comprehensive industrial constraint coverage. The present evidence supports a more precise characterization: MAWPF is a family of methods for collision-free multi-robot routing in warehouses, increasingly embedded in layout design, task coupling, and execution control, but still short of a unified warehouse execution theory.