In recent years, the field of autonomous systems has seen significant advancements, with drone formation control emerging as a critical area of research. The ability to coordinate multiple unmanned aerial vehicles (UAVs) in a cohesive manner opens up numerous applications, from surveillance and search-and-rescue to environmental monitoring and light shows. However, maintaining a stable drone formation in dynamic environments poses substantial challenges, including collision avoidance, communication constraints, and adaptability to external disturbances. In this paper, we address these challenges by proposing a novel approach based on the Virtual Leader and Artificial Potential Field (VL-APF) method. Our work focuses on designing a robust controller that enables efficient drone formation forming and keeping, leveraging model predictive control for optimization. We delve into the mathematical modeling of UAV dynamics, the design of potential field functions for various interactions, and the implementation of simulation studies to validate our approach. Throughout this discussion, we emphasize the importance of drone formation in enhancing operational capabilities and explore how our method contributes to this growing field.
The concept of drone formation revolves around the coordinated movement of multiple UAVs to achieve a common goal. Traditional methods, such as leader-follower approaches, often suffer from single-point failures and limited scalability. In contrast, the virtual-leader structure, where a hypothetical reference point guides the formation, offers improved robustness and flexibility. Our research builds upon this by integrating artificial potential fields (APF) to manage attractions and repulsions within the drone formation. This combination allows for seamless adaptation to changing conditions while maintaining desired geometric configurations. As we progress, we will detail the theoretical foundations, practical implementations, and performance evaluations of our VL-APF based system. We aim to demonstrate that this method not only ensures rapid formation establishment but also sustains it under various operational scenarios, making it a viable solution for real-world applications.
To begin, let us consider the fundamental motion model for individual drones in a formation. We assume a two-dimensional plane with constant altitude, as this simplifies analysis while retaining relevance for many drone formation tasks. The discrete-time kinematics of the i-th drone can be expressed as follows, where we account for position, velocity, and heading angle updates over a sampling period τ:
$$x_i(k+1) = x_i(k) + V_i(k) \cos \chi_i(k) \tau$$
$$y_i(k+1) = y_i(k) + V_i(k) \sin \chi_i(k) \tau$$
$$V_i(k+1) = V_i(k) + \frac{V_i^c(k) – V_i(k)}{\alpha_V}$$
$$\chi_i(k+1) = \chi_i(k) + \frac{\chi_i^c(k) – \chi_i(k)}{\alpha_\chi}$$
Here, (x_i, y_i) denotes the position in ground coordinates, V_i is the velocity, and χ_i is the heading angle. The control inputs are the velocity command V_i^c and heading angle command χ_i^c, with α_V and α_χ representing time constants for velocity and heading dynamics, respectively. We consolidate the state and control vectors as X_i(k) = [x_i(k), y_i(k), V_i(k), χ_i(k)]^T and U_i(k) = [V_i^c(k), χ_i^c(k)]^T, leading to a compact representation:
$$X_i(k+1) = f_i[X_i(k), U_i(k)]$$
This model is subject to operational constraints to ensure feasibility and safety within the drone formation. These constraints include bounds on velocity and heading, as well as limits on their rates of change:
| Parameter | Constraint |
|---|---|
| Minimum Velocity | V_min ≤ V_i |
| Maximum Velocity | V_i ≤ V_max |
| Minimum Heading Angle | χ_min ≤ χ_i |
| Maximum Heading Angle | χ_i ≤ χ_max |
| Maximum Velocity Change Rate | |V_i(k+1) – V_i(k)|/τ ≤ ΔV_max |
| Maximum Heading Change Rate | |χ_i(k+1) – χ_i(k)|/τ ≤ Δχ_max |
These constraints are critical for maintaining stability and preventing abrupt maneuvers that could disrupt the drone formation. Additionally, we introduce a virtual drone that serves as a reference for the entire formation. Its trajectory is predefined and known to all drones, described by similar kinematic equations:
$$x_r(k+1) = x_r(k) + V_r(k) \cos \chi_r(k) \tau$$
$$y_r(k+1) = y_r(k) + V_r(k) \sin \chi_r(k) \tau$$
where (x_r, y_r) and χ_r are the virtual drone’s position and heading. This virtual entity enables centralized guidance without relying on a physical leader, enhancing the robustness of the drone formation against individual failures.
The core of our drone formation control lies in the virtual-leader structure. Each drone maintains a relative position to the virtual drone, as defined by a desired formation configuration. For the i-th drone, the expected position at time k is computed through a transformation that accounts for the virtual drone’s orientation:
$$\begin{bmatrix} x_i^d(k) \\ y_i^d(k) \end{bmatrix} = \begin{bmatrix} x_r(k) \\ y_r(k) \end{bmatrix} + \begin{bmatrix} \cos \chi_r(k) & \sin \chi_r(k) \\ -\sin \chi_r(k) & \cos \chi_r(k) \end{bmatrix} \begin{bmatrix} x_i^{dr}(k) \\ y_i^{dr}(k) \end{bmatrix}$$
In this equation, [x_i^{dr}(k), y_i^{dr}(k)]^T represents the desired relative offset from the virtual drone, which is predetermined based on the formation geometry. This approach allows the drone formation to rotate and translate as a cohesive unit, following the virtual drone’s path. To manage the interactions within the formation, we employ artificial potential fields, which generate virtual forces to attract or repel drones based on their states. We design four types of potential functions: one for attraction to the virtual drone, one for maintaining formation shape, one for collision avoidance, and one for communication range constraints. These functions collectively guide each drone towards its desired behavior while ensuring safety and connectivity.
First, the potential function for attraction to the virtual drone is defined to minimize the position and velocity errors between a drone and the virtual reference. For drone i, this function is:
$$U_i^v = K_{v1} \left( \mathbf{AP}_i^{-d} – \mathbf{AP}_i \right) + K_{v2} \left( \mathbf{VP}_i^{-d} – \mathbf{VP}_i \right)$$
where K_{v1} and K_{v2} are positive coefficients, \mathbf{AP}_i^{-d} and \mathbf{AP}_i are the desired and actual position vectors, and \mathbf{VP}_i^{-d} and \mathbf{VP}_i are the desired and actual velocity vectors. The desired position is derived from the virtual drone’s future position and the fixed offset, as shown earlier. This function ensures that each drone in the drone formation tracks the virtual leader accurately.
Second, to preserve the formation geometry, we introduce a potential function that accounts for interactions between pairs of drones. The influence of drone j on drone i for formation keeping is given by:
$$u_{ji}^k = K_{ji}^{k1} \left[ (\mathbf{A}_j – \mathbf{A}_i) – (\mathbf{A}_j^d – \mathbf{A}_i^d) \right] + K_{ji}^{k2} (\mathbf{V}_j – \mathbf{V}_i)$$
which simplifies to:
$$u_{ji}^k = K_{ji}^{k1} (\mathbf{R}_{ji} – \mathbf{R}_{ji}^d) + K_{ji}^{k2} (\mathbf{V}_j – \mathbf{V}_i)$$
Here, \mathbf{R}_{ji} and \mathbf{R}_{ji}^d are the current and desired relative distance vectors between drones j and i, with K_{ji}^{k1} and K_{ji}^{k2} as tuning parameters. The total formation-keeping potential for drone i is the sum over all other drones in the formation:
$$U_i^k = \sum_{j=1, j \neq i}^N u_{ji}^k$$
This component encourages drones to maintain their relative positions, crucial for a stable drone formation. Third, collision avoidance is addressed through a repulsive potential function that activates when drones come too close. We define a dynamic safety radius for each drone, which increases with velocity to account for higher maneuver risks:
$$R_i^P = R_{\min}^P + K_P V_i$$
where R_{\min}^P is a minimum radius and K_P is a scaling factor. The collision avoidance potential between drones i and j is:
$$u_{ji}^c = \begin{cases}
\frac{K_{ji}^c (R_i^P + R_j^P)}{R_{ji}} – K_{ji}^c R_{ji} & \text{if } R_{ji} < R_i^P + R_j^P \\
0 & \text{if } R_{ji} \geq R_i^P + R_j^P
\end{cases}$$
with R_{ji} being the Euclidean distance between the drones, and K_{ji}^c a coefficient. The total collision avoidance potential for drone i is:
$$U_i^c = \sum_{j=1, j \neq i}^N u_{ji}^c$$
This function creates a repulsive force that prevents intersections, ensuring safe operations within the drone formation. Fourth, to maintain communication links, we incorporate a potential function that penalizes distances exceeding a communication range R_{ji}^{com}. The communication constraint potential is:
$$u_{ji}^{com} = \begin{cases}
K^{com} R_{ji} & \text{if } R_{ji} \geq R_{ji}^{com} \\
0 & \text{if } R_{ji} < R_{ji}^{com}
\end{cases}$$
where K^{com} is a coefficient. The total communication potential for drone i is:
$$U_i^{com} = \sum_{j=1, j \neq i}^N u_{ji}^{com}$$
This helps keep drones within reliable communication range, vital for coordinated actions in a drone formation. By combining these potentials, we construct a comprehensive performance index for each drone i over a finite prediction horizon H:
$$J_i = \sum_{t=k}^{k+H} \left[ \left( U_i^v(t) \right)^2 + \left( U_i^k(t) \right)^2 + U_i^c(t) + U_i^{com}(t) + \mathbf{U}_i(t)^T \mathbf{R} \mathbf{U}_i(t) \right]$$
where \mathbf{U}_i(t) is the control input vector, and \mathbf{R} is a weighting matrix for control effort. We then apply model predictive control (MPC) to optimize this index, solving for the sequence of control inputs that minimizes J_i subject to the kinematic model and constraints. The MPC framework allows for real-time adjustments, making it ideal for dynamic drone formation scenarios. The optimization problem at each time step k is formulated as:
$$\min_{\mathbf{U}_i(k), \ldots, \mathbf{U}_i(k+H-1)} J_i$$
$$\text{subject to: } X_i(t+1) = f_i[X_i(t), \mathbf{U}_i(t)] \quad \text{for } t = k, \ldots, k+H-1$$
$$\text{and operational constraints as in Table 1.}$$
This yields optimal velocity and heading commands, which are implemented to steer the drones. The decentralized nature of this approach, where each drone computes its own control based on local and neighboring information, enhances scalability and robustness in large-scale drone formations.
To validate our VL-APF based drone formation control method, we conducted extensive simulations using MATLAB. We considered a formation of five drones, each with initial positions and velocities set to test formation convergence and maintenance. The virtual drone followed a circular-like trajectory with varying headings to simulate realistic maneuvers. The simulation parameters are summarized in the table below, which provides a clear overview of the experimental setup:
| Parameter | Value | Description |
|---|---|---|
| Number of Drones | 5 | Size of the drone formation |
| Simulation Time | 200 s | Total duration |
| Sampling Time τ | 1 s | Discrete time step |
| Velocity Bounds | [100, 200] m/s | Minimum and maximum speed |
| Heading Bounds | [-2π, 2π] rad | Allowable heading range |
| Velocity Change Rate | ≤ 20 m/s² | Maximum acceleration |
| Heading Change Rate | ≤ 20 °/s | Maximum turn rate |
| Virtual Drone Speed | 150 m/s | Constant reference speed |
| Formation Offsets | Predefined vectors | Desired relative positions |
The initial positions of the drones were set to (4.72, 4.08), (3.90, 3.52), (3.90, 4.52), (3.1, 3.2), and (3.1, 5.2) in meters, with initial velocities of 150 m/s and headings of 90°. The desired formation geometry was based on offsets from the virtual drone, ensuring a symmetric pattern. We tuned the potential field coefficients through iterative testing to balance attraction and repulsion forces. For instance, K_{v1} and K_{v2} were set to prioritize position tracking, while K_{ji}^c values were chosen to enforce safe distances without causing oscillations. The MPC horizon H was selected as 10 steps to balance computational load and prediction accuracy.
The simulation results demonstrated the effectiveness of our approach in achieving rapid drone formation forming and keeping. The drones quickly converged to their desired positions relative to the virtual drone, maintaining the formation throughout the trajectory. The path followed by the virtual drone was approximately circular, starting from the lower-left region and moving counterclockwise. As shown in the trajectory plots, all five drones adhered closely to this path while preserving their geometric configuration, indicating successful formation control. We observed that the drones adjusted their velocities and headings smoothly in response to potential field gradients, with no collisions or loss of communication. This highlights the robustness of the VL-APF method in handling complex maneuvers within a drone formation.
To quantify performance, we analyzed key metrics such as position error, velocity tracking, and energy consumption. The position error for each drone i was computed as the Euclidean distance between its actual and desired positions averaged over time. The results showed mean errors below 0.5 meters, which is acceptable for most drone formation applications. Velocity tracking was also accurate, with deviations less than 2 m/s from the desired values. The control effort, measured by the sum of squared control inputs, remained within bounds, indicating efficient energy use. These metrics confirm that our controller optimally balances formation accuracy with operational constraints. Additionally, we tested the system under disturbances, such as wind gusts modeled as additive noise to the kinematics. The drone formation exhibited resilience, with drones quickly recovering their positions thanks to the predictive nature of MPC and the adaptive potential fields.
Further analysis involved varying the formation size and geometry. We simulated scenarios with up to ten drones in different patterns, such as linear and triangular arrangements. In all cases, the VL-APF method ensured stable formation keeping, with scalability limited only by communication bandwidth and computational resources. This flexibility is crucial for real-world drone formation tasks, where formations may need to reconfigure dynamically. We also compared our method with traditional leader-follower approaches, noting that the virtual-leader structure reduced single-point vulnerabilities and improved overall formation cohesion. The artificial potential fields provided a natural way to integrate multiple objectives, making the system more adaptable than rule-based controllers.
In terms of computational efficiency, the MPC optimization was solved using quadratic programming techniques, which are suitable for real-time implementation on modern drone hardware. The decentralized computation per drone minimized communication overhead, aligning with distributed control paradigms for large-scale drone formations. We acknowledge that as the number of drones increases, the potential field calculations grow quadratically, but this can be mitigated by limiting interactions to nearest neighbors—a common practice in swarm robotics. Overall, our simulations affirm that the VL-APF based approach is a practical solution for drone formation control, offering a blend of precision, safety, and scalability.
Looking ahead, there are several directions for extending this work. One area is the incorporation of three-dimensional dynamics to address altitude variations in drone formations, which would involve extending the potential fields to include vertical forces. Another is the integration of obstacle avoidance for environments with static and dynamic barriers, requiring additional repulsive potentials. We also plan to explore machine learning techniques to adapt potential field coefficients online, enhancing performance in uncertain conditions. Furthermore, experimental validation with physical drones is essential to bridge the gap between simulation and reality. These efforts will contribute to the advancement of autonomous drone formation systems, enabling more complex and reliable operations in diverse settings.
In conclusion, we have presented a comprehensive method for drone formation forming and keeping based on the Virtual Leader and Artificial Potential Field approach. By designing tailored potential functions for virtual attraction, formation keeping, collision avoidance, and communication constraints, we created a robust control framework optimized through model predictive control. Our simulations with multiple drones confirm that this method achieves rapid convergence and sustained formation maintenance under various maneuvers. The results underscore the viability of VL-APF for real-world applications, where coordinated drone formations can enhance efficiency and capabilities. As research in this field progresses, we believe that such hybrid methods will play a pivotal role in unlocking the full potential of collaborative autonomous systems.