Rapid Formation Assembly and Distributed Consensus Control for Fixed-Wing Drone Swarms

We address the challenges of formation assembly efficiency and geometric precision in small fixed-wing drone swarms. Our approach integrates a loose formation rapid assembly stage based on Dubins path planning with a fine formation control stage using a distributed consensus algorithm. This two-stage strategy enables fast and efficient assembly, stable formation keeping, and smooth formation transformation. We model the cooperative path planning problem using Dubins curves and solve it with a genetic algorithm (GA) to obtain near-optimal loose assembly trajectories. For the residual state errors after assembly, we design a virtual-leader-based consensus control law that generates accurate geometric formations while ensuring inter-drone safety distances. Simulations under oblique-line and V-shaped formations validate the effectiveness of the proposed method in terms of assembly time efficiency and formation control accuracy.

1. Introduction

Unmanned aerial vehicle (UAV) swarms have attracted significant attention in both military and civilian domains due to their extended operational range, high efficiency, and robustness. Among various platforms, fixed-wing drones offer advantages in endurance, speed, and payload capacity. However, their dynamics are constrained by minimum turning radius, limited thrust-to-weight ratio, and narrow speed envelope, making formation control more challenging than for rotorcraft. The problem of formation assembly can be approached via two main methods: speed-based coordination, which adjusts individual velocities to synchronize arrival times, and multi-UAV cooperative path planning, which modifies trajectories while maintaining constant cruise speed. The latter is more suitable for fixed-wing drones because large speed variations are often infeasible. Dubins curves provide a geometric way to plan shortest paths for constant-speed vehicles with bounded curvature, making them ideal for fixed-wing drone swarms.

After assembly, maintaining precise geometry requires a distributed control law. Consensus-based algorithms, particularly those using virtual leaders, allow each drone to locally interact with neighbors and follow a virtual reference, achieving formation stability even under dynamic conditions. In this work, we combine Dubins-based loose assembly with virtual-leader consensus control to achieve both rapid convergence and high-precision formation keeping. We also compare analytical solutions with GA-based numerical optimization for the Dubins planner. The rest of the paper is organized as follows. Section 2 describes the problem and control strategy. Section 3 presents the mathematical models, including fixed-wing drone dynamics, Dubins path planning, and consensus control. Section 4 provides simulation results and analysis. Section 5 concludes the paper.

2. Problem Description and Control Strategy

We consider a swarm of small fixed-wing drones launched sequentially from the same ground platform at fixed time intervals. After climbing to a constant altitude, they fly at a nearly constant speed within a limited range. The objective is to quickly assemble into a prescribed geometric formation and then maintain or transform the formation while avoiding collisions. The entire process is divided into two phases: loose formation assembly and fine formation control. In the loose phase, each drone follows a Dubins path (composed of circular arcs and straight segments) that coordinates arrival times without requiring communication. In the fine phase, a distributed consensus controller uses inter-drone communication to correct residual errors and achieve a rigid geometric configuration. The strategy is illustrated conceptually in the system architecture below:

3. Mathematical Models

3.1 Fixed-Wing Drone Dynamics

We model each fixed-wing drone as a rigid body moving in three-dimensional space, but we focus on steady, level flight at constant altitude, reducing the problem to two dimensions. Under the assumptions of no wind, zero sideslip, and coordinated turns, the kinematic and simplified dynamic equations are:

$$
\dot{x}_i = V_{gi} \cos \chi_i \cos \gamma_i
$$
$$
\dot{y}_i = V_{gi} \sin \chi_i \cos \gamma_i
$$
$$
\dot{z}_i = -V_{gi} \sin \gamma_i
$$
$$
\dot{V}_{gi} = \frac{T_{hi} – D_{gi}}{m_i} – g \sin \gamma_i
$$
$$
\dot{\gamma}_i = \frac{L_{fi} \cos \phi_i – m_i g \cos \gamma_i}{m_i V_{gi}}
$$
$$
\dot{\chi}_i = \frac{L_{fi} \sin \phi_i}{m_i V_{gi} \cos \gamma_i}
$$

where for drone \(i\), \((x_i, y_i, z_i)\) is the position, \(V_{gi}\) is the ground speed, \(\gamma_i\) the flight path angle, \(\chi_i\) the course angle, \(T_{hi}\) the thrust, \(D_{gi}\) the drag, \(L_{fi}\) the lift, \(m_i\) the mass, \(g\) gravity, and \(\phi_i\) the roll angle. For level flight (\(\gamma_i = 0, \dot{\gamma}_i = 0\)), the dynamics simplify to a second-order integrator model in the horizontal plane:

$$
\dot{\xi}_i = \zeta_i, \quad \dot{\zeta}_i = u_i
$$

where \(\xi_i = [x_i, y_i]^T\), \(\zeta_i = [v_{xi}, v_{yi}]^T\) is the velocity vector, and \(u_i\) is the acceleration control input. The actual control inputs (thrust, roll, etc.) can be recovered from \(u_i\) using inversion formulas. We impose constraints on speed range, acceleration magnitude, and turning rate:

$$
V_{\min} \le V_{gi} \le V_{\max}, \quad |a_i| \le a_{\max}, \quad |\dot{\psi}_i| \le \dot{\psi}_{\max}
$$

These constraints lead to a minimum turning radius \(r_{\min} = V_{gi}^2 / (g \tan \phi_{\max})\). In the following, we assume a constant cruise speed \(V_0\) for path planning, which is typical for small fixed-wing drones.

3.2 Dubins Path Planning for Rapid Assembly

We consider \(n\) fixed-wing drones launched sequentially with time interval \(\Delta t\) from the same point (0,0) with initial heading \(\chi_i(0) = \pi/2\). All drones fly at the same constant speed \(V_0\). The desired terminal formation is defined by relative positions \((\Delta x_i, \Delta y_i)\) with respect to a reference drone. Each drone follows a Dubins path of type CLC (circle-line-circle) with two circular arcs of the same curvature radius \(r_i\) (equal for both arcs) and one straight segment of length \(d_i\). The total path length for drone \(i\) is:

$$
L_i = r_i \phi_{i1} + d_i + r_i \phi_{i2}
$$

where \(\phi_{i1}\) and \(\phi_{i2}\) are the turning angles (both equal to \(\phi_i\) for simplicity). The terminal heading is \(\chi_f = \pi/2\). The path must satisfy geometric constraints in \(x\) and \(y\) directions relative to the first drone (\(i=1\)):

$$
r_1 (1 – \cos\phi_1) + d_1 \sin\phi_1 + r_1 (1 – \cos\phi_1) = 0 + \Delta x_1
$$
$$
r_1 \sin\phi_1 + d_1 \cos\phi_1 + r_1 \sin\phi_1 = V_0 t_1 + \Delta y_1
$$

For other drones \(i\), we add the initial time offset and position offset:

$$
r_i (1 – \cos\phi_i) + d_i \sin\phi_i + r_i (1 – \cos\phi_i) = (i-1)\Delta x + \Delta x_i
$$
$$
r_i \sin\phi_i + d_i \cos\phi_i + r_i \sin\phi_i = V_0 (t_1 + (i-1)\Delta t) + \Delta y_i
$$

The unknowns are \((r_i, \phi_i, d_i)\) for each drone and the reference time \(t_1\). We solve this system using two methods:

  • Analytical method: Choose one drone as reference, assume its variables known, solve for \(t_1\), then compute all others exactly. This gives minimal assembly time but requires offline computation.
  • Genetic algorithm (GA): Formulate a nonlinear constrained optimization to minimize total path length or assembly time, treating all variables as unknowns within bounds. GA provides a suboptimal solution near real-time, suitable for online re-planning.

The objective function for GA is:

$$
F(X) = \sum_{i=1}^n \left( L_i + \alpha_x|\tilde{x}_i| + \alpha_y|\tilde{y}_i| \right)
$$

where \(\tilde{x}_i, \tilde{y}_i\) are the deviations from the target relative positions, and \(\alpha_x, \alpha_y\) are penalty weights. The decision vector \(X\) includes all \(r_i, \phi_i, d_i\) and \(t_1\). Constraints include bounds on radius (\(r_{\min} \le r_i \le r_{\max}\)), turning angle (\(0 < \phi_i < \pi\)), and straight segment length (\(d_i \ge 0\)). We also enforce minimum inter-drone distance \(\rho_0\) during the whole path. If collisions are detected, the solution is discarded and another candidate is selected. Figure below shows the GA workflow.

The comparison between analytical and GA solutions for a three-drone oblique-line formation (spacing 100 m in both axes) is summarized in Table 1.

Table 1: Comparison of Analytical and GA Solutions for Three Drone Assembly
Parameter Analytical GA (suboptimal)
Radii \(r_i\) (m) [200.0, 221.8, 212.8] [628.0, 653.5, 677.7]
Turning angles \(\phi_i\) (rad) [1.745, 1.579, 1.116] [1.67, 1.58, 1.43]
Straight segments \(d_i\) (m) [300.0, 217.6, 363.0] [452, 412, 453]
Assembly time \(t_1\) (s) 24.95 63.8
Final position error (m) <0.1 <10

The analytical method yields shorter assembly time but requires a known reference and offline search. GA provides feasible, albeit longer, paths with acceptable errors for the subsequent fine control phase. The GA approach is more flexible for dynamic scenarios, such as when launch intervals or target positions change.

3.3 Distributed Consensus Control for Fine Formation

After the loose assembly, each fixed-wing drone has a residual position error due to the suboptimal paths. We now apply a virtual-leader-based consensus controller to achieve precise geometric formation and maintain it during flight. The communication topology is a fixed undirected graph where each drone can exchange state information with its neighbors and also receive the virtual leader’s state. Let \(\xi_0(t) = [x_0(t), y_0(t)]^T\) and \(\dot{\xi}_0 = \zeta_0\) be the virtual leader’s position and velocity (constant velocity in our case). The desired formation is defined by offset vectors \(d_i\) for each drone relative to the leader, and \(d_{ij} = d_j – d_i\) for relative offsets between drones. The control input for drone \(i\) is:

$$
\begin{aligned}
u_i = &- \alpha \left[ (\xi_i – \xi_0 – d_i) + \beta (\zeta_i – \zeta_0) \right] \\
&- \sum_{j \in \mathcal{N}_i} a_{ij} \left[ (\xi_i – \xi_j – d_{ij}) + \beta (\zeta_i – \zeta_j) \right]
\end{aligned}
$$

where \(\alpha > 0\) is the leader-follower gain (we take \(\alpha=1\)), \(\beta > 0\) is the velocity damping coefficient (\(\beta=2\)), \(a_{ij}=1\) if drone \(i\) can communicate with \(j\), and \(\mathcal{N}_i\) is the neighbor set. This control law drives the formation error to zero exponentially under fixed topology with no delays. To avoid collisions, we augment the control with a repulsive potential field when inter-drone distance falls below a safety threshold \(\rho_0=25\) m. The overall control becomes:

$$
u_i = u_i^{\text{consensus}} + u_i^{\text{repulsion}}
$$

The repulsion term is designed as a gradient of a potential function:

$$
u_i^{\text{repulsion}} = – \sum_{j \neq i} \nabla_{\xi_i} \frac{k}{(\|\xi_i – \xi_j\| – \rho_0)^2}
$$

where \(k\) is a positive gain. This ensures that drones maintain safe separation while tracking the formation. The combination of consensus and repulsion allows both precise geometry and safety.

4. Simulation Results

4.1 Simulation Setup

We simulate three identical fixed-wing drones with parameters listed in Table 2. The drones are launched at 2 s intervals from (0,0) with initial heading \(\pi/2\). The desired formation after assembly is an oblique line with longitudinal and lateral spacing of 100 m. After holding the formation for a while, they transform to a V-shaped (equilateral triangle) with sides 100 m.

Table 2: Simulation Parameters
Parameter Value
Minimum speed \(V_{\min}\) 35 m/s
Maximum speed \(V_{\max}\) 45 m/s
Cruise speed \(V_0\) 40 m/s
Maximum acceleration \(a_{\max}\) 0.5 m/s²
Minimum turning radius \(r_{\min}\) 200 m
Launch interval \(\Delta t\) 2 s
Starting position \((x_i, y_i)\) (0,0)
Starting heading \(\chi_i(0)\) π/2 rad
Desired lateral/longitudinal spacing (100 m, 100 m)
Safety distance \(\rho_0\) 25 m
Wingspan <5 m

4.2 Loose Assembly Results

We applied the GA-based Dubins planner to compute paths for the three fixed-wing drones. After 150 generations, a feasible solution without collisions was found. The resulting paths are shown in Figure 4 of the original paper (not reproduced here). The GA solution gave an assembly time of 63.8 s with final position errors below 10 m. The analytical solution (if computed offline) would yield a shorter time of 24.95 s, but requires precomputed reference and is not adaptable to real-time changes. In practice, we use the GA solution for online planning and rely on the consensus controller to correct the errors.

4.3 Fine Formation Control Results

Upon completion of the Dubins paths, the virtual-leader-based consensus controller is activated. The virtual leader moves at constant speed \(V_0\) along the x-axis: \(x_0 = V_0 t\), \(y_0 = 0\). The initial errors from the path planner are within 10 m. Within about 10 s, the formation converges to a precise oblique line with position errors less than \(10^{-3}\) m. Figure 5(a) in the original paper shows the trajectory and final geometry. The drones then maintain the formation for another 100 s. At \(t=150\) s, a command to transform into a V-shaped formation (equilateral triangle) is given. The consensus controller adaptively adjusts the offsets \(d_i\) to the new geometry. The transformation completes in about 350 s, as shown in Figure 5(b). During the whole process, the minimum inter-drone distance never falls below 25 m, as verified by the local plot in Figure 5(c). The repulsive potential successfully prevents collisions even during the transient when drones cross paths.

4.4 Discussion

The results demonstrate that the two-stage strategy works effectively for small fixed-wing drone swarms. The Dubins-based loose assembly provides a fast way to bring drones into a neighborhood of the desired formation without requiring extensive communication. The subsequent consensus controller refines the geometry to high precision. The transformation time (350 s) is relatively long due to the limited acceleration capability of fixed-wing drones; however this is acceptable for many surveillance or patrolling missions. Future work should incorporate dynamic obstacle avoidance using velocity obstacle methods, and consider communication delays, sensor noise, and input saturation. Additionally, tight formation flight with aerodynamic coupling would require more advanced consensus protocols.

5. Conclusion

We presented a practical method for rapid formation assembly and high-precision formation control of small fixed-wing drones. The key is to combine Dubins path planning with a virtual-leader consensus algorithm. The GA-based Dubins planner yields feasible suboptimal trajectories that can be computed online, while the analytical method provides a benchmark for minimal assembly time. The consensus controller ensures that residual errors are eliminated and formations are maintained or transformed safely. Simulation results confirm the effectiveness of the approach. Future research will focus on extending the method to scenarios with obstacles, time-varying communication graphs, and heterogeneous drone types.

Scroll to Top