CoordField: Coordination Field for Agentic UAV Task Allocation in Low-Altitude Urban Scenarios

Tengchao Zhang, Yonglin Tian, Fei Lin, Jun Huang, Patrik Péter Süli, Qinghua Ni, Rui Qin, Xuan Wang, Fei-Yue Wang

TL;DR

This paper introduces CoordField, a novel agentic system that leverages Large Language Models (LLMs) and a "coordination field" to enable heterogeneous UAV swarms to dynamically allocate tasks in complex urban environments. CoordField translates natural language commands into actionable tasks and guides UAVs using a continuous potential field, leading to more efficient and adaptive swarm coordination.

What Problem Does This Paper Address?

The increasing use of Unmanned Aerial Vehicle (UAV) swarms for tasks like surveillance, delivery, and inspection in urban settings presents significant challenges. These include understanding complex, high-level human instructions, planning and adapting to dynamic task requirements, and coordinating multiple heterogeneous UAVs (drones with different capabilities) in real-time. Existing methods often struggle with this complexity, being either too rigid, task-specific, or lacking the ability to adapt quickly to changing urban environments. This paper aims to bridge this gap by proposing a system that combines the semantic understanding power of LLMs with a decentralized, field-based coordination mechanism for robust and adaptive UAV swarm operation.

Methods

CoordField is built around an agentic AI paradigm, integrating three core modules:

  • Semantic Understanding Module: This module utilizes LLMs (specifically, the DeepSeek API is mentioned) to interpret natural language instructions. It converts these instructions into structured task tuples, each containing the target location, task type, priority, and other relevant semantic information. This acts as a crucial bridge between human intent and the system's operational logic.
  • Planning Module and Coordination Field: This is the heart of CoordField. It takes the structured task tuples and UAV state data to create a dynamic coordination strategy.
  • Perception Mapping: It constructs a time-varying "task landscape" represented by a scalar potential field. This field reflects the spatial distribution and urgency of tasks, effectively highlighting "hotspots" of demand. Obstacles like buildings are excluded from this field.
  • Task Decomposition (Vector Field): To guide UAVs, a time-varying vector field is generated. This field represents the desired velocity for UAVs, derived from fluid dynamics principles (Navier-Stokes equations). UAVs follow the local gradient of this field to move towards high-priority task areas. A diffusion term helps prevent over-concentration.
  • Task Assignment (Local Repulsion): To prevent resource redundancy and conflicts, a local vortex mechanism is introduced around each UAV. This creates a repulsive field, ensuring agents maintain a safe distance and distribute themselves effectively. The final control velocity for each UAV is a combination of the global guiding field and these local repulsive velocities.
  • Execution Module: This module translates the planning module's outputs (velocity vectors and task assignments) into executable control commands for the UAVs. It manages communication with the UAV platform via standardized APIs (like MAVSDK, ROS2, or AirSim). The system operates in a closed-loop, continuously updating the fields based on real-time UAV status and environmental changes.

Datasets

The paper describes experiments conducted in a custom-designed 2D urban simulation environment. This environment includes a structured road network, buildings, traffic lights, and dynamically moving pedestrians and vehicles. Task commands are issued via natural language and parsed by the DeepSeek API. No specific public datasets are mentioned for training or evaluation; the focus is on a simulated environment designed to mimic real-world urban complexity.

Key Insights and Ideas

  • LLMs as the "Brain" for UAV Swarms: The paper demonstrates how LLMs can effectively translate ambiguous natural language commands into precise, actionable tasks for UAVs, significantly simplifying human-swarm interaction.
  • Coordination Field for Decentralized Control: The novel "coordination field" approach offers a powerful way to achieve decentralized task allocation. By representing task demand and UAV influence as continuous fields, UAVs can make local decisions that contribute to global swarm objectives without centralized control.
  • Adaptive Response to Dynamic Environments: The field-based method, combined with real-time updates, allows the system to react swiftly to changing task requirements or environmental conditions, a critical capability for urban operations.
  • Balancing Attraction and Repulsion: The combination of an attractive potential field (for task seeking) and a repulsive vortex field (for inter-agent avoidance) is key to achieving both efficient task coverage and robust spatial distribution of the swarm.

Key Results

The experimental results, conducted in a simulated 2D urban environment with 20 UAVs, show that CoordField outperforms baseline methods across several metrics:

  • Task Parsing Accuracy (TPA): CoordField achieved 96% accuracy, significantly higher than LLM-based baselines like Deepseek-v3 (70%), GPT-40 (74%), Claude-3-7-Sonnet (76%), LLaMA-4-Scout (72%), and Gemini-2.5-Pro (73%).
  • Coverage Efficiency (CE): CoordField achieved 95% CE, also outperforming LLM baselines.
  • Task Load Balance (TLB): CoordField demonstrated a TLB of 0.8, indicating a more even distribution of tasks among UAVs compared to LLM baselines which had TLBs ranging from 1.5 to 2.1.
  • UAV Utilization Rate (UUR): CoordField achieved a high UUR of 97%.

When compared against traditional algorithms (ACO, GWO, WOA, A*) for the coordination field part, CoordField showed competitive performance, particularly excelling in Coverage Efficiency and Task Load Balance, while maintaining a high Completion Rate.

Limitations and Future Work

  • Simulation-Based: The current evaluation is based on a 2D simulation. The authors acknowledge the need to extend the system to three-dimensional scenarios and real-world UAV swarms to validate its practical applicability.
  • LLM Dependency: While LLMs are powerful, their performance can be influenced by prompt engineering and the specific model used. The paper highlights comparative LLM performance, suggesting that the choice of LLM can impact overall system effectiveness.

Future work will focus on extending the system to 3D environments and conducting real-world deployments to further validate its capabilities.

Summary updated

Abstract

With the increasing demand for heterogeneous Unmanned Aerial Vehicle (UAV) swarms to perform complex tasks in urban environments, system design now faces major challenges, including efficient semantic understanding, flexible task planning, and the ability to dynamically adjust coordination strategies in response to evolving environmental conditions and continuously changing task requirements. To address the limitations of existing methods, this paper proposes CoordField, a coordination field agent system for coordinating heterogeneous drone swarms in complex urban scenarios. In this system, large language models (LLMs) is responsible for interpreting high-level human instructions and converting them into executable commands for the UAV swarms, such as patrol and target tracking. Subsequently, a Coordination field mechanism is proposed to guide UAV motion and task selection, enabling decentralized and adaptive allocation of emergent tasks. A total of 50 rounds of comparative testing were conducted across different models in a 2D simulation space to evaluate their performance. Experimental results demonstrate that the proposed system achieves superior performance in terms of task coverage, response time, and adaptability to dynamic changes.
BibTeX
@misc{zhang2025coordfield,
  author = {Tengchao Zhang and Yonglin Tian and Fei Lin and Jun Huang and Patrik Péter Süli and Qinghua Ni and Rui Qin and Xuan Wang and Fei-Yue Wang},
  title = {{CoordField: Coordination Field for Agentic UAV Task Allocation in Low-Altitude Urban Scenarios}},
  year = {2025},
  eprint = {2505.00091},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.00091},
}
Figure