The coordinated control of multiple autonomous agents has emerged as a cornerstone of modern robotics, enabling tasks that are infeasible for a single entity. Among these agents, the quadrotor drone stands out due to its agility, mechanical simplicity, and vertical take-off and landing capability. However, the true potential of the quadrotor drone is unlocked when deployed in swarms or formations, allowing for distributed sensing, cooperative transport, and complex aerial maneuvers. The central challenge lies in designing control laws that enable a group of quadrotor drones to achieve and maintain a desired spatial configuration using only local communication and information from neighbors, a paradigm known as distributed formation control. This article, from the perspective of a control systems engineer, delves into the application of consensus theory—a powerful framework from multi-agent systems—to solve this very problem for fleets of quadrotor drones modeled as second-order dynamical systems.
Graph Theory: The Language of Interconnection
To analyze and design control strategies for a networked system, we first need a mathematical language to describe the communication or sensing links between agents. Graph theory provides this essential foundation. A directed graph (digraph) $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ is used to model the information flow within a formation of $N$ quadrotor drones.
- Vertex Set ($\mathcal{V}$): $\mathcal{V} = \{1, 2, …, N\}$, where each vertex $i$ represents a single quadrotor drone.
- Edge Set ($\mathcal{E}$): $\mathcal{E} \subseteq \mathcal{V} \times \mathcal{V}$. An ordered pair $(j, i) \in \mathcal{E}$ signifies that quadrotor drone $i$ can receive information from quadrotor drone $j$. Quadrotor $j$ is then a neighbor of $i$.
The structure of $\mathcal{G}$ is encoded algebraically using matrices. The adjacency matrix $\mathbf{A} = [a_{ij}] \in \mathbb{R}^{N \times N}$ is defined as:
$$a_{ij} = \begin{cases} 1, & \text{if } (j, i) \in \mathcal{E} \\ 0, & \text{otherwise} \end{cases}$$
The in-degree of node $i$ is $d_i^{in} = \sum_{j=1}^{N} a_{ij}$. A more fundamental matrix for consensus analysis is the Laplacian matrix $\mathbf{L} = [l_{ij}] \in \mathbb{R}^{N \times N}$, defined as:
$$l_{ij} = \begin{cases} \sum_{k=1}^{N} a_{ik}, & \text{if } i = j \\ -a_{ij}, & \text{if } i \neq j \end{cases}$$
An important property of $\mathbf{L}$ is that it always has at least one zero eigenvalue with a corresponding eigenvector of all ones, $\mathbf{1}_N$. A directed graph is said to contain a directed spanning tree if there exists a root node (e.g., a leader quadrotor drone) that has a directed path to every other node in the graph. This topological property is crucial for ensuring that information can propagate throughout the entire formation.
| Topology Name | Graph Type | Laplacian Eigenvalues | Key Property |
|---|---|---|---|
| Directed Chain | Directed | All eigenvalues have non-negative real parts. | Simple but fragile; a single link failure partitions the formation. |
| Ring | Undirected/Directed | Eigenvalues are distributed on a circle in the complex plane for the directed case. | Provides redundancy compared to a chain. |
| Complete Graph | Undirected | $\lambda_1=0$, $\lambda_2=\lambda_3=…=\lambda_N=N$. | All-to-all communication. Ensures fastest consensus but is impractical for large $N$. |
| Leader-Follower (with spanning tree) | Directed | One zero eigenvalue, all others have positive real parts. | Leader node has no incoming edges. Enables tracking of an external command. |
Consensus Theory for Multi-Agent Systems
Consensus is the process by which a group of agents, each with its own initial value (e.g., position, velocity, heading), comes to an agreement on a common value through a distributed protocol based on neighbor-to-neighbor interactions. For a formation of quadrotor drones, we are typically concerned with achieving consensus on their spatial states.
Consider a network of $N$ agents with first-order dynamics:
$$\dot{\xi}_i(t) = u_i(t), \quad i=1,…,N$$
where $\xi_i \in \mathbb{R}$ is the state and $u_i \in \mathbb{R}$ is the control input. A foundational continuous-time consensus protocol is:
$$u_i(t) = -\sum_{j \in \mathcal{N}_i} a_{ij}(t) (\xi_i(t) – \xi_j(t))$$
where $\mathcal{N}_i$ is the set of neighbors of agent $i$. In matrix form, this becomes $\dot{\boldsymbol{\xi}} = -\mathbf{L}(t) \boldsymbol{\xi}$. The system reaches consensus, $\xi_i \to \xi_j$ for all $i,j$, asymptotically if and only if the communication graph $\mathcal{G}(t)$ contains a directed spanning tree infinitely often over time intervals.
However, the dynamics of a quadrotor drone, particularly its translational motion, are more accurately modeled as a second-order system. We model each quadrotor drone in the formation control layer as a double integrator:
$$\begin{align}
\dot{\mathbf{p}}_i(t) &= \mathbf{v}_i(t) \\
\dot{\mathbf{v}}_i(t) &= \mathbf{u}_i(t)
\end{align}$$
where $\mathbf{p}_i \in \mathbb{R}^3$ and $\mathbf{v}_i \in \mathbb{R}^3$ are the position and velocity of the $i$-th quadrotor drone in an inertial frame, and $\mathbf{u}_i \in \mathbb{R}^3$ is the formation control input (an acceleration command). A standard second-order consensus protocol is:
$$\mathbf{u}_i(t) = -\sum_{j \in \mathcal{N}_i} a_{ij}\left[ (\mathbf{p}_i – \mathbf{p}_j) + \gamma (\mathbf{v}_i – \mathbf{v}_j) \right]$$
where $\gamma > 0$ is a damping gain. Under a fixed, connected undirected graph, this protocol drives both positions and velocities to a common agreement: $\mathbf{p}_i \to \mathbf{p}_j$ and $\mathbf{v}_i \to \mathbf{v}_j$ for all $i,j$.
Modeling the Quadrotor Drone for Formation Control
To integrate consensus theory with the physical reality of a quadrotor drone, we must understand its dynamics. A quadrotor drone is an underactuated system with six degrees of freedom (position $(x,y,z)$ and orientation $(\phi, \theta, \psi)$) but only four independent control inputs (the thrusts of the four rotors). The kinematic and dynamic equations are derived using Newton-Euler formalism. Define the inertial frame $\{O_X, O_Y, O_Z\}$ and the body-fixed frame $\{o_x, o_y, o_z\}$ attached to the center of mass of the quadrotor drone.
The rotation matrix $\mathbf{R} \in SO(3)$ from the body frame to the inertial frame, parameterized by ZYX Euler angles $(\phi, \theta, \psi)$, is:
$$\mathbf{R} = \begin{bmatrix}
c_\theta c_\psi & s_\phi s_\theta c_\psi – c_\phi s_\psi & c_\phi s_\theta c_\psi + s_\phi s_\psi \\
c_\theta s_\psi & s_\phi s_\theta s_\psi + c_\phi c_\psi & c_\phi s_\theta s_\psi – s_\phi c_\psi \\
-s_\theta & s_\phi c_\theta & c_\phi c_\theta
\end{bmatrix}$$
where $s_{\cdot}$ and $c_{\cdot}$ denote $\sin(\cdot)$ and $\cos(\cdot)$, respectively. Let $\mathbf{e}_3 = [0, 0, 1]^\top$. The total thrust $T$ acts along the negative body $z$-axis. The equations of motion are:
$$\begin{align}
\dot{\mathbf{p}} &= \mathbf{v} \\
m \dot{\mathbf{v}} &= m g \mathbf{e}_3 – T \mathbf{R} \mathbf{e}_3 \quad & \text{(Translational Dynamics)} \\
\dot{\mathbf{R}} &= \mathbf{R} S(\boldsymbol{\omega}) \\
\mathbf{J} \dot{\boldsymbol{\omega}} &= -\boldsymbol{\omega} \times \mathbf{J}\boldsymbol{\omega} + \boldsymbol{\tau} \quad & \text{(Rotational Dynamics)}
\end{align}$$
where $m$ is mass, $g$ is gravity, $\mathbf{J}$ is the inertia matrix, $\boldsymbol{\omega}$ is the angular velocity in the body frame, $\boldsymbol{\tau}$ is the control torque, and $S(\cdot)$ is the skew-symmetric cross-product matrix.
To manage this complexity, a hierarchical control architecture is standard. An inner-loop attitude controller stabilizes the orientation $(\phi, \theta, \psi)$ to track commands from an outer-loop position controller. The formation control algorithm sits above this level. It takes the states of neighboring quadrotor drones (and possibly a leader) as input and outputs a desired acceleration $\mathbf{u}_i$ for each quadrotor drone $i$. This desired acceleration is then integrated to produce reference position and velocity commands for the quadrotor drone’s own low-level flight controller.
Distributed Formation Control Design
The core objective is to design the acceleration command $\mathbf{u}_i$ for the second-order model of each quadrotor drone such that the fleet achieves a desired geometric formation while possibly tracking a collective trajectory. We combine the graph-based consensus approach with the leader-follower strategy. Let quadrotor drone $1$ be the designated leader. The desired formation shape is defined by a set of constant offset vectors $\mathbf{p}_i^* \in \mathbb{R}^3$ relative to a formation center. For the leader-follower structure, we define these offsets relative to the leader’s position.
We propose the following distributed formation control law for follower quadrotor drones ($i = 2, …, N$):
$$\mathbf{u}_i(t) = \dot{\mathbf{v}}_d(t) – \alpha\left[ (\mathbf{p}_i – \mathbf{p}_i^*) – \mathbf{p}_d(t) + \gamma (\mathbf{v}_i – \mathbf{v}_d(t)) \right] – \sum_{j \in \mathcal{N}_i} a_{ij}\left[ (\mathbf{p}_i – \mathbf{p}_i^*) – (\mathbf{p}_j – \mathbf{p}_j^*) + \gamma (\mathbf{v}_i – \mathbf{v}_j) \right]$$
where:
- $\mathbf{p}_d(t), \mathbf{v}_d(t), \dot{\mathbf{v}}_d(t)$ are the leader’s position, velocity, and acceleration (known to all followers).
- $\alpha > 0$ is the coupling gain to the leader (set to 1 if the follower receives the leader’s data).
- $\gamma > 0$ is the relative velocity damping gain.
- $a_{ij}$ are elements of the adjacency matrix defining communication among followers.
The control law has an intuitive interpretation. The first term, $\dot{\mathbf{v}}_d(t)$, is a feedforward acceleration that helps the follower anticipate the leader’s motion. The second term drives the follower’s formation-corrected position $(\mathbf{p}_i – \mathbf{p}_i^*)$ and its velocity towards the leader’s states $(\mathbf{p}_d, \mathbf{v}_d)$. The third term enforces consensus on these formation-corrected states among neighboring followers. For the leader quadrotor drone, we simply set $\mathbf{u}_1(t) = \dot{\mathbf{v}}_d(t)$, as it is tasked with following the pre-defined trajectory.
Let us analyze the one-dimensional scalar case for clarity. Define the formation tracking error for agent $i$ as $\tilde{p}_i = p_i – p_i^* – p_d$. The closed-loop error dynamics for the followers become:
$$\ddot{\tilde{p}}_i = -\alpha (\tilde{p}_i + \gamma \dot{\tilde{p}}_i) – \sum_{j=2}^{N} l_{ij} (\tilde{p}_j + \gamma \dot{\tilde{p}}_j)$$
where $l_{ij}$ are elements of the Laplacian matrix $\mathbf{L}_f$ for the follower subgraph. This can be written in state-space form for the stacked error vector $\tilde{\mathbf{p}}_f$:
$$\frac{d}{dt}\begin{bmatrix} \tilde{\mathbf{p}}_f \\ \dot{\tilde{\mathbf{p}}}_f \end{bmatrix} = \left( \boldsymbol{\Theta} \otimes \mathbf{I} \right) \begin{bmatrix} \tilde{\mathbf{p}}_f \\ \dot{\tilde{\mathbf{p}}}_f \end{bmatrix}, \quad \text{where } \boldsymbol{\Theta} = \begin{bmatrix} \mathbf{0} & \mathbf{I} \\ -(\alpha\mathbf{I} + \mathbf{L}_f) & -\gamma(\alpha\mathbf{I} + \mathbf{L}_f) \end{bmatrix}$$
The stability and convergence of this system are governed by the eigenvalues of the matrix $\boldsymbol{\Theta}$. A key result from second-order consensus theory ensures that if the graph containing the leader and followers has a directed spanning tree rooted at the leader, and the gain $\gamma$ is chosen sufficiently large, then the formation tracking errors converge to zero asymptotically:
$$\lim_{t \to \infty} (p_i(t) – p_i^*) = p_d(t), \quad \lim_{t \to \infty} v_i(t) = v_d(t) \quad \forall i.$$
This implies each quadrotor drone converges to its prescribed offset relative to the moving leader, thereby achieving the desired formation shape during collective motion.
| Algorithm Type | Agent Dynamics Model | Information Required | Key Advantages | Key Challenges |
|---|---|---|---|---|
| First-Order Consensus | $\dot{\mathbf{p}}_i = \mathbf{u}_i$ | Neighbor positions | Simple design and analysis. Suitable for velocity-controlled platforms. | Cannot directly control acceleration/force. May require high gains for tracking. |
| Second-Order Consensus (Velocity Consensus) | $\ddot{\mathbf{p}}_i = \mathbf{u}_i$ | Neighbor positions and velocities | More natural model for physical systems like a quadrotor drone. Provides smoother transient response. | Requires velocity measurement/estimation. Convergence conditions are stricter (requires $\gamma$ tuning). |
| Leader-Follower Second-Order (Proposed) | $\ddot{\mathbf{p}}_i = \mathbf{u}_i$ | Neighbor pos/vel, Leader pos/vel/acceleration | Enables precise formation trajectory tracking. Leader provides a global reference. | Relies on leader’s information. Leader’s acceleration must be known or estimated. |
Simulation Analysis and Results
To validate the proposed distributed formation control law, we consider a scenario with $N=5$ quadrotor drones. Quadrotor drone 1 is the leader. The communication topology is defined such that every follower receives information from the leader, and the followers are connected in a directed path: $2 \leftarrow 3 \leftarrow 4 \leftarrow 5$. The desired formation is a diamond shape around the leader, specified by the following offset vectors (in meters):
$$\mathbf{p}^*_2 = [0.5, 0, 0]^\top, \quad \mathbf{p}^*_3 = [0, 0.5, 0]^\top, \quad \mathbf{p}^*_4 = [-0.5, 0, 0]^\top, \quad \mathbf{p}^*_5 = [0, -0.5, 0]^\top$$
The control gains are set to $\alpha=1$ and $\gamma=2$.
Scenario 1: Formation Assembly (Static Leader). The leader hovered at a fixed point $\mathbf{p}_d = [0,0,5]^\top$, $\mathbf{v}_d = \mathbf{0}$. The four follower quadrotor drones started from random initial positions on the ground $(z=0)$. The formation control law commanded them to take off and maneuver to their assigned positions relative to the stationary leader. The simulation confirmed that all followers asymptotically converged to the precise diamond configuration around the hover point, demonstrating successful formation assembly.
Scenario 2: Formation Maneuvering (Moving Leader). The leader performed a helical trajectory: $\mathbf{p}_d(t) = [-\cos(t), \sin(t), 5]^\top$. Its velocity is $\mathbf{v}_d(t) = [\sin(t), \cos(t), 0]^\top$ and acceleration is $\dot{\mathbf{v}}_d(t) = [\cos(t), -\sin(t), 0]^\top$. The followers started from disparate ground locations. The results were compelling: each quadrotor drone, guided by the distributed algorithm, initiated a spiral ascent towards its dynamically computed target point $\mathbf{p}_d(t) + \mathbf{p}^*_i$. Within approximately 5 seconds, the fleet locked into the diamond formation and subsequently maintained it perfectly while tracking the leader’s circular motion in the horizontal plane and constant altitude. The figures below illustrate the trajectories and convergence of the $X$-position and velocity states for all five quadrotor drones.
The velocity plots show an initial transient period where the followers adjust their speeds. After convergence, the velocity profiles of all four follower quadrotor drones overlap perfectly with that of the leader, visually confirming that velocity consensus was achieved. The position plots show the followers converging to their respective, constant offsets from the leader’s sinusoidally varying position. These simulations robustly validate the effectiveness of the second-order, leader-based consensus algorithm for coordinated formation flight of quadrotor drones.
Extensions and Practical Considerations
While the presented framework is effective, real-world deployment of a quadrotor drone formation introduces additional challenges. The communication topology may be time-varying due to link failures or range constraints, requiring algorithms robust to switching graphs. Communication delays are inevitable and can degrade performance or even cause instability if not accounted for in the control design. Each individual quadrotor drone is subject to model uncertainties, external disturbances (e.g., wind), and actuator saturation. Integrating adaptive control or robust $H_\infty$ techniques with the consensus protocol is a vital area of research.
Furthermore, the assumption that every follower knows the leader’s acceleration $\dot{\mathbf{v}}_d(t)$ can be relaxed. Observers or estimators can be designed to allow followers to reconstruct this information from the leader’s position and velocity data. Finally, extending the consensus objective to include attitude synchronization (for coordinated camera pointing or aerodynamic efficiency in forward flight) for a fleet of quadrotor drones represents a more complex, higher-dimensional consensus problem on the Special Euclidean group $SE(3)$.
In conclusion, the fusion of graph theory and second-order consensus principles provides a powerful, scalable, and theoretically sound methodology for the distributed formation control of quadrotor drones. By modeling the quadrotor drone as a double-integrator agent and carefully designing the interaction protocol based on the communication topology, we can achieve complex collective behaviors such as dynamic formation assembly and precision trajectory tracking. This approach, validated through simulation, lays a critical foundation for the development of intelligent, resilient, and collaborative swarms of quadrotor drones capable of executing sophisticated missions in the real world.