The paper addresses multi-agent pathfinding on a four-connected grid, where agents must move from distinct starts to distinct goals without vertex collisions or opposite-direction edge collisions. At every timestep, an agent either waits or moves to an adjacent vertex, and a neural policy maps local observations and interactions to a distribution over five actions. The objective is to produce collision-free paths, with makespan and Sum-of-Costs used as standard cost notions. Unlike the dominant monolithic-policy setting, in which one policy is expected to handle every environment, the paper asks whether several map-specialized policies can be combined adaptively during execution.
The primary contribution is a map-distribution MoE framework that prepares experts for Maze, Warehouse, Room, and Random distributions and learns a router to fuse their outputs. The framework is instantiated with both MAGAT, a commonly used graph-neural-network policy, and ReCAT, a newly designed lightweight policy based on recursive action-token refinement. The paper also provides a reusable expert-training recipe that starts from a pretrained policy, applies supervised fine-tuning for each target distribution, and then applies reinforcement-learning fine-tuning. Its conclusion is deliberately qualified: MoE can improve a pretrained policy and compete with established methods, but it does not automatically beat a well-trained monolithic policy under a matched parameter budget.
Expert preparation begins with a pretrained base policy and performs supervised fine-tuning on demonstrations from each target map distribution. The experts are then refined with Group Relative Policy Optimisation, producing policies intended to be complementary rather than identical. During MoE execution, the router receives the experts’ action logits for an agent and assigns situation-dependent weights before producing the combined action distribution. The main training route first constructs the experts and then trains the router with the experts frozen; variants also jointly update experts and router and optionally apply reinforcement-learning fine-tuning afterward. ReCAT combines a spatial encoder that preserves field-of-view resolution, interaction modeling through recursive refinement of action tokens, and two projections from initial and final tokens to action logits.
The evaluation uses the POGEMA toolkit and runs on a single NVIDIA H200 GPU for both MAGAT and ReCAT. The pretraining data mix contains 20% Random and 80% Maze instances, with widths and heights from 17 to 21 cells and 16, 32, or 48 agents; expert fine-tuning covers four target distributions, while evaluation additionally includes Sortation, Stripes, and Empty to examine distribution shift. The reported measures are success rate and normalized Sum-of-Costs, with uncertainty shown for SoC, and the comparisons include pretrained policies, MoE policies, parameter-matched supervised-plus-reinforcement-learning monolithic policies, HMAGAT, and MAPF-GPT-DDG. The excerpt identifies several training procedures and dataset proportions, but it omits enough exact values that a complete numerical reconstruction of every experiment is not established by the supplied evidence.
Across both policy architectures, the results indicate that map-distribution experts can make a useful MoE: the approach generally improves over the pretrained policy and can outperform existing leading methods. ReCAT-MoE is described as strong in both in-distribution and out-of-distribution settings despite its relatively compact construction. Nevertheless, parameter-matched monolithic models remain stronger on average, so the results do not support treating MoE as a universal replacement for a single policy. The training ablation favors supervised router learning with frozen experts, followed by reinforcement-learning fine-tuning of both components, although the best MoE variant still falls behind the monolithic model in important cases, especially the out-of-distribution Stripes setting. The authors identify explicit dependence on predefined map distributions, the cost of evaluating all experts, the possibility of implicit specialization through joint learning, and automatic map-distribution generation as central limitations and future directions.
MAPF is a standard testbed for large-scale multi-agent learning because agents must coordinate densely while pursuing individual goals. It is also relevant to collective automation applications such as warehouse logistics. Neural MAPF policies have been trained with reinforcement learning and imitation or supervised learning, and some prior systems combine learned policies with search. Mixture-of-Experts is a broader machine-learning paradigm in which a router combines specialized subnetworks, with successful uses in language models and related ideas in robotics; this paper transfers that meta-level idea to map-dependent neural MAPF.
The formal setting consists of agents on a four-connected grid, each with a distinct start vertex and goal vertex. An agent can wait or move to an adjacent vertex, but no two agents may occupy the same vertex at the same time and no pair may traverse the same edge in opposite directions during one timestep. A solution is a sequence of collision-free configurations that brings every agent to its goal, after which it remains there. Makespan measures the overall plan length, while Sum-of-Costs sums the travel times of all agents, and the learned policy must make local decisions despite global coordination requirements.
The expert pipeline uses distribution-specific supervised fine-tuning followed by reinforcement-learning fine-tuning from a shared pretrained model. The router fuses expert action logits per agent and per local situation, rather than selecting one expert using a fixed rule. Training masks 80% of wait-action labels for agents that have already reached their goals, because such labels can weaken the learning signal, while those agents remain in the communication graph. ReCAT is introduced as a compact architecture that preserves spatial correspondence and recursively refines action tokens, whereas MAGAT provides a distinct GNN-based test case for the same MoE idea.
The experiments compare pretrained MAGAT and ReCAT, their map-distribution MoE versions, and monolithic versions trained with supervised and reinforcement-learning fine-tuning at approximately matched parameter scales. HMAGAT and MAPF-GPT-DDG are also included as established comparison methods. The evaluated distributions are Maze, Warehouse, Room, Random, Sortation, Stripes, and Empty, with distinctions between in-distribution and out-of-distribution evaluation. The study examines success rate, normalized SoC, equal-weight expert fusion, and alternative choices for freezing or jointly training experts and router; the excerpt does not establish every run count or exact hyperparameter.
The MoE variants generally improve on their pretrained starting policies and in some settings outperform the leading comparison methods. ReCAT-MoE is reported to achieve strong performance both on familiar distributions and under distribution shift. Among the routing-training variants, supervised router training with frozen experts followed by reinforcement-learning updates to both router and experts gives the largest gains in the reported comparison. However, a parameter-matched monolithic model remains better on average, with the out-of-distribution Stripes case contributing notably to the MoE gap; the supplied excerpt does not provide enough complete numerical values to state exact success rates or SoC differences.
Map-distribution MoE is a credible meta-level approach for neural MAPF, especially when experts can capture useful differences among environment distributions. The study shows that the idea transfers across two substantially different policy architectures and can improve a pretrained policy without relying on a single fixed expert choice. Its evidence also demonstrates the limits of the approach: under comparable parameter budgets, a well-trained monolithic model is still stronger. The practical value of MoE therefore depends on the balance among specialization, generalization under distribution shift, parameter allocation, and inference cost.
The current design requires explicitly constructing experts for predefined map distributions, so it does not yet discover all useful specializations automatically. Because each expert may perform its own interaction modeling, the usual implementation evaluates all experts for every agent, creating more inference overhead than a sparse MoE that activates only a subset. The average disadvantage relative to a matched monolithic policy suggests that the benefits of specialization and routing do not always compensate for duplicated capacity and computation. The excerpt also does not establish complete numerical results, broad real-world robustness, or statistical conclusions beyond the reported experimental comparisons.
The paper broadens neural MAPF research from designing an individual policy architecture to designing a collection of specialized policies and a mechanism for combining them. Its learned, situation-dependent fusion is distinct from fixed policy switching, simple algorithm selection, or additive combination of unrelated policy outputs. The study is important because it reports both positive and negative evidence: specialization can improve pretrained systems and help under distribution shift, yet a single carefully trained model may use the same capacity more effectively. This framing gives MAPF researchers a concrete basis for comparing specialization, generalization, and inference overhead rather than assuming that the success of MoE in other domains transfers automatically.
The paper is suitable for researchers working on neural MAPF, multi-agent reinforcement learning, imitation learning, map-distribution shift, and policy ensembles. It is especially relevant to readers interested in comparing GNN-based policies such as MAGAT with lightweight recursive-refinement policies such as ReCAT. Researchers considering an MoE implementation will find the expert-training and router-training choices useful, including the distinction between frozen-expert supervised routing and later joint reinforcement-learning fine-tuning. Readers who require exact benchmark numbers, complete hyperparameters, or extensive deployment evidence should consult the full paper because those details are not fully established by the supplied excerpt.