Pairwise is Not Enough: Hypergraph Neural Networks for Multi-Agent Pathfinding

Multi-Agent Path Finding (MAPF) is a representative multi-agent coordination problem, where multiple agents are required to navigate to their respective goals without collisions. Solving MAPF optimally is known to be NP-hard, leading to the adoption of learning-based approaches to alleviate the online computational burden. Prevailing approaches, such as Graph Neural Networks (GNNs), are typically constrained to pairwise message passing between agents. However, this limitation leads to suboptimal behaviours and critical issues, such as attention dilution, particularly in dense environments where group (i.e. beyond just two agents) coordination is most critical. Despite the importance of such higher-order interactions, existing approaches have not been able to fully explore them. To address this representational bottleneck, we introduce HMAGAT (Hypergraph Multi-Agent Attention Network), a novel architecture that leverages attentional mechanisms over directed hypergraphs to explicitly capture group dynamics. Empirically, HMAGAT establishes a new state-of-the-art among learning-based MAPF solvers: e.g., despite having just 1M parameters and being trained on 100$\times$ less data, it outperforms the current SoTA 85M parameter model. Through detailed analysis of HMAGAT’s attention values, we demonstrate how hypergraph representations mitigate the attention dilution inherent in GNNs and capture complex interactions where pairwise methods fail. Our results illustrate that appropriate inductive biases are often more critical than the training data size or sheer parameter count for multi-agent problems.

💡 Research Summary

The paper tackles the Multi‑Agent Path Finding (MAPF) problem, a classic NP‑hard coordination task where many agents must reach individual goals without colliding. While recent learning‑based MAPF solvers have adopted Graph Neural Networks (GNNs) or transformers, they are fundamentally limited to pairwise message passing. This limitation manifests as two critical issues: (1) “attention dilution” in dense scenarios, where the softmax‑normalized pairwise attention scores become spread thin across many irrelevant neighbours, and (2) an inability to capture genuine group‑level interactions that are essential for optimal MAPF solutions.

To overcome these shortcomings, the authors propose HMAGAT (Hypergraph Multi‑Agent Attention Network), a novel imitation‑learning architecture that replaces the GNN core of the state‑of‑the‑art MAGAT policy with a hypergraph‑based attention network. The key ideas are:

1. Directed Hypergraph Representation – Each hyperedge has a singleton head (the agent whose action is being decided) and a tail consisting of multiple agents that jointly influence that decision. Hyperedges are enriched with a 4‑dimensional feature vector encoding relative positions and Manhattan distance between tail agents and the head (or the centroid of the head agents).

2. Two‑Stage Hypergraph Attention – Message passing proceeds from tail nodes → hyperedge → head node. First, tail nodes compute attention weights (αᵉⱼ) toward the hyperedge; second, the hyperedge computes attention (αⁱᵉ) toward the head. Both stages use LeakyReLU followed by a softmax, but crucially the normalizations are performed locally within each hyperedge, preventing the dilution effect that occurs when a single softmax is applied over a large neighbourhood. The update equations (1‑4) show how node features, hyperedge features, and attention coefficients are combined.

3. Dynamic Hypergraph Construction – Three practical strategies are introduced:

   • Lloyd‑based Voronoi coloring with “soft borders” that allow vertices to belong to multiple colors, thereby forming overlapping groups.
   • k‑means‑based color diffusion, a faster alternative that diffuses color vectors over the graph and clusters them.
   • Shortest‑distance based hyperedges, where agents that can be encountered on a shortest‑path between two others are placed together in a tail.
     These methods generate hypergraphs on‑the‑fly at each timestep, respecting a communication radius R₍comm₎.

4. Learning Pipeline – Expert demonstrations are collected from a high‑quality MAPF solver over 21 K instances using the POGEMA toolkit. For each timestep, the current configuration is encoded by a CNN (obstacle map, agent map, goal direction, normalized cost‑to‑go) producing node embeddings. The HGNN layers then process the hypergraph, and an MLP decoder outputs a distribution over the five possible actions (stay or move to a neighboring cell). After the main training phase, a temperature‑sampler is fine‑tuned on intermediate‑difficulty instances to increase policy confidence.

Experimental Findings

• Parameter and Data Efficiency – HMAGAT uses ~1 M parameters and only ~1 % of the data required by the previous best model (an 85 M‑parameter MAPF‑GPT). Despite this, it achieves lower sum‑of‑costs (SoC) and higher success rates across a range of map sizes and agent densities.
• Robustness in High‑Density Scenarios – In “High AD” (high agent density) maps, pairwise GNN‑based policies suffer from severe attention dilution, leading to many collisions and inflated makespans. HMAGAT’s hypergraph attention maintains high weights on truly relevant agents, dramatically reducing collisions and improving makespan.
• Ablation and Attention Visualization – Heat‑maps of attention coefficients reveal that HMAGAT concentrates its focus on agents that share a hyperedge with the head, while ignoring distant, irrelevant agents. An informal proof in the appendix confirms that the two‑stage softmax prevents dilution.
• Comparison to Baselines – The paper benchmarks against MAGAT, G2RL, MAPF‑GPT, SCRIMP, and several recent transformer‑based approaches. HMAGAT consistently outperforms them on both solution quality (lower SoC) and scalability (larger maps, more agents).

Contributions Summarized

1. Introduction of directed hypergraphs with singleton heads to explicitly model group influences in MAPF.
2. Design of a hypergraph attention network that mitigates attention dilution and captures higher‑order interactions.
3. Practical, scalable hypergraph generation algorithms (Lloyd, k‑means, distance‑based) suitable for large, dense environments.
4. Demonstration that strong inductive biases (hypergraph structure) can outweigh raw parameter count and data volume in multi‑agent learning.

Implications – By showing that hypergraph‑based representations can efficiently encode complex multi‑agent dynamics, the work opens a pathway for applying similar ideas to other coordination‑intensive domains such as warehouse robotics, autonomous vehicle fleets, and collaborative drone swarms. The architecture’s lightweight nature also makes it attractive for real‑time deployment on embedded platforms where computational resources are limited. Overall, HMAGAT represents a significant step forward in bridging the gap between optimal MAPF theory and practical, scalable, learning‑driven solutions.