The night sky is no longer just a canvas for stars. Today, it becomes a dynamic, luminous stage for stories, logos, and breathtaking art, painted by swarms of synchronized drones. This is the world of formation drone light shows. My professional journey has been deeply intertwined with the evolution of this spectacle, moving from simple, pre-programmed sequences to complex, adaptive, and precisely timed aerial ballets. The core challenge, and the focus of my work, has always been the same: how to command a fleet of unmanned aerial vehicles (UAVs) to form and transition between intricate shapes with absolute temporal and spatial precision, despite wind, turbulence, and system uncertainties. It is a beautiful fusion of control theory, computer science, and artistic vision. The quest for perfection led my team and me beyond conventional methods, pushing us to develop control strategies where timing is not just estimated but guaranteed. This is an account of that technical pursuit.
At the heart of any formation drone light show lies a complex multi-agent control problem. Each drone is a nonlinear, underactuated system. For a quadrotor, the standard workhorse for such shows, the dynamics can be described by Newton-Euler equations. While simplified, these equations capture the essential challenge. The translational motion $(x, y, z)$ is governed by the total thrust and the vehicle’s orientation (roll $\phi$, pitch $\theta$, yaw $\psi$), while the rotational dynamics depend on the moments generated. A common model for the $i$-th drone is:
$$
\begin{align*}
m_i \ddot{x}_i &= U_{1,i} (\cos\phi_i \sin\theta_i \cos\psi_i + \sin\phi_i \sin\psi_i) + d_{x,i} \\
m_i \ddot{y}_i &= U_{1,i} (\cos\phi_i \sin\theta_i \sin\psi_i – \cos\phi_i \sin\psi_i) + d_{y,i} \\
m_i \ddot{z}_i &= U_{1,i} (\cos\theta_i \cos\phi_i) – m_i g + d_{z,i} \\
I_{x,i} \ddot{\phi}_i &= l_i U_{2,i} + d_{\phi,i} \\
I_{y,i} \ddot{\theta}_i &= l_i U_{3,i} + d_{\theta,i} \\
I_{z,i} \ddot{\psi}_i &= l_i U_{4,i} + d_{\psi,i}
\end{align*}
$$
Here, $m_i$ is mass, $g$ is gravity, $I$ terms are moments of inertia, $l_i$ is the arm length, $U_{1,i}$ is the total thrust command, and $U_{2,i}, U_{3,i}, U_{4,i}$ are the moment commands for attitude control. The critical terms $d_{*,i}$ represent lumped uncertainties and external disturbances—the nemesis of a perfect show, encompassing wind gusts, model inaccuracies, and actuator imperfections.
The control architecture is typically split into an inner attitude loop and an outer position loop. For a formation drone light show, the position controller is paramount. It must calculate the required thrust $U_{1,i}$ and desired attitude angles ($\phi_{d,i}, \theta_{d,i}$) to make the drone follow a specific point in the evolving formation pattern. The yaw $\psi_{d,i}$ is often controlled separately for visual effect. The formation task translates into each drone tracking a time-varying reference point $Q_{d,i}(t)$, which is offset from a virtual leader’s trajectory $Q_0(t)$ that defines the overall shape’s motion: $Q_{d,i}(t) = Q_0(t) – \sigma_i$, where $\sigma_i$ is the fixed formation offset for drone $i$.
Early shows used linear PID or backstepping controllers. They worked, but convergence was asymptotic. The formation would “mostly” be there after a few seconds, but precise timing for choreographed transitions was elusive. We then explored finite-time control, where convergence time depends on initial conditions—unacceptable for a show starting from arbitrary park positions. Fixed-time control was a breakthrough, offering a bounded convergence time independent of initial states. Yet, its upper bound is a complex function of controller parameters. Tuning to achieve a specific, desired convergence time (e.g., 500ms for a snap transition) was more art than science. This led us to the concept of predefined-time stability. Here, the convergence time $T_c$ can be set directly as an explicit, tunable parameter. This was the game-changer for a reliable formation drone light show.
A fundamental lemma underpins our approach: For a system with a Lyapunov function $V(\xi)$, if its derivative satisfies
$$
\dot{V}(\xi) \leq -\frac{1}{p T_c} \exp\left(V^p(\xi)\right) V^{1-p}(\xi)
$$
for $p \in (0,1)$ and $T_c > 0$, then the system’s origin is globally stable within the predefined time $T_c$. This explicit $T_c$ is what we directly plug into our controllers.
The first major subsystem we redesigned was the state estimator. In a real formation drone light show, drones communicate over a network. Not every drone receives the virtual leader’s trajectory directly. We need a distributed way for each drone to estimate $Q_{d,i}(t)$ and its derivative $v_{d,i}(t)$. We designed a predefined-time sliding mode estimator. For drone $i$, let $\hat{Q}_{d,i}$ and $\hat{v}_{d,i}$ be its estimates. Defining the estimation errors $\tilde{Q}_{d,i} = \hat{Q}_{d,i} – Q_{d,i}$ and $\tilde{v}_{d,i} = \hat{v}_{d,i} – v_{d,i}$, the estimator dynamics are crafted to drive these errors to zero in a predefined time $T_{c,est}$.
The collective error dynamics for a fleet of $n$ drones, under a communication graph with Laplacian $L$ and leader-access matrix $B$, can be written for the stacked vectors $\tilde{\mathbf{Q}}$ and $\tilde{\mathbf{v}}$:
$$
\begin{align*}
\dot{\tilde{\mathbf{Q}}} &= \tilde{\mathbf{v}} – \alpha(t) \lceil (L+B) \otimes I_3 \, \tilde{\mathbf{Q}} \rfloor^{1-2p} – \kappa_1 \text{sign}((L+B) \otimes I_3 \, \tilde{\mathbf{Q}}) \\
\dot{\tilde{\mathbf{v}}} &= -\beta(t) \lceil (L+B) \otimes I_3 \, \tilde{\mathbf{v}} \rfloor^{1-2p} – \delta \text{sign}((L+B) \otimes I_3 \, \tilde{\mathbf{v}}) – \mathbf{1}_n \otimes \dot{v}_0(t)
\end{align*}
$$
The gains $\alpha(t)$ and $\beta(t)$ are the key innovation. They are not constants but time-varying functions that incorporate the predefined convergence time:
$$
\alpha(t) = \frac{(\epsilon \varsigma)^{1-2p} \{ 2\lambda_{\min}[(L+B)\otimes I_3] \}^{1-p}}{p T_{c,2}} \exp\left( \left\{ \frac{1}{2} \tilde{\mathbf{Q}}^T [(L+B)\otimes I_3] \tilde{\mathbf{Q}} \right\}^p \right)
$$
A similar structure defines $\beta(t)$ with $T_{c,1}$. By constructing a Lyapunov function $V_{est} = \frac{1}{2} \tilde{\mathbf{Q}}^T [(L+B)\otimes I_3] \tilde{\mathbf{Q}}$, we can prove $\dot{V}_{est} \leq -\frac{1}{p T_{c,2}} \exp(V_{est}^p) V_{est}^{1-p}$, guaranteeing estimation convergence before $T_{c,est} = \max(T_{c,1}, T_{c,2})$. This means every drone knows exactly where it needs to be in the formation pattern within a known, designer-set time after show start or a pattern transition command.
With precise estimates of its target position $\hat{Q}_{d,i}$ and velocity $\hat{v}_{d,i}$ from the estimator, each drone must now track this reference. We employ a predefined-time backstepping controller with an adaptive law to kill disturbances. Consider the position tracking error $e_{p,i} = p_i – \hat{Q}_{d,i}$ and the velocity error $e_{v,i} = v_i – \hat{v}_{d,i}$, where $p_i = [x_i, y_i, z_i]^T$. The control objective is to drive $(e_{p,i}, e_{v,i})$ to zero in predefined times $T_{c,pos}$ and $T_{c,vel}$.
We define a predefined-time stabilization function $\Phi_{k,s}(\xi; T_c)$ for a vector $\xi$:
$$
\Phi_{k,s}(\xi; T_c) = \frac{1}{T_c k s} \exp\left( \|\xi\|^{k s} \right) \|\xi\|^{1 – k s}
$$
where $k \geq 1$, $0 < s \leq 1/k$, and $T_c > 0$ is the predefined time. This function is central to our controller.
For the position loop, we treat $e_{v,i}$ as a virtual control. We first design a desired velocity $v_{d,i}^* = -\Phi_{s_1, q_1}(e_{p,i}; T_{c1}) + \dot{\hat{Q}}_{d,i}$ to stabilize $e_{p,i}$ in time $T_{c1}$. Then, we design the actual thrust vector (virtual control input) to stabilize the velocity error $e_{v,i} = v_i – v_{d,i}^*$ in time $T_{c2}$. The resulting controller for the translational dynamics, after incorporating adaptive disturbance cancellation $\hat{d}_{p,i}$, takes the form:
$$
\mathbf{U}_{1,cmd,i} = m_i \left( g \mathbf{e}_z + \ddot{\hat{Q}}_{d,i} – \hat{d}_{p,i} – \frac{s_1}{s_2} \frac{e_{v,i}}{\|e_{v,i}\|^{2-s_2}} \frac{e_{p,i}^T}{\|e_{p,i}\|^{2-s_1}} e_{v,i} – \Phi_{s_2, q_2}(e_{v,i}; T_{c2}) \right)
$$
The total thrust $U_{1,i} = \|\mathbf{U}_{1,cmd,i}\|$ and the desired roll/pitch angles are derived from its orientation. The adaptive law for the disturbance estimate is $\dot{\hat{d}}_{p,i} = \gamma \, s_2 \, \frac{e_{v,i}}{\|e_{v,i}\|^{2-s_2}}$, where $\gamma$ is a positive adaptation gain. This law continuously estimates and compensates for wind and other disturbances in real-time, which is absolutely critical for maintaining a rock-solid formation drone light show outdoors.
Similarly, a predefined-time adaptive backstepping controller is designed for the attitude loop to track the desired $\phi_{d,i}, \theta_{d,i}, \psi_{d,i}$ generated by the position controller. The control moments $U_{2,i}, U_{3,i}, U_{4,i}$ are calculated to drive attitude and angular rate errors to zero in predefined times $T_{c,att}$ and $T_{c,rate}$.
The performance of different control strategies for a formation drone light show can be summarized in the following table:
| Control Method | Convergence Time | Dependence on Initial State | Robustness to Disturbances | Tuning for Desired Speed |
|---|---|---|---|---|
| Linear (PID) | Asymptotic (theoretically infinite) | Very High | Low | Very Difficult |
| Finite-Time | Bounded, but depends on initial error | Very High | Medium | Difficult |
| Fixed-Time | Bounded, independent of initial error | None | High | Complex (indirect) |
| Predefined-Time Adaptive (Our Approach) | Bounded, can be set directly by designer | None | Very High (via adaptation) | Straightforward (direct parameter $T_c$) |
To validate our approach, we simulate a squadron of drones performing a complex show. The parameters for a typical show drone are listed below:
| Parameter | Symbol | Value |
|---|---|---|
| Mass | $m_i$ | 1.2 kg |
| Arm Length | $l_i$ | 0.25 m |
| Moment of Inertia (X/Y) | $I_{x,i}, I_{y,i}$ | 0.008 kg·m² |
| Moment of Inertia (Z) | $I_{z,i}$ | 0.015 kg·m² |
| Predefined Formation Time | $T_{c,est}$ | 0.8 s |
| Predefined Position Tracking Time | $T_{c,pos}$ | 1.5 s |
| Disturbance Adaptation Gain | $\gamma$ | 5.0 |
In simulation, we command the fleet to form a rotating “star” pattern while following a sinusoidal leader trajectory, under significant simulated wind gusts. With our predefined-time adaptive backstepping control, the drones establish the formation within the guaranteed 0.8 seconds and maintain precise tracking. The position tracking errors $e_{p,i}$ converge to a near-zero neighborhood well before $T_{c,pos}=1.5$s. In a comparative test against a standard fixed-time sliding mode controller, the advantage is clear. While the fixed-time controller also converges, its settling time is longer and more variable with different initial conditions. Our controller’s convergence is crisp and exactly aligned with the predefined schedule, which is non-negotiable for choreographing a formation drone light show to music.
The implications of this technical framework extend far beyond entertainment. The guaranteed timing enables new artistic expressions: shapes can morph in perfect sync with musical beats. From an engineering standpoint, it enhances safety and reliability. Knowing the exact worst-case convergence time allows for better battery management, collision avoidance planning, and emergency procedure triggering. The adaptive component makes the show resilient to real-world conditions, allowing performances in weather that would ground lesser systems.
In conclusion, the modern formation drone light show is a symphony of advanced control theory. Moving from asymptotic to predefined-time stability transforms it from a technical demonstration into a reliable, high-precision performance art. By integrating a predefined-time distributed estimator with an adaptive predefined-time backstepping controller, we achieve what every show director dreams of: absolute command over the “when” and “where” of every drone in the fleet. This synergy of guaranteed timing and robust adaptation ensures that the spectacle in the sky is not only beautiful but also predictably perfect, show after show, under the ever-changing canvas of the night. The future will involve scaling to thousands of drones, integrating real-time trajectory planning for obstacle avoidance, and further refining the algorithms for even faster transition times. The journey to light up the sky with intelligent, synchronized swarms continues, one guaranteed time step at a time.