LaGAT: Graph Attention-Guided Dense MAPF Solver

• LaGAT is a hybrid solver for dense MAPF that combines decentralized graph-attention (MAGAT*) with complete search (LaCAM) to coordinate multiple agents in tight, occupied maps.
• It employs LaCAM as the outer search shell and PIBT for collision-free successor generation, while MAGAT* provides learned local action preferences for improved decision-making.
• The system incorporates deadlock detection and map-specific fine-tuning, ensuring completeness and robust performance even in challenging, high-density scenarios.

Searching arXiv for the exact term and closely related variants to verify current naming and support disambiguation. LaGAT is a hybrid solver for dense multi-agent pathfinding (MAPF) introduced in “Graph Attention-Guided Search for Dense Multi-Agent Pathfinding” (Jain et al., 20 Oct 2025). The name abbreviates Graph Attention-Guided Search and denotes a system that embeds an enhanced decentralized graph-attention policy, $\mathrm{MAGAT}^\star$, into the complete search-based planner LaCAM. Its target regime is dense MAPF on small or medium maps with high occupancy and tight obstacle structure, where purely search-based methods can be myopic and purely learning-based policies can be fast but incomplete, collision-prone, or deadlock-prone (Jain et al., 20 Oct 2025).

1. Problem setting and formal scope

LaGAT is defined for standard one-shot MAPF on a 4-connected grid graph $G=(V,E)$ with agents $A=\{1,\dots,n\}$, distinct starts $s_i$, and goals $g_i$. At each timestep, an agent either waits or moves to a neighboring cell; vertex collisions and edge swaps are forbidden. A solution is a set of collision-free paths

$\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$

and solution quality is measured by sum-of-costs (SoC), i.e. the total travel time until each agent reaches goal and stops moving. A global state is a configuration

$Q \in V^n,$

recording every agent’s current location (Jain et al., 20 Oct 2025).

The paper is explicitly concerned with dense MAPF, meaning settings with small or medium maps, high occupancy, tight obstacle structure, and many mutually interfering agents. In this regime, local motions are strongly coupled: an agent’s move often depends on several others moving first. The paper’s diagnosis is that this produces two complementary failure modes. Search-based planners such as LaCAM or LaCAM$^\ast$ can find solutions quickly, but their default local guidance is myopic and often yields feasible yet substantially suboptimal coordination. Decentralized learned policies can encode local interaction structure and run quickly, but they are imperfect, can deadlock or livelock, and lack completeness guarantees (Jain et al., 20 Oct 2025).

LaGAT is therefore designed as a division of labor. Learning supplies a strong local coordination prior in the form of action preferences, while search supplies reliability, systematic exploration, and completeness. The method is not a replacement of combinatorial search by a neural controller; it is a search algorithm whose successor generation is guided by a graph-attention policy (Jain et al., 20 Oct 2025).

2. Constituent mechanisms: LaCAM, PIBT, and $\mathrm{MAGAT}^\star$

LaGAT is built from three algorithmic components with distinct roles. LaCAM is the outer configuration-space search procedure. PIBT is the low-level transition generator that produces a collision-free next configuration from a current one, given per-agent action preferences. $\mathrm{MAGAT}^\star$ is the learned policy that replaces LaCAM’s default preference construction with graph-attention-based action ranking (Jain et al., 20 Oct 2025).

Component	Role in LaGAT	Key details
LaCAM	Complete outer search shell	Searches over configurations $G=(V,E)$0
PIBT	One-step successor generation	Uses sorted candidate moves per agent
$G=(V,E)$1	Learned local guidance	Outputs action distributions used to rank PIBT moves

LaCAM avoids exhaustive enumeration of successor configurations, which would be exponential in the number of agents. Instead, it lazily generates successors through PIBT under a search tree of constraints. The simplified procedure given in the paper initializes the root node with

$G=(V,E)$2

maintains an open stack and explored table, repeatedly pops a search node, checks whether the goal configuration $G=(V,E)$3 has been reached, and otherwise pops a constraint set, updates constraints, calls a configuration generator, and inserts a new search node if the generated configuration is valid and unexplored (Jain et al., 20 Oct 2025).

The default PIBT preference rule is goal-directed: candidate moves are sorted by increasing cost-to-go $G=(V,E)$4. LaGAT changes exactly this step. Instead of using only shortest-path distance to the goal, it supplies PIBT with learned action rankings derived from $G=(V,E)$5, while leaving LaCAM’s complete search structure intact (Jain et al., 20 Oct 2025).

The learned policy $G=(V,E)$6 extends MAGAT with three modifications: an extra cost-to-go observation channel, edge features in graph attention, and three stacked graph-attention layers instead of one. Each agent $G=(V,E)$7 receives a local observation tensor $G=(V,E)$8 centered at its current position $G=(V,E)$9, of shape

$A=\{1,\dots,n\}$0

with $A=\{1,\dots,n\}$1. The four channels are obstacle occupancy, other-agent occupancy, projected goal direction, and a normalized cost-to-go channel storing

$A=\{1,\dots,n\}$2

for each location $A=\{1,\dots,n\}$3 in the field of view. Agents communicate through a graph $A=\{1,\dots,n\}$4 defined by spatial proximity within $A=\{1,\dots,n\}$5. Each edge carries a 3D feature $A=\{1,\dots,n\}$6 containing relative $A=\{1,\dots,n\}$7-offset, relative $A=\{1,\dots,n\}$8-offset, and Manhattan distance (Jain et al., 20 Oct 2025).

3. Graph-attention-guided successor generation

The central technical contribution of LaGAT is not a global heuristic over search nodes but a learned mechanism for constructing per-agent action preferences before PIBT is invoked. This is why the paper describes LaGAT as guided search rather than as a neural replacement for search (Jain et al., 20 Oct 2025).

The graph-attention update in $A=\{1,\dots,n\}$9 uses both node and edge features. Let $s_i$0 denote the feature of agent $s_i$1 at GNN layer $s_i$2, and let $s_i$3 be the processed edge feature. The layer is

$s_i$4

$s_i$5

$s_i$6

$s_i$7

These equations make the attention both message-aware and edge-aware: neighbor importance depends on sender features and relative geometry (Jain et al., 20 Oct 2025).

The output of $s_i$8 is an action distribution

$s_i$9

over candidate next target vertices $g_i$0, including wait and adjacent cells. In LaGAT, this distribution is not used as a direct executor, not converted into a scalar heuristic $g_i$1, and not used to score LaCAM’s open list. Instead, it is used to sort PIBT’s candidate moves. For guided agents, LaGAT applies the ranking

$g_i$2

whereas unguided agents retain the default PIBT order

$g_i$3

The network therefore intervenes only at the configuration generation stage (Jain et al., 20 Oct 2025).

The paper reports that deterministic use of the neural ranking yields higher-quality actions than stochastic sampling, although deterministic deployment also exposes repetitive failure modes such as oscillation and local traps. This observation directly motivates the deadlock detection mechanism that distinguishes LaGAT from a simple “policy-inside-search” construction (Jain et al., 20 Oct 2025).

4. Deadlock detection, fallback override, and completeness

A major contribution of LaGAT is a deadlock/livelock detection mechanism designed to compensate for imperfect neural guidance under deterministic deployment. The system maintains, for each search node, a set of agents that are no longer guided by the neural model. Initially this set is empty (Jain et al., 20 Oct 2025).

After generating a new configuration $g_i$4, LaGAT traces back through up to depth $g_i$5 ancestors, with $g_i$6 typically less than 3. For any unfinished agent $g_i$7, it tests whether the location is unchanged,

$g_i$8

and whether the agent’s local neighborhood is unchanged. The paper defines the local neighborhood relation as

$g_i$9

If

$\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$0

the agent is deemed stuck in a repeated local situation, i.e. a deadlock or livelock (Jain et al., 20 Oct 2025).

When such repetition is detected, the affected agent is inserted into the ancestor node’s unguided set; that ancestor’s constraint tree is reset to the initial unconstrained state; and the ancestor node is reinserted into the open structure. The effect is twofold: future expansions from that state no longer trust the learned policy for that agent, and the search is not trapped by already explored constraint configurations that may have reinforced the bad behavior (Jain et al., 20 Oct 2025).

The paper proves the following result: LaCAM with deadlock detection is complete. The proof sketch is that LaCAM is complete regardless of PIBT preferences, and deadlock detection causes only finite resets because once an agent is placed into a node’s unguided set, it stays unguided thereafter. Since the number of agents is finite, the number of such updates is finite, so all successors remain eventually explorable. This is central to the method’s interpretation: LaGAT preserves completeness while using learned guidance aggressively where that guidance is useful (Jain et al., 20 Oct 2025).

The paper does not claim optimality for the base algorithm. Rather, LaGAT is presented as a fast complete search method for obtaining higher-quality real-time solutions in dense regimes. The paper states that a LaCAM$\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$1-style eventually optimal variant remains readily attainable (Jain et al., 20 Oct 2025).

5. Training regime, specialization, and implementation

The learned component of LaGAT is trained in two stages: general pretraining and map-specific fine-tuning. This training strategy is treated as one of the core reasons prior learning-guided MAPF search methods underperformed and LaGAT does not (Jain et al., 20 Oct 2025).

In the pretraining stage, the authors generate 21K MAPF instances with the POGEMA toolkit: 20% with random obstacles and 80% maze-like maps, with map sizes from $\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$2 to $\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$3, and agent counts $\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$4. Expert demonstrations are produced with LaCAM3 using the timeout schedule

$\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$5

The paper emphasizes that most pretraining instances are solved within 1 second. $\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$6 is then trained by imitation learning with cross-entropy loss on expert actions. The setup uses 200 epochs, AdamW, batch size 64, cosine learning-rate decay from $\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$7 to $\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$8, and about 100 hours on an NVIDIA L40S GPU (Jain et al., 20 Oct 2025).

To mitigate distribution shift, the training loop performs dataset aggregation every four epochs. It samples 500 training instances from the original dataset, runs the current $\pi_i=(v^0=s_i,v^1,\dots,v^T=g_i),$9, and for failed instances invokes LaCAM3 as an online expert from the final configuration; those expert trajectories are then added to the training set. The paper explicitly describes this as on-demand dataset aggregation in the spirit of Ross et al. (2011) and DAgger (Jain et al., 20 Oct 2025).

The second stage is map-wise fine-tuning. For a target deployment map $Q \in V^n,$0, the system generates 1000 additional instances on that exact map and solves them with LaCAM3. These map-specific instances use higher densities, with $Q \in V^n,$1 agents, and the pretrained policy is fine-tuned for 52 epochs, usually taking 4–8 hours. This map-specific specialization is presented as practically important for fixed-layout domains such as warehouses and rail scheduling, where coordination patterns depend strongly on the exact topology of the deployment map (Jain et al., 20 Oct 2025).

Implementation choices are also operationally important. $Q \in V^n,$2 is implemented in Python and translated to C++ for faster inference; LaGAT itself is implemented in C++. The paper uses $Q \in V^n,$3, $Q \in V^n,$4, and optional LNS postprocessing for the 30-second planner comparison (Jain et al., 20 Oct 2025).

6. Empirical behavior and performance regime

The empirical results distinguish between the standalone learned policy and the full hybrid planner. For the standalone policy benchmark on 3456 instances from POGEMA over Warehouse, Dense Warehouse, Room, Dense Room, Maze, and Dense Maze, the paper reports several qualitative conclusions: imitation learning outperforms reinforcement learning on these policy benchmarks, CS-PIBT materially improves policy effectiveness, and $Q \in V^n,$5 attains success rates comparable to MAPF-GPT despite far fewer parameters and much smaller training data (Jain et al., 20 Oct 2025).

The reported parameter counts are $Q \in V^n,$6 at 760K, MAPF-GPT at 2M / 6M / 85M, SSIL at 160K, SCRIMP at 8M, and DCC at 2M. Inference speed is especially relevant because the model is called repeatedly during search. The paper states that the smallest MAPF-GPT model requires about 50 ms per timestep for 128 agents, while $Q \in V^n,$7 runs in $Q \in V^n,$8 ms (Jain et al., 20 Oct 2025).

The main planner comparison uses a 30-second deadline and focuses on normalized cost

$Q \in V^n,$9

The strongest gains occur in the dense settings Dense Warehouse, Dense Room, and Dense Maze. The paper states that LaGAT takes more time to derive the first solution than pure search because of neural inference overhead, but its initial solution quality is consistently much better than both MAPF-GPT and LaCAM3, and its final quality at 30 seconds is markedly better than LaCAM3 after both are allowed postprocessing (Jain et al., 20 Oct 2025).

A particularly informative average tradeoff plot for the densest instances of Dense Warehouse, Dense Maze, and Dense Room combined gives the following example values: $^\ast$0 standalone has initial normalized cost about 11.64 at about 3.06 s; LaCAM3 initial is about 13.08 at 4.43 s; LaCAM3 after 30 s is about 7.33; LaGAT initial is about 4.99 at 9.92 s; and LaGAT after 30 s is about 4.38. These values capture the paper’s central claim: LaGAT spends more time up front but lands in a much better region of solution space, and LaCAM3’s anytime improvements do not catch up by 30 seconds in the hardest dense cases (Jain et al., 20 Oct 2025).

On larger and sparser maps such as ost003d and lak303d, the picture is more nuanced. The paper reports that LaCAM3 already produces near-optimal solutions quickly; LaGAT’s initial solutions can be slightly worse; and LNS postprocessing usually makes final performance comparable, though LaGAT sometimes ends slightly ahead. This is presented as a regime with less room for learned guidance because greedy collision avoidance and global map structure are already sufficient (Jain et al., 20 Oct 2025).

Success-rate improvements also appear in some difficult cases. On Dense Random at the highest density, LaGAT achieves 93.75% success within 30 seconds, whereas LaCAM3 achieves 75%. The paper’s ablation narrative further states that all components are necessary for the best dense-performance point: removing map-specific fine-tuning, pretraining, neural guidance, deadlock detection, or search each degrades either solvability or solution quality (Jain et al., 20 Oct 2025).

7. Interpretation, limitations, and disambiguation

LaGAT is best understood as a tightly integrated hybrid of three named elements: LaCAM is the complete configuration-space search algorithm; MAGAT is the neural graph-attention policy family; and LaGAT is the system that combines LaCAM with $^\ast$1, map-wise fine-tuning, and deadlock detection. In the paper’s own framing, it is not merely “MAGAT inside LaCAM,” but a particular recipe in which a compact, fast, edge-aware graph-attention policy is specialized to the target map and then safely overridable during search (Jain et al., 20 Oct 2025).

The principal limitations are also explicit. Strong performance depends on map-specific fine-tuning, which is practical for fixed-layout deployments but reduces plug-and-play universality. $^\ast$2 still operates with bounded observation and communication radii, so its local nature can produce suboptimality on much larger maps. LaGAT is not the fastest method to a first feasible solution because neural inference introduces overhead. Finally, the largest gains are density-dependent: in sparse or easy instances, strong search already performs near-optimally, leaving little room for learned guidance (Jain et al., 20 Oct 2025).

The term LaGAT should also be distinguished from several nearby acronyms and names in other literatures. LA-GAT denotes a Lane-Aware Graph Attention Network for multi-vehicle trajectory prediction in expressway merge zones and is unrelated to MAPF search (Laughter, 12 May 2026). L-CAT denotes a learnable graph convolutional attention network that interpolates among GCN-, GAT-, and CAT-like behavior in graph neural networks, and it is a distinct graph-learning architecture rather than a MAPF solver (Javaloy et al., 2022). AC-GATE is a lag-gated framework for panel time series and is conceptually adjacent only in the very broad sense of using a gate over structured dependencies (Xu, 20 May 2026). Work on LGAD gain-layer design concerns radiation-hard detector engineering and is unrelated to graph attention or MAPF (Sola et al., 2022).

In its exact technical sense, then, LaGAT refers to the dense-MAPF method of (Jain et al., 20 Oct 2025): a complete LaCAM-based search procedure whose local successor generation is guided by an enhanced graph-attention policy and protected by deadlock-aware fallback.