PC2P: Multi-Agent Path Finding via Personalized-Enhanced Communication and Crowd Perception

TL;DR

PC$^2$P addresses distributed MAPF under partial observability by introducing personalization in inter-agent communication and crowd-aware perception, combined with a region-based deadlock-breaking strategy. It integrates a Graph Attention Network-based personalized communication module with a crowdsourced local observation channel and uses a D3QN-style Q-network to derive policies; deadlocks are resolved via expert-guided local planning. Ablation studies show each component contributes to robustness, particularly in high-density and large-scale maps, achieving superior success rates and competitive path lengths compared to DHC, DCC, SCRIMP, and EPH*. The work demonstrates improved scalability and real-time performance, with future directions toward CTDE to stabilize learning.

Abstract

Distributed Multi-Agent Path Finding (MAPF) integrated with Multi-Agent Reinforcement Learning (MARL) has emerged as a prominent research focus, enabling real-time cooperative decision-making in partially observable environments through inter-agent communication. However, due to insufficient collaborative and perceptual capabilities, existing methods are inadequate for scaling across diverse environmental conditions. To address these challenges, we propose PC2P, a novel distributed MAPF method derived from a Q-learning-based MARL framework. Initially, we introduce a personalized-enhanced communication mechanism based on dynamic graph topology, which ascertains the core aspects of ``who" and ``what" in interactive process through three-stage operations: selection, generation, and aggregation. Concurrently, we incorporate local crowd perception to enrich agents' heuristic observation, thereby strengthening the model's guidance for effective actions via the integration of static spatial constraints and dynamic occupancy changes. To resolve extreme deadlock issues, we propose a region-based deadlock-breaking strategy that leverages expert guidance to implement efficient coordination within confined areas. Experimental results demonstrate that PC2P achieves superior performance compared to state-of-the-art distributed MAPF methods in varied environments. Ablation studies further confirm the effectiveness of each module for overall performance.

Figures (7)

• Figure 1: The PC$^2$P framework comprises four components: Observation Module (green), Communication Module (blue), Q-network Module (pink), and Deadlock Resolution Module (purple). Agents gather local observations (red dashed box), include basic and crowd-augmented information, which are then encoded by the Observation Encoder. The Communication Module selects relevant agents and aggregates personalized information (detailed above), enabling the Q-network Module to generate policies. Finally, upon detecting extreme deadlocks, PC$^2$P activates the Deadlock Resolution Module to output alternative strategies.
• Figure 2: Illustration of the Personalized-Enhanced Communication Mechanism. (1) Agent $i$ selects Agent $j$ for communication via the Decision Causal Unit. (2) Agent $i$ sends information $m_t^i$ to agent $j$ to generate the personalized message $\hat{m}_t^{ij}$. (3) Agent $i$ aggregates the received $\hat{m}_t^{ij}$ with its own information $m_t^i$ to complete the communication process.
• Figure 3: Illustration of how agent $i$ acquires the crowd-augmented channel. To enhance clarity, both FOV (blue bounding box) and the crowd perception region (red bounding box) in the figure are set to a $3 \times 3$ area as an example. Crowding density is calculated based on the number of agents and obstacles within the $3 \times 3$ region surrounding $i$. Each directional channel (Up, Down, Left, Right) is marked as 1 if the corresponding action reduces the comprehensive congestion degree for $i$; otherwise, it is 0.
• Figure 4: Illustration of a typical deadlock-breaking example. First, agents $j, i, k$ are identified as a deadlocked group. Next, a Local Map is constructed, and a subgoal is generated using A* for agent $i$. With expert guidance, the agents collaborate to resolve the deadlock. Finally, at time step t+2, the task is continued by the RL planner.
• Figure 5: Experimental Results of SR and EL on random maps.