The coordinated flight of a drone formation presents significant advantages over single-unit operations across numerous domains, including surveillance, search and rescue, and precision delivery. A well-coordinated drone formation enhances area coverage, improves system redundancy, and can execute complex maneuvers with greater efficiency. At the heart of enabling such autonomous, cohesive flight lies the critical problem of real-time path planning and formation control. This task involves navigating a dynamic or partially known environment filled with obstacles while maintaining a predefined geometric configuration among the drones. The challenge is compounded by the stringent kinematic constraints of the vehicles, such as maximum turn rates, and the imperative for collision avoidance both with environmental threats and between members of the drone formation itself.
Among the plethora of path-planning algorithms, the Artificial Potential Field (APF) method stands out for its conceptual elegance and computational efficiency. Its principle is intuitively appealing: the target exerts an attractive force on the drone, while obstacles exert repulsive forces. The drone then moves according to the resultant force vector. This approach is highly reactive and suitable for real-time applications. However, the classical APF formulation is plagued by well-documented shortcomings that can lead to mission failure, particularly in complex environments relevant to a drone formation. These include the local minima problem, where the drone gets trapped in a force equilibrium point before reaching the goal; goal non-reachability, which occurs when an obstacle is very close to or surrounds the target, creating a repulsive barrier; and oscillatory behavior in narrow passages or near obstacle boundaries.
This article presents a comprehensive, integrated framework for drone formation route planning based on a significantly enhanced APF model. Our methodology directly addresses the classic limitations through three core innovations: a modified repulsive potential field function that guarantees goal reachability, a stochastic perturbation technique to escape local minima, and an obstacle-linking strategy to dampen oscillations. Furthermore, we seamlessly integrate formation-keeping control within the same APF paradigm, treating desired relative positions between drones as attractive potentials. To ensure the safety of the entire drone formation, we incorporate an obstacle inflation technique. The efficacy of this holistic approach is validated through extensive simulation in a complex 2D environment with multiple static threats. The results demonstrate robust obstacle avoidance, stable formation maintenance, and successful goal attainment for the multi-drone team.
Theoretical Foundations and Limitations of Classical APF
The foundation of the APF method is the construction of a scalar potential field $U_{total}(\mathbf{q})$ defined over the configuration space of the drone, where $\mathbf{q}$ denotes the drone’s position vector. This total field is a superposition of an attractive field $U_{att}$ pulling the drone toward the goal $G$, and repulsive fields $U_{rep}^i$ pushing it away from obstacles $O_i$.
The classical attractive potential and its corresponding force are given by:
$$U_{att}(\mathbf{q}) = \frac{1}{2} \zeta \rho_G^2(\mathbf{q})$$
$$F_{att}(\mathbf{q}) = -\nabla U_{att}(\mathbf{q}) = -\zeta \rho_G(\mathbf{q}) \cdot \nabla \rho_G(\mathbf{q}) = \zeta (\mathbf{q}_G – \mathbf{q})$$
where $\zeta$ is a positive scaling gain, and $\rho_G(\mathbf{q}) = ||\mathbf{q}_G – \mathbf{q}||$ is the Euclidean distance to the goal. The force is thus a vector linearly proportional to the distance from the goal, pointing toward it.
The classical repulsive potential for an obstacle $i$ is defined within its region of influence:
$$U_{rep}^i(\mathbf{q}) = \begin{cases}
\frac{1}{2} \eta \left( \frac{1}{\rho_i(\mathbf{q})} – \frac{1}{\rho_0} \right)^2, & \text{if } \rho_i(\mathbf{q}) \le \rho_0 \\
0, & \text{if } \rho_i(\mathbf{q}) > \rho_0
\end{cases}$$
where $\eta$ is a positive repulsive gain, $\rho_i(\mathbf{q})$ is the distance to obstacle $i$, and $\rho_0$ is the constant radius of influence of the obstacle. The corresponding repulsive force is the negative gradient:
$$F_{rep}^i(\mathbf{q}) = -\nabla U_{rep}^i(\mathbf{q}) = \begin{cases}
\eta \left( \frac{1}{\rho_i(\mathbf{q})} – \frac{1}{\rho_0} \right) \frac{1}{\rho_i^2(\mathbf{q})} \frac{\partial \rho_i(\mathbf{q})}{\partial \mathbf{q}}, & \text{if } \rho_i(\mathbf{q}) \le \rho_0 \\
0, & \text{if } \rho_i(\mathbf{q}) > \rho_0
\end{cases}$$
This force points directly away from the obstacle.
The total force acting on the drone is the vector sum:
$$F_{total}(\mathbf{q}) = F_{att}(\mathbf{q}) + \sum_{i} F_{rep}^i(\mathbf{q})$$
The drone’s path is then generated by iteratively moving in the direction of this resultant force with a fixed step size $\Delta s$: $\mathbf{q}_{k+1} = \mathbf{q}_k + \Delta s \cdot \frac{F_{total}(\mathbf{q}_k)}{||F_{total}(\mathbf{q}_k)||}$.
The fundamental limitations of this classical model are summarized in the table below:
| Limitation | Description | Consequence for Drone Formation |
|---|---|---|
| Local Minima | Occurs at points $\mathbf{q}$ where $F_{total}(\mathbf{q}) = 0$ but $\mathbf{q} \neq \mathbf{q}_G$. Common in symmetric obstacle configurations or when attractive and repulsive forces are collinear and opposite. | The lead drone, and consequently the entire formation, halts progress, leading to mission failure. |
| Goal Non-Reachability | When the goal lies within the influence region of an obstacle ($\rho_G < \rho_0$), the repulsive force at the goal is non-zero. As the drone approaches, $F_{att} \to 0$ but $F_{rep} \nrightarrow 0$, potentially creating a local minimum at the goal itself. | The formation cannot achieve its final objective, even if the path is clear. |
| Oscillations | In narrow corridors or near obstacle boundaries, small movements can cause large swings in the repulsive force direction, leading to oscillatory, non-smooth, or unstable paths. | Creatues unpredictable and energy-inefficient paths, complicates follower tracking, and increases risk of intra-formation collision. |
The Enhanced APF Model for Robust Single-Drone Planning
Our enhanced APF model introduces targeted modifications to each component of the classical method, specifically designed to overcome the aforementioned limitations for reliable navigation, which forms the basis for our drone formation strategy.
1. Modified Repulsive Potential for Guaranteed Goal Reachability
To solve the goal non-reachability problem, we reformulate the repulsive potential to ensure it vanishes as the drone approaches the target. This is achieved by incorporating the distance to the goal $\rho_G(\mathbf{q})$ as a multiplicative factor.
The enhanced repulsive potential is defined as:
$$U_{rep, enh}^i(\mathbf{q}) = \begin{cases}
\frac{1}{2} \eta \left( \frac{1}{\rho_i(\mathbf{q})} – \frac{1}{\rho_0} \right)^2 \rho_G^n(\mathbf{q}), & \text{if } \rho_i(\mathbf{q}) \le \rho_0 \\
0, & \text{if } \rho_i(\mathbf{q}) > \rho_0
\end{cases}$$
where $n > 0$ is a tuning parameter (typically $n=2$). The key property is that $\lim_{\mathbf{q} \to \mathbf{q}_G} U_{rep, enh}^i(\mathbf{q}) = 0$, guaranteeing the goal point is the global minimum of the total potential field.
The corresponding repulsive force is derived as the negative gradient:
$$F_{rep, enh}^i(\mathbf{q}) = -\nabla U_{rep, enh}^i(\mathbf{q}) = F_{rep1}^i(\mathbf{q}) + F_{rep2}^i(\mathbf{q}) \quad \text{for } \rho_i(\mathbf{q}) \le \rho_0$$
where
$$F_{rep1}^i(\mathbf{q}) = \eta \left( \frac{1}{\rho_i(\mathbf{q})} – \frac{1}{\rho_0} \right) \frac{1}{\rho_i^2(\mathbf{q})} \rho_G^n(\mathbf{q}) \cdot \frac{\partial \rho_i(\mathbf{q})}{\partial \mathbf{q}}$$
$$F_{rep2}^i(\mathbf{q}) = -\frac{n}{2} \eta \left( \frac{1}{\rho_i(\mathbf{q})} – \frac{1}{\rho_0} \right)^2 \rho_G^{n-1}(\mathbf{q}) \cdot \frac{\partial \rho_G(\mathbf{q})}{\partial \mathbf{q}}$$
Here, $F_{rep1}^i$ points directly away from the obstacle, while $F_{rep2}^i$ points toward the goal. At the goal ($\rho_G=0$), both components are zero. This modification ensures that obstacles very close to the target do not perpetually repel the drone, allowing successful mission completion for the lead agent in the drone formation.
2. Stochastic Perturbation for Escaping Local Minima
To address local minima, we employ a reactive strategy. When the algorithm detects a near-zero resultant force ($||F_{total}|| < \epsilon$) and the goal has not been reached, we infer the drone is trapped. A temporary stochastic perturbation force $F_{pert}$ is applied. This force is designed to nudge the drone out of the equilibrium point.
We define $F_{pert}$ as:
$$F_{pert} = \omega \cdot ||F_{att}(\mathbf{q})|| \cdot \mathbf{\hat{n}}$$
where $\omega$ is a positive gain coefficient, and $\mathbf{\hat{n}}$ is a unit vector perpendicular to the instantaneous goal-direction vector $(\mathbf{q}_G – \mathbf{q})$. The direction ($\pm \mathbf{\hat{n}}$) is chosen randomly, adhering to the right-hand rule convention. The magnitude is proportional to the current attractive force, ensuring the perturbation is context-aware—larger when far from the goal (where escape may need more energy) and smaller when near. After applying this force for a few iterations, the standard APF forces take over again, typically allowing the drone to proceed on a new path away from the trap. This mechanism is crucial for ensuring the continuous progress of the drone formation leader.
3. Obstacle-Linking Strategy to Mitigate Oscillations
Oscillatory behavior frequently arises when navigating through narrow passages between obstacles or skirting tight corners. The rapid change in repulsive force direction from two closely spaced obstacles causes the drone to “jitter.” Our solution is the obstacle-linking strategy. When two or more obstacles are in close proximity (e.g., the gap between them is less than a safety threshold $d_{safe}$), the algorithm virtually merges them into a single, larger composite obstacle for the purpose of force calculation.
If obstacles $O_i$ and $O_j$ are to be linked, we define a new virtual obstacle. A simple and effective method is to treat the line segment connecting their centers (or the convex hull for more than two) as a forbidden zone. The distance $\rho_{link}(\mathbf{q})$ from the drone to this linked obstacle (e.g., the shortest distance to the line segment) is used in a single, combined repulsive potential function instead of calculating separate forces from each individual obstacle. This smoothens the transition of the repulsive force field in the narrow passage, eliminating the oscillatory instability and producing a clean, viable path through tight spaces for the drone formation.
Integrated Path Planning and Formation Control for Drone Formation
We adopt a leader-follower architecture for the drone formation. The leader drone is responsible for primary obstacle-aware navigation to the global target using the enhanced APF described above. The follower drones are tasked with maintaining a predefined geometric relationship with the leader (or with a designated reference point in the formation) while also avoiding obstacles.
Leader Drone Path Planning Algorithm
The algorithm for planning the leader’s path incorporates all enhancements and practical constraints.
Step 1: Environment Preprocessing. Analyze the threat map. Apply the obstacle-linking strategy to any groups of obstacles separated by less than $d_{safe}$ to create a simplified environment model and prevent oscillations.
Step 2: Initialize APF Parameters. Set the gains $\zeta$, $\eta$, influence radius $\rho_0$, step size $\Delta s$, perturbation gain $\omega$, and local minima threshold $\epsilon$. Initialize the leader’s position $\mathbf{q}_L^{(0)}$.
Step 3: Force Calculation at Current Position $\mathbf{q}_L^{(k)}$.
– Calculate attractive force: $F_{att}^{(k)} = \zeta (\mathbf{q}_G – \mathbf{q}_L^{(k)})$.
– For all obstacles (including linked ones), calculate enhanced repulsive forces $F_{rep, enh}^{i,(k)}$.
– Compute total force: $F_{total}^{(k)} = F_{att}^{(k)} + \sum_i F_{rep, enh}^{i,(k)}$.
Step 4: Local Minima Check & Perturbation.
If $||F_{total}^{(k)}|| < \epsilon$ and $||\mathbf{q}_G – \mathbf{q}_L^{(k)}|| > \delta_{goal}$:
– Generate perturbation force $F_{pert}$ as defined.
– Set $F_{total}^{(k)} = F_{pert}$.
Step 5: Compute Candidate Next Position.
$\mathbf{q}_{candidate} = \mathbf{q}_L^{(k)} + \Delta s \cdot \frac{F_{total}^{(k)}}{||F_{total}^{(k)}||}$.
Step 6: Enforce Kinematic Constraint (Maximum Turn Angle). A critical practical constraint for fixed-wing drones is the maximum allowable turn angle $\theta_{max}$ between path segments. Let $\mathbf{a} = \mathbf{q}_L^{(k)} – \mathbf{q}_L^{(k-1)}$ and $\mathbf{b} = \mathbf{q}_{candidate} – \mathbf{q}_L^{(k)}$. The turn angle $\theta$ is:
$$\theta = \arccos\left( \frac{\mathbf{a} \cdot \mathbf{b}}{||\mathbf{a}|| \cdot ||\mathbf{b}||} \right)$$
If $\theta > \theta_{max}$, the candidate point is invalid. We compute a smoothed waypoint:
$$\mathbf{q}_L^{(k)} \leftarrow \mathbf{q}_L^{(k-1)} + l \cdot \mathbf{\hat{e}}$$
$$\text{where } \mathbf{e} = \frac{(\mathbf{q}_L^{(k)} + \mathbf{q}_{candidate})}{2} – \mathbf{q}_L^{(k-1)}, \quad \mathbf{\hat{e}} = \frac{\mathbf{e}}{||\mathbf{e}||}$$
and $l$ is a small step length. Then, return to Step 3 using the updated $\mathbf{q}_L^{(k)}$. If $\theta \le \theta_{max}$, accept the candidate: $\mathbf{q}_L^{(k+1)} = \mathbf{q}_{candidate}$.
Step 7: Goal Check. If $||\mathbf{q}_G – \mathbf{q}_L^{(k+1)}|| < \delta_{goal}$, stop. Otherwise, set $k = k+1$ and go to Step 3.
This process generates a sequence of waypoints $\{\mathbf{q}_L^{(0)}, \mathbf{q}_L^{(1)}, \dots, \mathbf{q}_L^{(N)}\}$ constituting the leader’s path. A simulation result for the leader in a cluttered environment with parameters $\zeta=1$, $\eta=1$, $\rho_0=R_{obs}/2$, $\omega=1$, $\theta_{max}=45^\circ$, $\Delta s = 0.1$ km demonstrates successful avoidance of multiple threats, including one very close to the goal, with a smooth trajectory respecting the turn constraint.
Follower Drone Control via Formation Potentials
Each follower drone $F_j$ must maintain a specific offset $(l_j, \phi_j)$ relative to the leader’s current position, where $l_j$ is the desired distance and $\phi_j$ is the bearing angle in the leader’s body frame. The follower’s desired position $\mathbf{q}_{F_j}^d(k)$ is calculated in real-time based on the leader’s current state $\mathbf{q}_L(k)$ and heading $\psi_L(k)$:
$$\mathbf{q}_{F_j}^d(k) = \mathbf{q}_L(k) + l_j \cdot [\cos(\psi_L(k)+\phi_j), \sin(\psi_L(k)+\phi_j)]^T$$
We embed this formation-keeping objective into an APF framework for the follower. The primary attractive force for follower $F_j$ is not toward the global goal, but toward its instantaneous desired position relative to the leader. This is termed the formation potential.
The formation potential and force for follower $j$ are:
$$U_{form}^j(\mathbf{q}_{F_j}) = \frac{1}{2} \kappa_j ||\mathbf{q}_{F_j} – \mathbf{q}_{F_j}^d||^2$$
$$F_{form}^j(\mathbf{q}_{F_j}) = -\nabla U_{form}^j(\mathbf{q}_{F_j}) = \kappa_j (\mathbf{q}_{F_j}^d – \mathbf{q}_{F_j})$$
where $\kappa_j$ is a formation gain. Simultaneously, the follower must avoid obstacles using the same enhanced repulsive potential $U_{rep, enh}$ as the leader. Therefore, the total force on follower $j$ is:
$$F_{total}^{F_j} = F_{form}^j + \sum_i F_{rep, enh}^{i}$$
The follower’s path is generated by iteratively moving under $F_{total}^{F_j}$, subject to the same turn angle constraint. This approach causes the follower to be “dragged along” by the moving desired point while actively avoiding obstacles, effectively maintaining the drone formation shape during transit. In open spaces, the formation is precisely kept. When the leader maneuvers around an obstacle, the follower’s desired point moves accordingly, and the follower’s own repulsive forces ensure it finds a safe local path to regain its position, causing temporary, elastic deformation of the formation which recovers once the threat is passed.
Collision Avoidance for the Entire Drone Formation via Obstacle Inflation
The follower control strategy above ensures followers avoid obstacles based on their own position. However, this does not guarantee that the entire volume occupied by the drone formation is clear, especially when the leader’s path passes close to an obstacle’s edge. A follower maintaining its offset could fly into the obstacle even if the leader is clear. To ensure collective safety, we treat the entire drone formation as a single entity with a circumscribed radius $R_{form}$. This radius is determined by the maximum distance from the formation centroid to any follower’s desired position: $R_{form} = \max( l_j ) + r_{drone}$.
We then perform obstacle inflation for path planning. Every obstacle’s geometric representation is expanded (inflated) outward by $R_{form}$. The leader’s path is planned using the enhanced APF in this inflated environment. Since the leader is now planning for the formation’s footprint, any path that remains clear of inflated obstacles guarantees that the entire drone formation, with followers at their nominal offsets, will remain clear of the actual, physical obstacles. This elegantly reduces the complex multi-body collision avoidance problem to a single-agent path planning problem in a modified environment. The simulation confirms that with this inflation, follower drones skirt the actual obstacle boundaries safely even when their nominal paths would have intersected the threat zone.
The challenges of real-world drone formation flight extend beyond algorithmic simulation. As illustrated, dense formations operating in complex environments must manage physical aerodynamic interactions, precise relative localization (often relying on GNSS and vision systems), and robust inter-drone communication links. The light trails in the image visually represent the kind of coordinated, tightly-coupled paths our enhanced APF algorithm seeks to generate computationally, but achieving this in practice requires seamless integration of planning, control, and estimation under real-world uncertainties like wind and sensor noise.
Parameter Selection and Performance Considerations
The performance of the enhanced APF framework is highly dependent on the chosen parameters. The following table provides a guide for tuning based on simulation experience and theoretical rationale.
| Parameter | Symbol | Recommended Range / Effect | Guideline |
|---|---|---|---|
| Attractive Gain | $\zeta$ | 0.5 – 5 | Higher values make the drone more aggressive towards the goal but can lead to overshoot. Set to ensure $F_{att}$ dominates at long ranges. |
| Repulsive Gain | $\eta$ | 1 – 20 | Higher values increase safety margin from obstacles but can exacerbate local minima. Balance with $\zeta$. Start with $\eta \approx \zeta$. |
| Obstacle Influence | $\rho_0$ | 1.5$\cdot R_{obs}$ – 3$\cdot R_{obs}$ | Must be > $R_{form}$. Larger values cause earlier, smoother deviations but can clutter the potential field. |
| Goal Distance Exponent | $n$ | 2 | Typically $n=2$ provides a good balance, ensuring repulsive force decays quadratically with goal proximity. |
| Perturbation Gain | $\omega$ | 0.5 – 2 | Controls escape energy from local minima. Too small may not escape; too large causes erratic detours. |
| Formation Gain | $\kappa_j$ | 5 – 15 | Should be significantly higher than $\zeta$ to ensure formation keeping is a stronger objective than going to the global goal for followers. |
| Step Size | $\Delta s$ | 0.01$D_{env}$ – 0.05$D_{env}$ | Where $D_{env}$ is environment scale. Small for precision, large for speed. Must satisfy $\Delta s < \rho_0$. |
| Max Turn Angle | $\theta_{max}$ | 30° – 60° | Dictated by drone dynamics (e.g., 45° for many fixed-wing UAVs). The constraint is enforced in Step 6 of the leader algorithm. |
The computational complexity of the algorithm per iteration is $O(M)$ for the leader and $O(M)$ per follower, where $M$ is the number of (potentially linked) obstacles. This linear scaling makes it highly suitable for real-time implementation in a drone formation, even with dozens of obstacles. The most computationally expensive steps are distance calculations and gradient computations for each obstacle.
Discussion: Applications and Future Extensions
The proposed enhanced APF framework is particularly suited for applications requiring autonomous, cohesive movement of a drone formation in semi-structured or dynamically changing environments. Key application domains include:
- Military Reconnaissance/Surveillance: A drone formation can cover a wider area, provide multiple viewing angles, and maintain operation if one unit is lost.
- Precision Agriculture: Formations can efficiently survey large fields, with different drones carrying different sensors (e.g., multispectral, thermal).
- Infrastructure Inspection: A drone formation can simultaneously inspect different sections of a large structure like a bridge or dam.
- Public Safety & Disaster Response: Formations can be deployed for searching large areas, mapping disaster zones, or establishing communication relays.
While the current model handles static obstacles effectively, real-world operations demand extensions for full autonomy. Future work will focus on integrating dynamic obstacle prediction and avoidance. This can be achieved by modeling moving obstacles with velocity vectors and adding a repulsive potential that is a function of both relative position and velocity (like the Velocity Obstacle concept). Another vital extension is 3D path planning, incorporating terrain elevation and constraints like minimum/maximum altitude. The potential field formalism naturally extends to 3D, with forces acting in $(x, y, z)$ space. Finally, decentralizing the leader-follower model to a more cooperative scheme, where each drone in the drone formation calculates its path while optimizing a shared formation objective, could enhance robustness and flexibility, especially in communication-denied environments.
Conclusion
This article has presented a holistic and robust framework for the route planning and control of a drone formation based on an enhanced Artificial Potential Field method. By systematically addressing the classical APF shortcomings—through a modified repulsive potential for guaranteed goal reachability, a stochastic perturbation method for escaping local minima, and an obstacle-linking strategy to suppress oscillations—we establish a reliable foundation for single-drone navigation. This core algorithm is then integrated with a formation potential field approach within a leader-follower architecture to maintain the desired geometric configuration. The critical issue of collective collision avoidance is solved elegantly via obstacle inflation, transforming the multi-agent problem into a single-agent planning task.
The simulation results validate the framework’s effectiveness, showing smooth, constraint-compliant paths for the leader, stable formation keeping and local obstacle avoidance for followers, and guaranteed clearance for the entire drone formation from all threats. The method maintains the computational efficiency and reactive nature of classical APF while overcoming its most significant drawbacks. This makes it a highly suitable candidate for the on-line, real-time path planning needs of autonomous drone formations operating in complex, obstacle-rich environments. The work provides a practical and extensible foundation upon which more advanced features, such as dynamic obstacle handling and fully decentralized control, can be built.