The setting is decentralized MAPF on a four-connected grid, where multiple agents must move from distinct starts to distinct goals over discrete time without collisions. Each agent acts from local observations and communication rather than from a centralized planner with complete joint information. The evaluation uses success rate, sum of costs (SoC), and makespan: respectively, the fraction of fully solved instances, the total arrival cost over agents, and the arrival time of the last agent. The supplied excerpt does not fully establish every observation limit, communication bandwidth, or horizon detail.
The central contribution is to identify a factorization gap in decentralized policies: even if each agent learns its expert marginal action distribution correctly, independently sampling those marginals can combine pieces of different valid joint actions into an incompatible result. DMM addresses this by communicating and refining intermediate action intents before simultaneous commitment, rather than sampling final actions independently once. The paper also introduces MICPO, a critic-free group-relative reinforcement-learning method for fine-tuning the resulting multi-round decision process. At the solver level, DMM-MICPO-3M solves 1,598 of 1,600 MovingAI tasks, the highest coverage among the evaluated methods.
DMM initializes each agent with a random action intent and updates that intent over several local communication rounds. The communicated object is the intermediate refinement state, not only a completed action or a context from which an action has already been formed; agents can therefore condition their evolving choices on one another while those choices are still being constructed. Each agent commits to an action simultaneously after refinement, reducing the chance that independently plausible local choices will be recombined inconsistently. Training has two stages: imitation pretraining on expert MAPF solutions, followed by MICPO fine-tuning using shared task-level outcomes. MICPO keeps a reference policy anchored to the imitation-pretrained behavior at each round while allowing the communication pathway to remain trainable, so coordination can be reshaped without unrestricted drift.
The evaluation combines POGEMA, which tests learned policies as team size increases, MovingAI, which compares MAPF methods and their action-processing mechanisms across maps and populations, a large-scale execution study, and targeted corridor experiments. POGEMA compares DMM-3M, DMM-0.8M, their MICPO-tuned versions, and five learned baselines. The MovingAI protocol uses the maximum-agent instance from each scenario on 32 of 33 maps, producing 1,600 tasks, with populations from 32 to 8,000 agents. The excerpt does not provide every baseline value, a statistical significance analysis, or a compute-equalized runtime comparison.
On POGEMA, success rates generally decrease as teams become larger, but DMM-MICPO-3M retains a success rate of 1.000 on Warehouse at the largest evaluated size of 192 agents and 0.922 on Cities-Tiles at 256 agents. LC-MAPF-3M, the highest-success baseline at those maximum sizes, reaches 0.938 and 0.805, respectively, so the reported gap is especially visible in those domains. More refinement rounds can improve performance, but gains beyond four rounds become smaller; the MICPO checkpoint reaches a comparable level with fewer rounds. In the corridor experiment, a nonzero round-level teacher-forcing floor is important: setting it to zero removes concentration on the two valid resolutions and brings the frequencies closer to the near-uniform behavior of independent-sampling baselines.
Classical MAPF approaches include explicit search, reductions to optimization problems, and reactive coordination; the paper cites Conflict-Based Search, minimum-cost flow, and SAT as representative families. Learning-based MAPF instead acquires decision rules from data, but decentralized policies can lose coordination at the final independent sampling step even when communication has supplied richer context. DMM takes the position that communication should couple the intents while they are forming, not merely provide information before independent decisions are made.
In MAPF, every agent must reach its goal while avoiding vertex and edge conflicts with the other agents. When several joint actions are valid in the same context, the correct joint distribution may contain dependencies that are absent from its per-agent marginals. Independent sampling can therefore mix locally plausible choices from different valid resolutions. The resulting error is not only imperfect learning; it can be an irreducible consequence of factorizing the executor at commitment time.
The paper instantiates the same intent-refinement mechanism as DMM-3M and DMM-0.8M. The compact DMM-0.8M moves most observation-dependent computation outside the refinement loop, reuses it across rounds, and has 763,296 trainable parameters. Each agent receives a tokenized observation of up to 256 tokens encoding its local field of view, agent-specific attributes, and spatial context. DMM-3M retains the LC-MAPF Transformer encoder-decoder for controlled comparison, while the excerpt does not fully specify all layers, message dimensions, or training hyperparameters.
The corridor study isolates a symmetric ambiguity in which two agents start at opposite corridor ends and must exchange positions. Independent sampling allows four joint outcomes; LC-MAPF, MAGAT+, and HMAGAT distribute probability nearly uniformly over them, yielding about 50 percent valid joint actions. DMM refinement instead concentrates probability toward the two valid resolutions. In the scalability study, DMM-MICPO-0.8M solves every tested large-scale instance up to 1,048,576 simultaneously acting agents in obstacle-rich environments.
On MovingAI, DMM-MICPO-3M solves 1,598 of 1,600 tasks and has the highest reported coverage among the evaluated methods. On POGEMA, DMM generally obtains higher success rates and lower pooled SoC ratios than LC-MAPF-3M, with the pattern appearing for both the 0.8M and 3M variants. MICPO further improves both DMM variants over their imitation-pretrained counterparts. The paper reports solution costs close to those of the strongest baselines, but the supplied excerpt does not give the complete mean values for every condition.
The paper concludes that independently sampling final actions is a structural weakness of decentralized MAPF: individually plausible decisions can still form an incompatible joint action. DMM couples stochastic intents through local communication before simultaneous commitment, and MICPO fine-tunes this multi-round process from shared task outcomes. The reported results support improving decentralized coordination beyond one-shot independent sampling while preserving decentralized execution.
DMM does not enforce joint-action feasibility as a hard constraint, so refinement can still produce an incompatible joint action. Although the paper demonstrates large-scale execution, the excerpt does not establish performance under all communication failures, observation noise, environmental distributions, or out-of-distribution maps. More rounds can increase coordination quality but also add computation and delay, and the reported marginal improvements become smaller beyond four rounds. The authors identify explicit feasibility constraints inside the refinement process as a natural direction for future work.
The work is important because it attributes decentralized MAPF failures not only to inaccurate policy learning, but also to the final reconstruction of a joint action from correct-looking marginal choices. Sharing an intermediate intent state provides a way to model cross-agent dependence without replacing decentralized execution with a centralized joint-action selector. The reported ability to solve instances with more than one million simultaneously acting agents also makes the interaction between coordination design and execution scalability particularly relevant.
This paper is relevant to researchers in decentralized MAPF, learned pathfinding, multi-agent reinforcement learning, and communication-based coordination. It is especially suitable for readers studying multimodality caused by independent sampling, iterative intent formation, imitation-to-reinforcement-learning fine-tuning, and scalable multi-agent execution.