In recent years, the control of drone formation has emerged as a critical focus in the field of autonomous systems, driven by applications in surveillance, logistics, and environmental monitoring. As a researcher in this domain, I have been particularly intrigued by the challenges posed by real-world constraints such as communication delays and resource limitations. In this article, I present a comprehensive study on event-triggered control strategies for drone formation in the presence of random time delays and noise, aiming to enhance efficiency and robustness. The core idea is to reduce communication overhead while ensuring stable formation tracking, which is essential for scalable multi-drone systems. Throughout this work, the term “drone formation” will be emphasized to underscore its significance in collaborative aerial tasks.
The motivation stems from the observation that traditional continuous control methods often lead to excessive resource consumption, especially in noisy environments where delays are unpredictable. Inspired by advancements in multi-agent systems, we explore a distributed approach where each drone decides when to update its control input based on local events, rather than relying on periodic communication. This event-triggered mechanism not only conserves bandwidth but also mitigates the effects of stochastic delays modeled via Bernoulli distributions. Our goal is to derive stability conditions that guarantee convergence to a desired drone formation, even under adverse conditions.
To set the stage, consider a group of $N$ drones operating in a two-dimensional plane, where each drone is modeled as a double-integrator system. Let $p_i(t) = [x_i(t), y_i(t)]^T$ and $v_i(t) = [v_{xi}(t), v_{yi}(t)]^T$ represent the position and velocity of drone $i$, respectively. The dynamics are given by:
$$ \dot{p}_i(t) = v_i(t), $$
$$ \dot{v}_i(t) = u_i(t) \mu(t), $$
where $u_i(t)$ is the control input, and $\mu(t)$ encapsulates factors like airspeed and wind effects, assumed to be bounded. The objective is to achieve a predefined drone formation specified by offsets $d_i$ relative to a virtual center, while synchronizing velocities. Formally, we require:
$$ \lim_{t \to \infty} \| p_j(t) – d_j – (p_i(t) – d_i) \| = 0, $$
$$ \lim_{t \to \infty} \| v_j(t) – v_i(t) \| = 0, $$
for all drones $i$ and $j$ in the network. This defines the drone formation control problem, which becomes challenging when communication links suffer from random delays and noise.
In our framework, the communication topology among drones is represented by a directed graph $\mathcal{G} = (\mathcal{V}, \mathcal{E}, A)$, where $\mathcal{V}$ is the set of nodes (drones), $\mathcal{E}$ denotes edges (communication links), and $A = [a_{ij}]$ is the adjacency matrix. The Laplacian matrix $L = D – A$ captures the connectivity, with $D$ being the degree matrix. We assume the graph is strongly connected, ensuring information propagation across the drone formation. To model random delays, we introduce a Bernoulli random variable $\beta(t)$ with probability $c$ of delay occurrence, i.e., $\text{Prob}(\beta(t) = 1) = c$ and $\text{Prob}(\beta(t) = 0) = 1 – c$, where $c \in [0,1]$. The delay magnitude is denoted by $\tau > 0$.
The control input for drone $i$ is designed as a distributed event-triggered law:
$$ u_i(t) = k \sum_{j \in \mathcal{N}_i} a_{ij} \left[ \tilde{p}_j(t_k^{‘j}) – \tilde{p}_i(t_k^i) + \beta(t) (\tilde{p}_j(t_k^{‘j} – \tau) – \tilde{p}_i(t_k^i – \tau)) \right] + k \sum_{j \in \mathcal{N}_i} a_{ij} \left[ \tilde{v}_j(t_k^{‘j}) – \tilde{v}_i(t_k^i) + \beta(t) (\tilde{v}_j(t_k^{‘j} – \tau) – \tilde{v}_i(t_k^i – \tau)) \right] + k \sigma_i n_i(t), $$
where $k > 0$ is the control gain, $\tilde{p}_i(t) = p_i(t) – p_d – d_i$ and $\tilde{v}_i(t) = v_i(t) – v_d$ are formation errors relative to a virtual leader with desired trajectory $p_d$ and velocity $v_d$, $t_k^i$ and $t_k^{‘j}$ are the latest event-triggering times for drones $i$ and $j$, $\sigma_i$ is noise intensity, and $n_i(t)$ is white noise. This formulation integrates both current and delayed neighbor information, adapting to random communication delays in the drone formation.
The event-triggering condition is defined to reduce updates: for drone $i$, an event is triggered when the error $e_i(t) = [e_{pi}(t), e_{vi}(t)]^T$, with $e_{pi}(t) = \tilde{p}_i(t_k^i) – \tilde{p}_i(t)$ and $e_{vi}(t) = \tilde{v}_i(t_k^i) – \tilde{v}_i(t)$, exceeds a threshold relative to the system state $\delta_i(t) = [\tilde{p}_i(t), \tilde{v}_i(t)]^T$. Specifically, the trigger function is:
$$ f(e_i(t), \delta_i(t)) = e_i^T(t) e_i(t) – \rho \delta_i^T(t) \delta_i(t) > 0, $$
where $\rho > 0$ is a design parameter. This ensures that control updates occur only when necessary, conserving resources in the drone formation network.
To analyze stability, we transform the system into a compact form. Let $\delta(t) = [\tilde{p}^T(t), \tilde{v}^T(t)]^T$ and $e(t) = [e_p^T(t), e_v^T(t)]^T$, where $\tilde{p}(t)$ and $\tilde{v}(t)$ stack all drones’ errors. The closed-loop dynamics can be written as a stochastic differential equation:
$$ d\delta(t) = [A \delta(t) + \beta(t) B \delta(t – \tau) + B e(t) + \beta(t) B e(t – \tau)] dt + C dw(t), $$
with matrices $A, B, C$ defined appropriately based on the Laplacian $L$ and control gain. Here, $w(t)$ is a Wiener process representing noise. We then construct a Lyapunov-Krasovskii functional to assess mean-square stability of the drone formation.
Consider the functional:
$$ V(t) = \delta^T(t) P \delta(t) + \int_{t-\tau}^{t} \delta^T(s) Q \delta(s) ds, $$
where $P$ and $Q$ are positive definite matrices. Applying Itô’s lemma and taking expectations, we derive sufficient conditions for stability in terms of linear matrix inequalities (LMIs). Specifically, after lengthy derivations, we obtain the following LMIs that ensure the drone formation converges:
$$ \begin{bmatrix} -Q & \epsilon_1 I & \epsilon_2 c I & \epsilon_4 c \rho I \\ * & -\epsilon_1 I & 0 & 0 \\ * & * & -\epsilon_2 c I & 0 \\ * & * & * & -\epsilon_4 c \rho I \end{bmatrix} < 0, $$
$$ \begin{bmatrix} -Q & P A & \epsilon_1 I & c P B & P B & \epsilon_3 \rho I & c P B \\ * & -\epsilon_1 I & 0 & 0 & 0 & 0 & 0 \\ * & * & -\epsilon_1 I & 0 & 0 & 0 & 0 \\ * & * & * & -\epsilon_2 I & 0 & 0 & 0 \\ * & * & * & * & -\epsilon_3 \rho I & 0 & 0 \\ * & * & * & * & * & -\epsilon_3 \rho I & 0 \\ * & * & * & * & * & * & -\epsilon_4 I \end{bmatrix} < 0, $$
where $\epsilon_1, \epsilon_2, \epsilon_3, \epsilon_4$ are positive scalars, and $*$ denotes symmetric entries. These LMIs can be solved numerically to find feasible matrices $P$ and $Q$, guaranteeing that the drone formation system is stable under event-triggered control with random delays. Moreover, we prove the absence of Zeno behavior, ensuring a minimum inter-event time $h = \frac{\sqrt{\rho}}{(1 + \sqrt{\rho}) \|M\|}$, where $\|M\|$ is a matrix norm bound, thus preventing infinite triggering in the drone formation.
To validate our approach, we conducted simulation experiments with a quadrotor drone formation of four units. The communication topology is a directed cycle, as shown in the graph representation, with adjacency matrix entries $a_{ij} = 1$ for specific links. Key parameters are summarized in Table 1, which includes initial positions, velocities, and drone masses. The desired formation offsets $d_i$ are set to create a geometric pattern, and the virtual leader moves at a constant velocity $v_d$. We assume wind effects with speed $v_w = 2 \, \text{m/s}$ and direction $\theta_w = \pi/6$, incorporated into $\mu(t)$.
| Drone ID | Mass (kg) | Initial Position (m) | Initial Velocity (m/s) | Heading Angle (rad) | Noise Intensity $\sigma_i$ |
|---|---|---|---|---|---|
| 1 | 2.5 | (6.6, 4.6) | (4.2, -2.7) | 1.012 | 0.1 |
| 2 | 2.5 | (7.0, 5.7) | (-3.3, 1.8) | 1.047 | 0.1 |
| 3 | 2.5 | (4.2, 5.0) | (2.5, -1.5) | 0.977 | 0.1 |
| 4 | 2.5 | (4.8, 5.7) | (-1.2, 3.0) | 1.082 | 0.1 |
In the simulations, we set control gain $k = 1.2$, event-triggering parameter $\rho = 0.8$, delay probability $c = 0.6$, and delay magnitude $\tau = 0.1 \, \text{s}$. Solving the LMIs yields feasible solutions with $\epsilon_1 = 0.64$, $\epsilon_2 = 1.33$, $\epsilon_3 = 0.81$, $\epsilon_4 = 1.40$. The results demonstrate that the drone formation achieves the desired configuration within approximately 6.8 seconds, as shown in Figure 1 (though not referenced directly, described textually). The position and velocity errors converge to zero, confirming stability. For comparison, we tested scenarios with no delays ($c = 0$) and deterministic delays, highlighting that random delays can sometimes accelerate convergence due to stochastic effects, but our method consistently ensures robust drone formation control.
A key aspect is the efficiency of event-triggering. Over a 10-second simulation, the average number of control updates per drone was around 40, significantly lower than traditional time-triggered methods that require updates at every time step (e.g., 200 updates). This reduction in communication underscores the resource-saving benefits of our approach for large-scale drone formation systems. To quantify performance, we define a formation error metric $E(t) = \sum_{i=1}^N \| \tilde{p}_i(t) \|^2 + \| \tilde{v}_i(t) \|^2$, which decays exponentially as per the stability analysis. The error dynamics can be approximated by:
$$ \frac{dE(t)}{dt} \leq -\lambda E(t) + q, $$
where $\lambda > 0$ and $q$ is a constant related to noise, leading to bounded steady-state error. This aligns with the mean-square stability guarantees derived from the Lyapunov analysis.
Further insights can be gained by examining the impact of delay probability $c$ on convergence time. Table 2 summarizes results for different $c$ values, showing that higher probabilities (up to a point) may reduce convergence time due to increased stochastic excitation, but excessive delays can degrade performance. This trade-off is crucial for designing resilient drone formation protocols.
| Delay Probability $c$ | Convergence Time (s) | Average Trigger Events per Drone | Final Formation Error $E(\infty)$ |
|---|---|---|---|
| 0.0 | 19.1 | 35 | 0.05 |
| 0.2 | 11.0 | 38 | 0.06 |
| 0.6 | 6.8 | 41 | 0.07 |
| 0.8 | 7.5 | 43 | 0.08 |
The mathematical foundation of our controller relies on robust stochastic stability theory. To elaborate, the closed-loop system can be viewed as a switched stochastic delay system, where modes correspond to delay occurrences. Using a common Lyapunov function, we ensure stability across all modes. The event-triggering condition introduces a bounded error $e(t)$, which is treated as a disturbance in the LMI formulation. The scalability of this approach to larger drone formations is promising, as the LMIs depend only on local connectivity and can be decomposed via distributed optimization techniques.
In practice, implementing event-triggered control for drone formation requires onboard computation of trigger conditions. We propose an algorithm where each drone monitors its own state and neighbor information, updating the control input only when $f(e_i(t), \delta_i(t)) > 0$. This decentralized logic reduces communication bursts and enhances scalability. For instance, in a formation of $N$ drones, the total communication load scales linearly with event frequency rather than quadratically with network size, as in continuous schemes.
To further illustrate the controller’s behavior, consider the error dynamics in the frequency domain. Taking Laplace transforms of the linearized system around the equilibrium yields transfer functions that characterize response to disturbances. The inclusion of random delays modifies these transfer functions, introducing phase shifts that our controller compensates via the gain $k$. The stability margins can be analyzed using Nyquist criteria, but the stochastic nature complicates this; hence, our time-domain Lyapunov approach is more suitable.
Another important consideration is noise rejection. The term $\sigma_i n_i(t)$ in the controller represents measurement or environmental noise. Our analysis shows that the mean-square error remains bounded, with bound proportional to $\sum_i \sigma_i^2$. This implies that the drone formation can tolerate moderate noise levels without destabilizing, which is vital for outdoor operations. In simulations, we injected Gaussian white noise with variance $0.01$, and the formation maintained cohesion, demonstrating robustness.
The event-triggering mechanism also has implications for energy consumption. By reducing control updates, drones save battery power otherwise spent on communication and computation. This is especially beneficial for long-duration missions where drone formation must be maintained over hours. We estimate energy savings of up to 30% compared to time-triggered baselines, based on simplified power models where each transmission costs a fixed energy amount.
Extensions of this work could involve adaptive event-triggering thresholds, where $\rho$ adjusts dynamically based on network conditions. For example, in high-noise environments, $\rho$ could increase to prevent excessive triggering, while in calm conditions, it could decrease for finer control. This adaptability would enhance the autonomy of drone formation systems. Additionally, incorporating obstacle avoidance into the formation control law is a natural next step, requiring modifications to the Lyapunov function to include potential fields.
From a theoretical perspective, our stability conditions are conservative due to the use of fixed matrices $P$ and $Q$. Less conservative LMIs could be derived using delay-dependent techniques or parameter-dependent Lyapunov functions, potentially allowing for larger delays or higher noise. However, the current formulation offers a tractable solution for real-time implementation, which is a priority for practical drone formation applications.
In conclusion, this article has presented a robust event-triggered control strategy for drone formation in the presence of random time delays and noise. By leveraging stochastic stability theory and LMI-based design, we ensured convergence to desired formations while significantly reducing communication overhead. Simulation results validated the effectiveness, showing fast convergence and low trigger rates. The emphasis on “drone formation” throughout highlights the applicability to collaborative aerial systems. Future work will focus on experimental validation with physical drones and integration with higher-level planning algorithms. Overall, this approach paves the way for efficient and resilient multi-drone operations in dynamic environments.
To summarize key equations, the overall system dynamics are encapsulated in:
$$ d\delta(t) = [A \delta(t) + \beta(t) B \delta(t – \tau) + B e(t) + \beta(t) B e(t – \tau)] dt + C dw(t), $$
with stability guaranteed if the LMIs hold. The event-triggering condition ensures efficient resource use, and the minimum inter-event time prevents Zeno behavior. This framework provides a solid foundation for advanced drone formation control, balancing performance and practicality in real-world scenarios.