The coordination of multiple Unmanned Aerial Vehicles (UAVs), or drone formations, has transitioned from a research concept to a critical capability for numerous civilian and military applications. These include automated logistics, infrastructure inspection, precision agriculture, and aerial surveillance. A well-executed drone formation leverages the strengths of distributed systems—enhanced robustness, scalability, and flexibility—compared to a single, complex UAV. However, maintaining a precise drone formation in dynamic, three-dimensional environments fraught with static and moving obstacles presents a significant control challenge. The core objectives are twofold: to achieve and sustain a desired geometric configuration and to ensure collision-free navigation for all members of the fleet.
Among the various control strategies, the Artificial Potential Field (APF) method offers a computationally efficient and intuitively appealing solution for drone formation and obstacle avoidance. The fundamental principle involves modeling the UAV’s environment with virtual forces. Attractive potentials pull drones towards their goals or desired relative positions, while repulsive potentials push them away from other drones and obstacles. The net force dictates the UAV’s motion. Despite its advantages, the classical APF method is notoriously plagued by local minima—situations where attractive and repulsive forces balance, trapping the drone before it reaches its target. Furthermore, traditional 2D APF models are insufficient for 3D aerial navigation, and their formulation often lacks explicit mechanisms for smooth, dynamic reconfiguration of the drone formation shape.
This article presents a comprehensive and enhanced 3D Artificial Potential Field (3D-APF) framework specifically designed for distributed multi-UAV systems. The proposed method systematically addresses the aforementioned limitations. It extends the potential field into three dimensions, incorporates communication topology from consensus theory for superior formation coordination, and introduces a novel rotational auxiliary field to definitively resolve local minima, particularly against moving obstacles. The subsequent sections will detail the mathematical formulation, control algorithms, and provide extensive simulation evidence of its efficacy.
Foundations of the Potential Field Approach
The kinematics and dynamics of a quadrotor UAV, a standard platform for drone formation research, are described by a set of nonlinear equations. The position of the i-th drone in the inertial frame is given by $\mathbf{P}_i = [x_i, y_i, z_i]^T$. A simplified translational dynamics model often used for high-level path planning is:
$$ m_i \ddot{\mathbf{P}}_i = \mathbf{F}_{control,i} – \mathbf{F}_{drag,i} + \mathbf{F}_{PF,i} $$
where $m_i$ is the mass, $\mathbf{F}_{control,i}$ is the vehicle’s thrust vector, $\mathbf{F}_{drag,i}$ is the aerodynamic damping force, and $\mathbf{F}_{PF,i}$ is the resultant virtual force derived from the artificial potential field, which is the focus of our control design.
In a distributed drone formation, information flow is limited. This structure is modeled as an undirected graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$, where $\mathcal{V} = {1, 2, …, n}$ is the set of drones, and $\mathcal{E}$ defines the communication links. The adjacency matrix $\mathbf{A} = [a_{ij}]$ encodes this topology, where $a_{ij} > 0$ if drones $i$ and $j$ can exchange data, and $a_{ij}=0$ otherwise. The degree of a node is $N_i = \sum_{j=1}^{n} a_{ij}$. Our enhanced APF framework explicitly utilizes this graph structure to calculate cooperative forces.
The Enhanced 3D Potential Field Formulation
Our framework employs a “leader-follower” drone formation architecture within a distributed communication setting. Only the leader has knowledge of the global mission target. The key innovation lies in the design of separate yet interconnected potential fields for formation keeping and obstacle avoidance, both extended and modified for 3D space.
Formation Control Potential Fields
The formation control potential is decomposed into a leader target attraction field and an inter-drone formation attraction field. This decomposition allows the fleet to first converge to a common operational altitude before refining the planar formation shape, enhancing efficiency and safety during deployment.
1. Leader Attraction Field: For the leader drone $L$, an attractive potential $U_L^{att}$ pulls it towards the target point $\mathbf{P}_{goal}=[x_g, y_g, z_g]^T$. Crucially, this field is designed with independent gains for horizontal and vertical convergence:
$$ U_L^{att} = \frac{1}{2} k_{p}^{L} \| \mathbf{P}_L^{horz} – \mathbf{P}_{goal}^{horz} \|^2 + \frac{1}{2} k_{z}^{L} (z_L – z_g)^2 $$
Here, $k_{p}^{L} > 0$ and $k_{z}^{L} > 0$ are tuning gains. The superscript $horz$ denotes the projection onto the horizontal (x-y) plane. The corresponding attractive force is the negative gradient:
$$ \mathbf{F}_L^{att} = -\nabla U_L^{att} = -k_{p}^{L} (\mathbf{P}_L^{horz} – \mathbf{P}_{goal}^{horz}) – k_{z}^{L} (z_L – z_g)\hat{\mathbf{k}} $$
where $\hat{\mathbf{k}}$ is the unit vector along the z-axis. This design prioritizes quick altitude alignment, a critical first step in 3D drone formation assembly.
2. Inter-Drone Formation Field: For a follower drone $i$, the formation potential is built relative to its neighboring drone $j$. Unlike simple distance-based attraction, this field incorporates the desired relative position $\mathbf{P}_{ij}^{des}$, enabling explicit shape control and reconfiguration. The pairwise potential is:
$$ U_{ij}^{form} = \frac{1}{2} k_{p}^{F} \| (\mathbf{P}_i – \mathbf{P}_j) – \mathbf{P}_{ij}^{des} \|^2 + \frac{1}{2} k_{v}^{F} \| \dot{\mathbf{P}}_i – \dot{\mathbf{P}}_j \|^2 $$
The term $k_{p}^{F} \| (\mathbf{P}_i – \mathbf{P}_j) – \mathbf{P}_{ij}^{des} \|^2$ minimizes position error relative to the desired formation geometry. The term $k_{v}^{F} \| \dot{\mathbf{P}}_i – \dot{\mathbf{P}}_j \|^2$ promotes velocity consensus, dampening oscillations and smoothing the drone formation motion. To account for the distributed communication topology, the total formation potential for drone $i$ is a weighted sum over its neighbors:
$$ U_i^{form} = \frac{1}{N_i} \sum_{j=1}^{n} a_{ij} \cdot U_{ij}^{form} $$
The resulting formation force on drone $i$ is:
$$ \mathbf{F}_i^{form} = -\nabla U_i^{form} = -\frac{1}{N_i} \sum_{j=1}^{n} a_{ij} \left[ k_{p}^{F} ((\mathbf{P}_i – \mathbf{P}_j) – \mathbf{P}_{ij}^{des}) + k_{v}^{F} (\dot{\mathbf{P}}_i – \dot{\mathbf{P}}_j) \right] $$
Collision Avoidance Potential Fields
The avoidance module consists of two repulsive fields: one for inter-drone collision prevention and a novel composite field for obstacle avoidance.
1. Inter-Drone Repulsion Field: A short-range repulsive field prevents collisions within the drone formation. For drones $i$ and $j$, the potential is active only within a critical distance $\rho_0$:
$$
U_{ij}^{rep} =
\begin{cases}
\frac{1}{2} \eta_{d} \left( \frac{1}{ \|\mathbf{P}_i – \mathbf{P}_j\| } – \frac{1}{\rho_0} \right)^2, & \text{if } \|\mathbf{P}_i – \mathbf{P}_j\| \leq \rho_0 \\
0, & \text{if } \|\mathbf{P}_i – \mathbf{P}_j\| > \rho_0
\end{cases}
$$
where $\eta_d > 0$ is a repulsive gain. The collective repulsive force on drone $i$ from all neighbors is:
$$ \mathbf{F}_i^{d-rep} = -\nabla \left( \frac{1}{N_i} \sum_{j=1}^{n} a_{ij} U_{ij}^{rep} \right) $$
2. Obstacle Avoidance Field: This is the core innovation to solve local minima. The field for drone $i$ regarding obstacle $o$ has two components: a Primary Repulsive Field and an Auxiliary Rotational Field.
- Primary Repulsive Field ($U_{io}^{pri}$): This is a standard repulsive field, but modulated by a function $g(v_o)$ of the obstacle’s velocity $v_o$ to increase its influence zone for fast-approaching obstacles.
$$ U_{io}^{pri} = \frac{1}{2} \eta_o \cdot g(v_o) \left( \frac{1}{d_{io}} – \frac{1}{\rho_o} \right)^2, \quad \text{for } d_{io} \leq \rho_o $$
$$ g(v_o) = k_1 + \frac{k_2}{1 + e^{-|v_o|}} $$
Here, $d_{io} = \|\mathbf{P}_i – \mathbf{P}_o\|$, $\rho_o$ is the obstacle’s sensing range, $\eta_o>0$, and $k_1, k_2$ are constants. - Auxiliary Rotational Field ($U_{io}^{aux}$): This field generates a force perpendicular to the obstacle’s velocity direction $\hat{\mathbf{v}}_o$. It acts as a “steering” field to guide the drone around the obstacle, breaking potential force equilibria. Its magnitude depends on the relative angle $\theta$ between the drone-obstacle line and $\hat{\mathbf{v}}_o$.
$$ U_{io}^{aux} = \frac{1}{2} \gamma(\theta) \cdot \left( \frac{1}{d_{io}} – \frac{1}{\rho_o} \right)^2, \quad \text{for } d_{io} \leq \rho_o $$
$$ \gamma(\theta) = \eta_{aux} \cdot \left( 1 + \sin^2(\theta – \pi/2) \right) $$
The direction of the resulting force is given by a rotation matrix $\mathbf{R}(\beta)$, where $\beta$ is $\pm \pi/2$ based on the relative geometry, ensuring the force pushes the drone to the safer side of the obstacle’s path.
The total obstacle avoidance force is the sum of the negative gradients of both fields:
$$ \mathbf{F}_i^{o} = -\nabla U_{io}^{pri} – \nabla U_{io}^{aux} $$
The following table summarizes the key potential fields and their roles in the drone formation control system.
| Field Type | Mathematical Form (Key Part) | Primary Purpose | Key Feature |
|---|---|---|---|
| Leader Attraction | $ \frac{1}{2} k_{z}^{L} (z_L – z_g)^2 + … $ | Guide leader to target | Independent altitude control |
| Formation Attraction | $ \frac{1}{2} k_{p}^{F} \| (\mathbf{P}_i – \mathbf{P}_j) – \mathbf{P}_{ij}^{des} \|^2 $ | Enforce desired drone formation shape | Uses desired relative position & communication weights |
| Inter-Drone Repulsion | $ \eta_d \left( \frac{1}{ \|\mathbf{P}_i – \mathbf{P}_j\| } – \frac{1}{\rho_0} \right)^2 $ | Prevent intra-formation collisions | Short-range, active within $\rho_0$ |
| Obstacle Avoidance (Composite) | $U_{io}^{pri} + U_{io}^{aux}$ | Avoid static/moving obstacles | Velocity-modulated primary field + rotational auxiliary field to escape local minima |
Integrated Control Law
The total virtual force acting on each drone is the vector sum of all relevant potential field forces. For the leader drone $L$:
$$ \mathbf{F}_{L}^{total} = \mathbf{F}_L^{att} + \mathbf{F}_L^{d-rep} + \sum_{o} \mathbf{F}_L^{o} $$
For a follower drone $i$:
$$ \mathbf{F}_{i}^{total} = \mathbf{F}_i^{form} + \mathbf{F}_i^{d-rep} + \sum_{o} \mathbf{F}_i^{o} $$
This total force $\mathbf{F}_{i}^{total} = [F_{i,x}, F_{i,y}, F_{i,z}]^T$ is then translated into a velocity command or acceleration input for the drone’s low-level flight controller. A common integration method is:
$$ \dot{\mathbf{P}}_i^{cmd} = \dot{\mathbf{P}}_i^{current} + \frac{1}{m_i} \mathbf{F}_{i}^{total} \Delta t $$
where $\Delta t$ is the control time step. The drones’ onboard controllers track this commanded velocity to achieve cohesive and safe drone formation flight.
Simulation and Analysis
To validate the proposed enhanced 3D-APF framework, a simulation was conducted with a fleet of five drones: one leader (L) and four followers (F1-F4). The communication topology is shown below, with bidirectional links enabling distributed control.
The mission involved navigating from a scattered initial state to a target point while forming a defined shape, avoiding obstacles, and dynamically changing the formation scale mid-mission. The environment contained three static obstacles (two cylinders, one sphere) and one moving obstacle designed to induce a local minimum for traditional APF methods.
| Drone | Initial Position (x, y, z) [m] | Initial Desired Relative Position to Leader [m] | Reconfigured Desired Relative Position [m] (after t=90s) |
|---|---|---|---|
| Leader (L) | (50, 50, 20) | (0, 0, 0) | (0, 0, 0) |
| Follower 1 (F1) | (33, 37, 8) | (5, 5, 0) | (10, 10, 0) |
| Follower 2 (F2) | (38, 12, 36) | (5, -5, 0) | (10, -10, 0) |
| Follower 3 (F3) | (25, 15, 14) | (0, 10, 0) | (0, 20, 0) |
| Follower 4 (F4) | (23, 33, 22) | (0, -10, 0) | (0, -20, 0) |
The simulation results demonstrated the full capabilities of the algorithm. The drone formation successfully performed the following sequence:
- Altitude Convergence and Initial Formation: The fleet first aligned to a common altitude by approximately t=24.3s, thanks to the independent z-axis potential. The initial tight V-formation was established by t=39.3s.
- Obstacle Avoidance and Local Minima Escape: The fleet encountered obstacles starting at t=59.5s. Critically, the moving obstacle, which would trap a traditional APF, was successfully avoided. The auxiliary rotational field provided the necessary lateral steering force, allowing the leader to bypass the obstacle without stopping, as shown in the trajectory plot. All minimum distances to obstacles remained safe (>1.87m for the moving obstacle).
- Dynamic Formation Reconfiguration: At t=90s, the desired relative positions were commanded to double, transforming the tight V-formation into a wider one. The drone formation smoothly expanded and re-stabilized into the new configuration by t=118.5s, showcasing the flexibility of the desired-position-based attraction field.
- Target Convergence: The entire fleet, maintaining its formation, reached the target vicinity at t=133.1s.
The following table quantifies the collision avoidance performance during the most critical phases.
| Metric | Value | Details |
|---|---|---|
| Minimum Inter-Drone Distance | 1.62 m | Occurred between L and F2 during obstacle avoidance at t=76.19s. |
| Minimum Distance to Moving Obstacle | 1.87 m | Occurred for leader L at t=60.29s, confirming successful avoidance. |
| Formation Reconfiguration Time | ~28.5 s | Time from command (t=90s) to stable new formation (t=118.5s). |
| Altitude Alignment Time | ~24.3 s | Time for all drones to converge within a close vertical band. |
The trajectories were smooth, and velocities remained within realistic bounds (≤10 m/s), confirming the stability and practicality of the approach. The integrated use of communication weights ensured that formation forces were calculated based on available local information, validating the distributed nature of the control law.
Conclusion
This article has presented a robust and enhanced 3D Artificial Potential Field framework for multi-UAV formation control and obstacle avoidance. The proposed method directly addresses the critical shortcomings of classical APF: local minima and inflexibility. By constructing a true 3D potential field with independent altitude control, the framework ensures efficient initial fleet coordination. The incorporation of graph-based communication topology and explicit desired relative positions enables flexible and distributed drone formation control. Most significantly, the introduction of a velocity-modulated primary obstacle field coupled with a directional auxiliary rotational field provides a deterministic mechanism to escape local minima, particularly for challenging moving obstacles.
The simulation of a five-drone fleet performing complex tasks—formation assembly, obstacle navigation amidst local minima, and dynamic in-flight reconfiguration—validates the effectiveness and reliability of the algorithm. The results demonstrate safe operation with maintained minimum distances, smooth trajectories, and successful completion of all mission phases. This enhanced 3D-APF framework provides a solid foundation for deploying intelligent, resilient, and adaptive drone formations in real-world applications where environmental uncertainty and dynamic obstacles are the norm. Future work may focus on integrating wind disturbances, refining the low-level controller coupling, and experimentally validating the approach with physical drone swarms.