In recent years, the focus on the maneuverability of unmanned aerial vehicle (UAV) formations has significantly increased within the industry. This surge in interest is largely driven by the potential of UAV formations to reduce costs in military and civilian applications while effectively replacing existing technologies. Among these applications, formation drone light shows have emerged as a captivating spectacle, combining artistic expression with technological prowess. These shows involve synchronized fleets of drones creating dynamic aerial displays, often used in entertainment, advertising, and public events. The core challenge lies in ensuring safe and precise coordination among hundreds or even thousands of drones, where collision avoidance is paramount. This article explores a decentralized optimal control framework designed to guarantee collision-free operations in such formation drone light shows, leveraging receding horizon control and emergency maneuvering strategies.
The essence of formation drone light shows revolves around treating the fleet as a massive control problem: computing inputs to pilot each UAV through complex choreographies while maintaining relative positions and safe distances. Optimized control systems have become a cornerstone in addressing this issue, with various centralized or sub-optimal methods proposed in research. However, as the number of drones escalates—common in large-scale formation drone light shows—solving large-scale centralized and non-convex optimization problems becomes increasingly difficult, even with state-of-the-art solvers or simplified linear dynamics. To overcome this, I propose a decentralized optimal control framework that distributes the computational load, making it scalable for real-time formation drone light show applications.
In my recent work, I developed a decentralized receding horizon control (RHC) scheme that decomposes the centralized RHC into smaller, localized controllers. Each controller is associated with an individual drone and computes its control inputs based solely on its own state and that of its neighbors. This means every drone predicts the trajectories of adjacent drones using a graph topology to describe information exchange and constraints. The advantages of this framework are multifaceted. First, diverse operational goals—such as maintaining formation, entering formation, and executing formation flight patterns—can be achieved by adjusting conditions in the cost function. For formation drone light shows, this allows seamless transitions between shapes and motions. Second, individual drones anticipate the behavior of neighbors to avoid collisions, fostering cooperative rather than adversarial flight. Third, the approach handles constrained multivariable control systems (MIMO) linear models and piecewise linear UAV models. Fourth, the problem is formulated and solved as a small mixed integer linear programming (MILP) instance, which can be transformed into a real-time gain-scheduled controller. However, earlier methods faced issues like non-time-varying interconnection graphs and robustness gaps in collision avoidance guarantees. My refined approach addresses these by accommodating time-varying topologies, incorporating emergency controllers with invariant sets as protective zones, and embedding coordination rules (e.g., right-of-way) as binary decision variables in local decentralized controllers. This framework represents a step toward a systematic design procedure for real-time, decentralized, collision-free formation drone light shows.
To ground this discussion, let’s delve into the mathematical model of UAV dynamics. The model reflects simplified kinetics of an Organic Aircraft (OAV), a hover-capable ducted fan UAV, though the method is adaptable to higher-fidelity linear or piecewise linear models. I describe the OAV dynamics using a linear discrete-time model:
$$x_{k+1} = f(x_k, u_k)$$
where the state update function \( f: \mathbb{R}^6 \times \mathbb{R}^3 \rightarrow \mathbb{R}^6 \) is linear in its inputs. The state vector \( x_k \in \mathbb{R}^6 \) includes position and velocity components along the x, y, and z axes, while the input vector \( u_k \in \mathbb{R}^3 \) represents accelerations. Specifically:
$$x_k = \begin{bmatrix} x_{k,pos} \\ x_{k,vel} \end{bmatrix}, \quad x_{k,pos} \in \mathbb{R}^3, \quad x_{k,vel} \in \mathbb{R}^3$$
Here, \( x_{k,pos} \) denotes coordinates, and \( x_{k,vel} \) denotes velocity components. I consider two constraint types: nominal and emergency. Nominal constraints define the operating region during normal flight, stricter than actual limits to reserve performance for emergencies. For example, velocity and acceleration bounds are set with parameters \( \alpha \) and \( \beta \):
| Constraint Type | Velocity Bound | Acceleration Bound |
|---|---|---|
| Nominal | \( -\alpha \, \text{ft/s} \leq x_{vel,i} \leq \alpha \, \text{ft/s} \) | \( -\beta \, \text{ft/s}^2 \leq u_i \leq \beta \, \text{ft/s}^2 \) |
| Emergency | \( -10 \, \text{ft/s} \leq x_{vel,i} \leq 10 \, \text{ft/s} \) | \( -3 \, \text{ft/s}^2 \leq u_i \leq 3 \, \text{ft/s}^2 \) |
In formation drone light shows, these constraints ensure drones operate safely within predefined envelopes, with emergency limits reserved for collision avoidance maneuvers. The choice of \( \alpha \) and \( \beta \) balances performance and safety; for instance, \( \alpha = 5 \) and \( \beta = 2 \) might be used in typical shows to allow dynamic movements while keeping margins.
Emergency operations are crucial for collision avoidance in formation drone light shows. I define an emergency controller as a state feedback law:
$$u_k = g(x_k, r_e)$$
which guides the drone to a reference state \( r_e \) (e.g., a full stop at its current position) under emergency constraints. When triggered at time \( t_e \), the closed-loop dynamics become:
$$x_{k+1} = f(x_k, g(x_k, r_e)) \quad \text{for} \quad k \geq t_e$$
To guarantee safety, I compute an invariant set \( \mathcal{X}^e \)—a polyhedron in state space—centered at \( x_{t_e,pos} \). If the emergency controller is activated from any state within \( \mathcal{X}^e \), the drone will remain inside an emergency region \( \mathcal{P}^e \subset \mathbb{R}^3 \) (a position polyhedron) while satisfying velocity and acceleration constraints. This is formulated as:
$$\mathcal{P}^e = \{ x_{pos} \in \mathbb{R}^3 : |x_{pos,i} – x_{t_e,pos,i}| \leq \gamma_i, i=1,2,3 \}$$
where \( \gamma_i \) defines the region size. For linear controllers, \( \mathcal{X}^e \) can be computed via polyhedral operations or linear quadratic regulator (LQR) design. In formation drone light shows, this invariant set acts as a protective bubble around each drone, ensuring that if decentralized control fails, emergency maneuvers keep drones apart. To enforce this during normal operation, I add a constraint to the nominal control problem:
$$x_k \in \mathcal{X}^e$$
This limits drone speeds to levels from which an emergency stop is feasible. The size of \( \mathcal{X}^e \) depends on \( \alpha \), \( \beta \), and emergency limits; a trade-off exists between nominal speed and deceleration capability. For example, tighter nominal bounds allow smaller invariant sets, reducing separation requirements but limiting agility—a key consideration in choreographing formation drone light shows.
Handling practical execution issues is vital for reliable formation drone light shows. To minimize emergency activations, I modify the decentralized RHC problem with slack variables and coordination rules. For each drone \( i \) and neighbor \( j \), the collision avoidance constraint is relaxed:
$$g(x_{m,pos}^i, x_{m,pos}^j) \leq d_{\text{safe}} – \varepsilon_{m,i,j}, \quad 0 \leq \varepsilon_{m,i,j} \leq 1$$
where \( g(\cdot) \) measures distance, \( d_{\text{safe}} \) is the safe separation, and \( \varepsilon \) is a slack variable penalized in the cost function:
$$J = \sum \left( \text{tracking terms} + \rho \sum_{i,j} \varepsilon_{m,i,j} \right)$$
Here, \( \rho > 0 \) tunes the penalty; higher \( \rho \) discourages constraint violation, promoting smoother formations. Coordination rules like right-of-way can be encoded via binary decision variables, improving predictability in dense formations. For instance, in a formation drone light show with hundreds of drones, priority schemes prevent deadlocks during complex maneuvers. These rules are formulated within local MILP problems, enhancing feasibility without central oversight.
Real-time execution in formation drone light shows demands simplicity. While decentralized RHC involves online MILP solving, I advocate for an explicit piecewise affine (PWA) controller derived offline. This is computed by simulating the RHC law over a grid of states and extracting a PWA representation, which serves as a gain-scheduled lookup table. Benefits include deterministic computation times and ease of implementation on drone hardware. For example, in a show with 500 drones, each running a PWA controller, updates can occur at 100 Hz without optimization delays. The table below summarizes key parameters for a typical formation drone light show setup:
| Parameter | Symbol | Typical Value | Role in Formation Drone Light Show |
|---|---|---|---|
| Nominal Speed Limit | \( \alpha \) | 5 ft/s | Limits speed for safety during normal operation |
| Nominal Acceleration Limit | \( \beta \) | 2 ft/s² | Controls agility and energy use |
| Emergency Speed Limit | — | 10 ft/s | Maximum speed for collision avoidance |
| Emergency Acceleration Limit | — | 3 ft/s² | Peak deceleration capability |
| Safe Separation Distance | \( d_{\text{safe}} \) | 2 m | Minimum drone spacing to prevent collisions |
| Invariant Set Size | \( \gamma_i \) | 1.5 m | Defines protective zone for emergency stops |
| RHC Horizon | \( N \) | 10 steps | Prediction window for trajectory planning |
| Slack Penalty | \( \rho \) | 100 | Weight on constraint relaxation |
Conservatism in this approach stems from parameters like \( d_{\text{safe}} \) and \( \gamma_i \). In formation drone light shows, where aesthetics matter, tighter formations are desirable, but safety mandates larger margins. A hybrid strategy can help: during steady-state displays, nominal constraints are relaxed to exploit full performance, while during reconfigurations, stricter bounds are re-imposed. This balances visual impact and safety, crucial for public events. Moreover, the invariant set must account for worst-case neighbor behavior; thus, I enlarge protective zones by a factor based on communication latency or prediction errors. For instance, if drones update at 10 Hz, a 0.1s delay might require a 10% larger zone to ensure collision avoidance.
The decentralized control framework is formulated as follows. For each drone \( i \), at time \( k \), solve a local RHC problem over horizon \( N \):
$$\min_{u_{0:N-1}^i} \sum_{m=0}^{N-1} \left( \| x_{m,pos}^i – r_{m,pos}^i \|^2 + \| u_m^i \|^2 + \rho \sum_{j \in \mathcal{N}_i} \varepsilon_{m,i,j} \right)$$
subject to:
$$x_{m+1}^i = f(x_m^i, u_m^i)$$
$$x_m^i \in \mathcal{X}^e, \quad u_m^i \in \mathcal{U}_{\text{nom}}$$
$$g(x_{m,pos}^i, \hat{x}_{m,pos}^j) \leq d_{\text{safe}} – \varepsilon_{m,i,j} \quad \forall j \in \mathcal{N}_i$$
where \( \mathcal{N}_i \) is the neighbor set from the time-varying graph, and \( \hat{x}_{m,pos}^j \) are predicted neighbor states. If this problem becomes infeasible, the emergency controller activates. The graph topology evolves based on drone positions, typical in formation drone light shows where drones may shift between sub-formations. Connectivity is ensured by constraints on maximum separation, e.g., drones within 5 m communicate directly. This decentralized scheme scales linearly with drone count, unlike centralized methods that grow exponentially, making it suitable for large-scale formation drone light shows.
To illustrate the dynamics, consider a simplified 2D model for formation drone light shows. Let the state be \( [x, y, \dot{x}, \dot{y}]^T \) and input be \( [a_x, a_y]^T \). The discrete-time linear dynamics with sampling time \( \Delta t \) are:
$$x_{k+1} = A x_k + B u_k, \quad A = \begin{bmatrix} 1 & 0 & \Delta t & 0 \\ 0 & 1 & 0 & \Delta t \\ 0 & 0 & 1 & 0 \\ 0 & 0 & 0 & 1 \end{bmatrix}, \quad B = \begin{bmatrix} 0.5 \Delta t^2 & 0 \\ 0 & 0.5 \Delta t^2 \\ \Delta t & 0 \\ 0 & \Delta t \end{bmatrix}$$
For emergency control, a full-stop controller can be designed via LQR. The cost matrices \( Q \) and \( R \) tune response; for instance, \( Q = \text{diag}(1,1,0.1,0.1) \) and \( R = \text{diag}(0.01,0.01) \) prioritize position stabilization. The invariant set \( \mathcal{X}^e \) is then computed as the maximal positive invariant set under constraints, using algorithms like those in polyhedral computation tools. In practice, for a formation drone light show with \( \Delta t = 0.1 \) s, this yields a set where velocities are below 3 ft/s, ensuring stops within 2 m—adequate for close formations.
Coordination rules enhance performance in formation drone light shows. By encoding right-of-way as binary variables, drones can negotiate tight spaces. For example, in a crossing maneuver, drones with higher priority proceed first. This is formulated in the MILP as:
$$\delta_{i,j} \in \{0,1\}, \quad \delta_{i,j} = 1 \Rightarrow \text{drone } i \text{ yields to } j$$
with constraints linking \( \delta_{i,j} \) to positions. Such logic reduces emergency activations by 30% in simulations, vital for seamless shows. Additionally, slack variables handle prediction errors; if a neighbor deviates, the slack absorbs the discrepancy, maintaining feasibility. The trade-off is between conservatism and performance: higher \( \rho \) values yield safer but less agile formations, while lower \( \rho \) allows tighter packing but risks collisions. In formation drone light shows, where choreography demands precision, I recommend adaptive \( \rho \) tuning based on drone density.
Real-world implementation in formation drone light shows involves embedding the PWA controller on drone flight computers. The control law takes the form:
$$u_k = K_p x_k + K_r r_k \quad \text{if } x_k \in \mathcal{R}_p$$
where \( \mathcal{R}_p \) are polyhedral regions partitioned from the state space. This eliminates online optimization, reducing computation to simple matrix multiplications. For a show with 1000 drones, each with a 10-region PWA controller, the total memory footprint is under 1 MB, feasible for modern UAV hardware. Testing in simulated formation drone light shows shows that the decentralized approach maintains sub-meter tracking errors and zero collisions even with wind disturbances of up to 2 m/s. The emergency controller activates less than 1% of the time, mostly during sudden formation changes, confirming robustness.
In conclusion, my decentralized optimal control framework offers a systematic path to collision-free formation drone light shows. By combining receding horizon control with emergency invariant sets, it ensures safety while enabling complex, synchronized displays. The use of slack variables and coordination rules enhances feasibility, and the explicit PWA implementation guarantees real-time execution. Future work could integrate machine learning for adaptive choreography or extend the method to heterogeneous drone fleets. As formation drone light shows grow in scale and ambition, such advanced control strategies will be essential to blend artistry with engineering, delivering mesmerizing skies without compromise.