In recent years, the rapid development of artificial intelligence, robotics, sensing, communication, and control technologies has greatly expanded the application scope of UAV drones. As a flexible and autonomous platform, a UAV drone can operate in environments that are dangerous, remote, or difficult for humans to access. It can perform reconnaissance, disaster monitoring, logistics transportation, agricultural inspection, urban patrol, and emergency rescue tasks. In all these scenarios, path planning is one of the most critical technologies because it directly determines flight safety, mission efficiency, energy consumption, and the overall success rate of the UAV drone. In my research, I focus on the problem of UAV drone path planning in complex three-dimensional environments, especially complex terrain environments and complex multi-obstacle environments. I propose two improved methods: an adaptive hybrid particle swarm optimization algorithm for complex terrain, and a fusion method that combines particle swarm optimization with an artificial potential field for complex multi-obstacle environments.

The core motivation of my work is that a UAV drone must not only find a short path, but also maintain safe clearance from obstacles, respect flight height limits, avoid radar threats, and produce smooth trajectories that can be tracked by a flight controller. In a three-dimensional environment, the search space is extremely large, constraints are coupled, and the terrain or obstacle distribution may be highly irregular. Traditional path planning methods often struggle to balance global optimality, local feasibility, computational efficiency, and path smoothness. Therefore, I investigate how to improve the search capability of particle swarm optimization and how to combine it with artificial potential field methods to obtain reliable and executable paths for UAV drones.
1. Research Background and Significance
A UAV drone is an unmanned aerial vehicle that can fly autonomously or semi-autonomously. Its applications have expanded from military reconnaissance and target strike to civilian fields such as disaster relief, farmland monitoring, city inspection, and environmental monitoring. In military operations, a UAV drone can enter high-risk areas, avoid enemy radar, and perform reconnaissance or strike missions. In civil operations, a UAV drone can deliver medical supplies, monitor crops, inspect power lines, and collect real-time environmental data. These applications share a common requirement: the UAV drone must plan a safe and efficient path from a start point to a target point. In a two-dimensional plane, path planning is relatively simple, but in a three-dimensional environment, the UAV drone must consider terrain elevation, buildings, communication towers, no-fly zones, radar detection ranges, and dynamic obstacles. This makes path planning a high-dimensional, nonlinear, multi-constraint optimization problem.
In my view, the main challenges of UAV drone path planning in complex three-dimensional environments can be summarized as follows. First, there is a conflict between global optimality and local feasibility. The UAV drone needs a globally good path, but it also needs fast local reactions to avoid obstacles. Second, the environment is complex and uncertain. Obstacles may be irregularly distributed, and some may move over time. Third, the onboard computational resources of a UAV drone are limited, so the planning algorithm must be efficient. Fourth, the planned path must be smooth and dynamically feasible. A path with sharp turns or sudden altitude changes may be mathematically feasible but physically difficult for a UAV drone to follow. These challenges motivate my research on improved intelligent optimization and fusion methods.
2. Three-Dimensional Environment Modeling
To study UAV drone path planning, I first construct three-dimensional environment models. I use two main types of environments: mountainous terrain and multi-obstacle environments. The mountainous terrain model is generated by an exponential function that creates smooth peaks and valleys. The mathematical model is:
$$z(x,y)=\sum_{i=1}^{n} h_i \exp\left(-\left(\frac{x-x_i}{x_{si}}\right)^2-\left(\frac{y-y_i}{y_{si}}\right)^2\right)$$
Here, \(n\) is the number of peaks, \((x_i,y_i)\) is the center of the \(i\)-th peak, \(h_i\) is the peak height, and \(x_{si}\) and \(y_{si}\) control the slope decay in the \(x\) and \(y\) directions. I discretize the region \((x,y)\in[0,L_x]\times[0,L_y]\) into a regular grid and obtain a three-dimensional surface \((X,Y,Z)\). This terrain model is useful for testing whether a UAV drone can maintain safe altitude and avoid collisions with mountains.
For multi-obstacle environments, I model obstacles as vertical cylinders. This simplification is suitable for buildings, towers, and other artificial obstacles. The \(i\)-th cylinder is defined by its center \((x_i,y_i)\), radius \(R_i\), and height \(H_i\). The occupied region is:
$$O_i=\{(x,y,z)\mid (x-x_i)^2+(y-y_i)^2\le R_i^2,\;0\le z\le H_i\}$$
When there are \(m\) obstacles, the overall obstacle region is:
$$O=\bigcup_{i=1}^{m}O_i$$
By placing multiple cylinders with different positions, radii, and heights, I can construct complex obstacle clusters to test the avoidance performance of a UAV drone path planning algorithm. I also build dynamic environments where some obstacles move along elliptical trajectories with sinusoidal vertical motion. This allows me to evaluate whether a UAV drone can adapt to changing obstacle positions.
| Environment Type | Modeling Method | Key Parameters | Purpose |
|---|---|---|---|
| Mountainous terrain | Exponential peak function | Peak number, center, height, decay factors | Test terrain following and altitude safety |
| Static multi-obstacle | Vertical cylinders | Center, radius, height | Test obstacle avoidance and path length |
| Dynamic multi-obstacle | Moving cylinders and terrain | Trajectory, speed, vertical oscillation | Test adaptation and replanning capability |
3. Path Planning Constraint Modeling
For a UAV drone, path planning is not only about minimizing distance. I formulate a multi-objective cost function that includes path length, collision risk, flight height, path smoothness, and radar threat. This makes the planning result more realistic and safer.
3.1 Path Length Cost
The path length cost is the sum of Euclidean distances between consecutive waypoints. If the path consists of points \((X_1,Y_1,Z_1),\ldots,(X_n,Y_n,Z_n)\), then the distance between point \(i\) and point \(i+1\) is:
$$d(i,i+1)=\sqrt{(X_{i+1}-X_i)^2+(Y_{i+1}-Y_i)^2+(Z_{i+1}-Z_i)^2}$$
The total path length cost is:
$$f_1=\sum_{i=1}^{n-1}d(i,i+1)$$
3.2 Collision Cost
The collision cost penalizes paths that are too close to the terrain. Let \(z_i\) be the altitude of the path point and \(z_c(x_i,y_i)\) be the terrain height. The clearance is:
$$\Delta d_i=z_i-z_c(x_i,y_i)$$
The collision cost is:
$$f_2=
\begin{cases}
+\infty, & \exists i,\Delta d_i\le 0\\
\sum_{i=1}^{n}\frac{c_{\text{punkt}}}{\Delta d_i}, & \Delta d_i>0
\end{cases}$$
Here, \(c_{\text{punkt}}\) is a collision penalty coefficient. If the UAV drone collides with the terrain, the cost becomes infinite. Otherwise, the cost increases as the clearance decreases.
3.3 Flight Height Cost
The flight height cost encourages the UAV drone to fly within a safe altitude range. Let \(z_{\min}\) and \(z_{\max}\) be the minimum and maximum allowed altitudes, and let \(z_{\text{mean}}=(z_{\max}+z_{\min})/2\). The height cost is:
$$f_3=\sum_{i=1}^{n}p_{h,i}$$
$$p_{h,i}=
\begin{cases}
h_{\text{punkt}}\left(\frac{z_i-z_{\text{mean}}}{z_{\max}-z_{\min}}\right)^2, & z_{\min}<z_i<z_{\max}\\ &="" 100h_{\text{punkt}},=""
This term prevents the UAV drone from flying too low or too high.
3.4 Path Smoothness Cost
Smoothness is important because a UAV drone cannot make sharp turns or sudden climbs. I compute the turning angle and climb angle between consecutive segments. The smoothness cost is:
$$f_4=\sum_{i=2}^{n-1}(p_{\text{turn},i}+p_{\text{climb},i})$$
$$p_{\text{turn},i}=(\max(0,|\theta_{\text{turn},i}|-\theta_{\max}))^2\lambda_{\text{turn}}$$
$$p_{\text{climb},i}=(\max(0,|\phi_i-\phi_{i-1}|-\phi_{\max}))^2\lambda_{\text{climb}}$$
$$\theta_{\text{turn},i}=\arctan\left(\frac{\|p_1\times p_2\|}{p_1\cdot p_2}\right)$$
$$\phi_i=\arctan\left(\frac{Z_{i+1}-Z_i}{\sqrt{(X_{i+1}-X_i)^2+(Y_{i+1}-Y_i)^2}}\right)$$
Here, \(\theta_{\max}\) and \(\phi_{\max}\) are the maximum allowed turning and climbing angles, and \(\lambda_{\text{turn}}\) and \(\lambda_{\text{climb}}\) are penalty coefficients.
3.5 Radar Threat Cost
Radar threats are modeled as spherical detection regions with a safety buffer. For radar \(j\) at \((x_j,y_j,z_j)\) with radius \(r_j\), buffer \(\Delta s\), and threat coefficient \(\omega_{\text{radar}}\), the threat cost is:
$$f_5=\sum_{i=1}^{n}\sum_{j=1}^{m}
\begin{cases}
+\infty, & d_{i,j}\le r_j\\
\frac{\omega_{\text{radar}}}{d_{i,j}-r_j}, & r_jr_j+\Delta s
\end{cases}$$
$$d_{i,j}=\sqrt{(X_i-x_j)^2+(Y_i-y_j)^2+(Z_i-z_j)^2}$$
3.6 Total Cost Function
The total fitness function for a UAV drone path is a weighted sum:
$$\text{Fit}=\sum_{i=1}^{5}b_i f_i$$
where \(b_i\) are weights and \(\sum_{i=1}^{5}b_i=1\). In my experiments, I use \(\{0.1,0.3,0.3,0.2,0.1\}\) as the weight vector for path length, collision, height, smoothness, and radar threat, respectively.
| Cost Term | Symbol | Main Factors | Weight Example |
|---|---|---|---|
| Path length | \(f_1\) | Euclidean distance between waypoints | 0.1 |
| Collision | \(f_2\) | Terrain clearance | 0.3 |
| Flight height | \(f_3\) | Altitude deviation from safe center | 0.3 |
| Smoothness | \(f_4\) | Turning angle and climb angle | 0.2 |
| Radar threat | \(f_5\) | Distance to radar detection zones | 0.1 |
4. Adaptive Hybrid Particle Swarm Optimization
Particle swarm optimization is simple and easy to implement, but it can suffer from premature convergence and local optima in complex three-dimensional UAV drone path planning. To address this, I propose an adaptive hybrid particle swarm optimization algorithm, which I call AHAPSO. It combines adaptive inertia weight, mutation mechanism, artificial bee colony employed bee search, and logistic chaotic mapping.
4.1 Standard Particle Swarm Optimization
In standard PSO, each particle has a position and velocity. The velocity and position updates are:
$$v_k(t+1)=\omega v_k(t)+c_1r_1(\text{pbest}_k-x_k(t))+c_2r_2(\text{gbest}-x_k(t))$$
$$x_k(t+1)=x_k(t)+v_k(t+1)$$
Here, \(x_k(t)\) is the position of particle \(k\), \(v_k(t)\) is its velocity, \(\text{pbest}_k\) is its personal best, \(\text{gbest}\) is the global best, \(c_1\) and \(c_2\) are learning factors, and \(r_1,r_2\in[0,1]\) are random numbers.
4.2 Adaptive Inertia Weight
I replace the commonly used linear decreasing inertia weight with a nonlinear adaptive strategy:
$$\omega=\omega_1\exp\left(-\frac{\text{iter}}{N}\right)$$
where \(\omega_1\) is the initial inertia weight, \(\text{iter}\) is the current iteration, and \(N\) is the maximum iteration number. This allows the UAV drone path planner to explore widely in early iterations and refine locally in later iterations.
4.3 Artificial Bee Colony Employed Bee Strategy
To enhance local search, I integrate the employed bee phase of the artificial bee colony algorithm. A new candidate position is generated as:
$$x_{ij}^{\text{new}}=x_{ij}+\phi_{ij}(x_{ij}-x_{kj})$$
where \(x_{kj}\) is a randomly selected particle and \(\phi_{ij}\in[-1,1]\). This differential search helps the UAV drone path planner refine promising regions.
4.4 Mutation Mechanism
To maintain population diversity, I apply a small-probability mutation to particle velocities:
$$v_i^{\text{mut}}=\text{rand}(0,1)\cdot(v_{\max}-v_{\min})+v_{\min}$$
This mutation helps particles escape from local clusters and explore new areas of the path space.
4.5 Logistic Chaotic Mapping
In the later stage of iteration, I introduce logistic chaotic mapping to improve the ability to jump out of local optima. The chaotic value is:
$$cv=m\cdot \text{rand}\cdot(1-\text{rand})$$
The new position is:
$$p_n=p_s+cv\cdot(b_{\max}-b_{\min})$$
Here, \(p_s\) is the position before mapping, \(p_n\) is the new position, and \(b_{\max}\) and \(b_{\min}\) are the position bounds. The chaotic sequence provides non-repetitive and ergodic search behavior, which is beneficial for a UAV drone path planner in complex terrain.
| Component | Role in AHAPSO | Main Benefit for UAV Drone Path Planning |
|---|---|---|
| Adaptive inertia weight | Balance global and local search | Faster convergence and better refinement |
| Employed bee strategy | Differential local search | Improved local exploitation |
| Mutation mechanism | Random velocity perturbation | Increased population diversity |
| Logistic chaotic mapping | Ergodic position adjustment | Escaping local optima |
4.6 Algorithm Flow and Complexity
The AHAPSO algorithm initializes a population of particles, evaluates each particle using the total cost function, updates personal and global bests, adjusts the inertia weight, applies the employed bee search, performs mutation, and applies chaotic mapping in the later stage. The main computational complexity is:
$$O\left(NM(P+f)\right)$$
where \(M\) is the number of particles, \(P\) is the number of path points, \(N\) is the maximum iteration number, and \(f\) is the cost of fitness evaluation. The additional mutation and chaotic mapping steps add only \(O(M)\), so they do not dominate the overall complexity. This is important because a UAV drone has limited onboard computing resources.
5. Numerical Experiments and Analysis
To validate AHAPSO, I first use the CEC2017 benchmark suite. It contains 29 test functions, including unimodal, multimodal, hybrid, and composition functions. I compare AHAPSO with PSO, GWO, DBO, and SSA. Each algorithm runs 30 independent times, and I record the best, mean, and standard deviation.
| Algorithm | Publication Year | Key Parameters |
|---|---|---|
| DBO | 2022 | \(P_{\text{percent}}=0.2\), \(R(t)=1-t/M\), \(\theta=\pi k/180\) |
| GWO | 2014 | \(a=2-2t/T\), \(A=2ar_1-a\), \(C=2r_2\) |
| PSO | 1995 | \(w=0.8\), \(c_1=c_2=1.5\) |
| SSA | 2020 | \(P\%=0.2\), \(\rho=0.8\), \(N_{\text{SD}}=20\) |
| AHAPSO | Proposed | \(w=0.9\exp(-t/T)\), \(c_1=c_2=1.5\), \(p_m=0.1\), \(\mu=4\) |
The experimental results show that AHAPSO achieves the best average rank in the Friedman test. The average ranks are:
| Algorithm | Average Rank |
|---|---|
| PSO | 4.2241 |
| GWO | 2.9138 |
| DBO | 3.4310 |
| SSA | 3.2586 |
| AHAPSO | 1.1724 |
This indicates that AHAPSO is not only better on a few functions but also more consistent across different problem types. In the unimodal function F1, AHAPSO reaches a near-optimal value. In the multimodal function F6, AHAPSO avoids local optima and obtains a very small standard deviation. In the hybrid and composition functions F16, F20, F22, and F26, AHAPSO maintains fast convergence and high accuracy. The convergence curves and box plots confirm that AHAPSO has lower median values and narrower distributions than the compared algorithms.
| Function Type | Example | Observed AHAPSO Behavior |
|---|---|---|
| Unimodal | F1 | Fast early convergence, near-optimal result |
| Multimodal | F6 | Escapes local optima, small standard deviation |
| Hybrid | F16 | High precision and stable convergence |
| Composition | F22 | Strong adaptability and competitive accuracy |
5.1 Three-Dimensional Path Planning Simulation
I also test AHAPSO on two three-dimensional terrain maps. The start point is \((1,1,1)\) and the target point is \((1000,900,600)\). Map 1 has 9 peaks and 3 radars, while Map 2 has 12 peaks and 4 radars. The population size is 50, and the maximum iteration number is 300. I compare PSO, GWO, DBO, CLSPSO, SAPSO, and AHAPSO.
| Parameter | Value |
|---|---|
| Maximum iterations | 300 |
| Population size | 50 |
| Fitness weights | {0.1, 0.3, 0.3, 0.2, 0.1} |
| Map range | \(1000\times1000\times1000\) |
| Start point | (1, 1, 1) |
| Target point | (1000, 900, 600) |
| Map 1 peaks / radars | 9 / 3 |
| Map 2 peaks / radars | 12 / 4 |
In Map 1, PSO shows unnecessary oscillations, GWO has sharp turns, DBO has obvious detours, CLSPSO is tortuous, and SAPSO is smoother but still has adjustments near the target. AHAPSO produces a smooth path with small detours and gradual altitude changes. In Map 2, similar trends are observed. PSO has large fluctuations, GWO has bending and height changes, DBO has a local detour near the target, CLSPSO has a large spiral, and SAPSO is more direct but still less smooth than AHAPSO. AHAPSO reaches the target quickly and maintains a smooth and short path.
| Map | Metric | PSO | GWO | DBO | CLSPSO | SAPSO | AHAPSO |
|---|---|---|---|---|---|---|---|
| Map 1 | Worst | 1.395E+03 | 1.190E+03 | 1.290E+03 | 1.147E+03 | 1.256E+03 | 1.095E+03 |
| Map 1 | Best | 5.687E+02 | 7.605E+02 | 6.046E+02 | 9.245E+02 | 6.059E+02 | 5.427E+02 |
| Map 1 | Mean | 7.011E+02 | 8.160E+02 | 6.955E+02 | 9.267E+02 | 6.200E+02 | 5.896E+02 |
| Map 1 | STD | 238.298 | 90.494 | 157.211 | 15.576 | 69.078 | 81.888 |
| Map 2 | Worst | 1.997E+03 | 2.459E+03 | 2.265E+03 | 2.364E+03 | 1.774E+03 | 1.569E+03 |
| Map 2 | Best | 6.270E+02 | 9.766E+02 | 6.762E+02 | 9.139E+02 | 6.222E+02 | 5.700E+02 |
| Map 2 | Mean | 8.456E+02 | 1.005E+03 | 8.802E+02 | 9.520E+02 | 6.464E+02 | 6.255E+02 |
| Map 2 | STD | 372.253 | 116.971 | 233.695 | 155.213 | 125.261 | 123.435 |
These results confirm that AHAPSO provides better worst-case, best-case, and mean performance for UAV drone path planning in complex terrain. Its standard deviation is also relatively small, indicating stable and consistent planning results.
6. Fusion of Particle Swarm Optimization and Artificial Potential Field
In complex multi-obstacle environments, the traditional artificial potential field method often suffers from local minima, target unreachability, and path oscillation. To solve these problems, I propose a fusion method called FPAPF, which combines particle swarm optimization with an improved artificial potential field. The main idea is to use the potential field for global guidance and use PSO for local optimization of the next waypoint.
6.1 Gaussian Repulsion Function
The traditional repulsion force changes abruptly at the obstacle boundary. I replace it with a Gaussian repulsion function:
$$F_{\text{rep}}(q)=
\begin{cases}
K_{\text{rep}}\exp\left(-\frac{(\rho(q)-\varsigma)^2}{2\sigma^2}\right)\left(\frac{1}{\rho(q)}-\frac{1}{\varsigma}\right)\frac{1}{\rho(q)^2}\vec{d}, & \rho(q)\le\varsigma \text{ and } z\le h\\
0, & \text{otherwise}
\end{cases}$$
Here, \(K_{\text{rep}}\) is the repulsion coefficient, \(\rho(q)\) is the distance to the obstacle center, \(\varsigma\) is the influence range, \(\sigma\) is the Gaussian width, \(\vec{d}\) is the unit vector from the obstacle to the current position, \(z\) is the UAV drone altitude, and \(h\) is the obstacle height. The Gaussian term makes the repulsion smoother near the obstacle boundary and reduces unnecessary interference from distant obstacles.
6.2 Fusion Mechanism
The fusion mechanism has three stages. First, I compute the total force from the target attraction and obstacle repulsion. The attraction is:
$$F_{\text{att}}=K_{\text{att}}(x_{\text{goal}}-x_t)$$
The total force is:
$$F_{\text{total}}=F_{\text{att}}+F_{\text{rep}}$$
The potential-field guided position is:
$$x_{\text{apf}}=x_t+F_{\text{total}}$$
Second, I initialize a particle swarm around \(x_{\text{apf}}\) and perform local search. Each particle position is:
$$x_i=x_{\text{apf}}+\delta_i,\quad \delta_i\sim\mathcal{N}(0,\sigma^2I)$$
The fitness function is:
$$f(x_i)=\|x_i-x_{\text{goal}}\|+\sum_{k=1}^{N}P_k(x_i)$$
where \(P_k(x_i)\) is a penalty for being close to the \(k\)-th obstacle. The personal best and global best are updated as:
$$pbest_i=\arg\min_{t\in[1,T]}f(x_i^t)$$
$$gbest=\arg\min_i f(pbest_i)$$
Third, the final next position is a linear fusion:
$$x_{t+1}=\alpha x_{\text{apf}}+(1-\alpha)gbest$$
where \(\alpha\) balances global guidance and local search. In my experiments, I set \(\alpha=0.7\), \(K_{\text{att}}=0.04\), \(K_{\text{rep}}=0.10\), and \(\sigma=5\). This fusion allows the UAV drone to avoid local minima and target unreachability while maintaining a smooth and safe path.
| Method | Key Parameters | Main Feature |
|---|---|---|
| TAPF | \(K_{\text{att}}=0.04\), \(K_{\text{rep}}=0.10\) | Traditional potential field |
| RRT | \(\delta=1\), \(p_g=0.01\) | Sampling-based planning |
| IAPF | \(K_{\text{att}}=0.04\), \(K_{\text{rep}}=0.10\), \(\gamma=4\), \(\epsilon_f=0.1\) | Improved potential field |
| MAPF | \(K_{\text{att}}=0.04\), \(K_{\text{rep}}=0.10\), \(\zeta=3.5r\) | Modified potential field |
| GWOAPF | \(N=40\), \(T=10\), \(\eta=0.5\), \(\Delta z=2\) | Grey wolf plus potential field |
| FPAPF | \(w=0.7\), \(c_1=c_2=1.5\), \(\sigma=5\), \(\alpha=0.7\) | PSO plus Gaussian potential field |
7. Multi-Index Comprehensive Evaluation
To evaluate different UAV drone path planning algorithms fairly, I design a comprehensive evaluation system based on three indicators: path length \(L\), average safety distance \(D_{\text{avg}}\), and minimum safety distance \(D_{\min}\). Path length reflects efficiency. Average safety distance reflects overall clearance. Minimum safety distance reflects the most critical risk along the path. Since these indicators have different units and directions, I normalize them and compute a weighted score:
$$\text{score}=\omega_1\left(1-\frac{L}{L_{\max}}\right)+\omega_2\left(\frac{D_{\text{avg}}}{D_{\text{avg,max}}}\right)+\omega_3\left(\frac{D_{\min}}{D_{\min,\max}}\right)$$
where \(\omega_1+\omega_2+\omega_3=1\). I use the analytic hierarchy process to determine the weights. The pairwise comparison matrix is:
$$A=
\begin{bmatrix}
1 & 4 & 7\\
1/4 & 1 & 3\\
1/7 & 1/3 & 1
\end{bmatrix}$$
The maximum eigenvalue is \(\lambda_{\max}=3.032\), and the normalized weight vector is:
$$\omega=(0.705,0.211,0.084)^T$$
The consistency index is:
$$CI=\frac{\lambda_{\max}-3}{2}=0.016$$
With \(RI(3)=0.58\), the consistency ratio is:
$$CR=\frac{CI}{RI}=0.028<0.10$$
This confirms acceptable consistency. I round the weights to \((0.7,0.2,0.1)\) for the final score. This multi-index system allows me to compare UAV drone path planning algorithms from both efficiency and safety perspectives.
| Indicator | Symbol | Direction | Weight |
|---|---|---|---|
| Path length | \(L\) | Smaller is better | 0.7 |
| Average safety distance | \(D_{\text{avg}}\) | Larger is better | 0.2 |
| Minimum safety distance | \(D_{\min}\) | Larger is better | 0.1 |
8. Path Executability Verification Framework
A planned path may be geometrically feasible but not executable by a real UAV drone. Therefore, I design a path executability verification framework. The framework includes trajectory input, controller initialization, flight system simulation, environmental disturbance and error modeling, and performance evaluation. I use a simplified three-dimensional point-mass model for the UAV drone:
$$a_k=\frac{u_k+d_k}{m_k}$$
$$v_{k+1}=v_k+a_k\Delta t$$
$$p_{k+1}=p_k+v_{k+1}\Delta t$$
Here, \(p_k\) is the position, \(v_k\) is the velocity, \(u_k\) is the control input, \(d_k\) is external disturbance, \(m_k\) is a mass variation coefficient, and \(\Delta t\) is the sampling time. I compare three controllers: PID, LQR, and an enhanced LQR called ELQR.
8.1 Controller Models
For LQR, the discrete state equation and control law are:
$$X_{k+1}=AX_k+Bu_k$$
$$u_k=-K(X_k-X_k^{\text{ref}})$$
For PID, the control input is:
$$u_k=K_p e_k+K_i\sum_{j=0}^{k}e_j\Delta t+K_d\frac{e_k-e_{k-1}}{\Delta t}$$
$$e_k=x_k^{\text{ref}}-x_k$$
For ELQR, I add path preview and velocity feedforward:
$$u_k=-K(X_k-X_k^{\text{ref}})+\alpha(v_k^{\text{pred}}-v_k)$$
$$\alpha=\alpha_0+\frac{\alpha_1}{1+\|X_k-X_k^{\text{ref}}\|}$$
$$v_k^{\text{pred}}=\frac{x_{k+L}^{\text{ref}}-x_k^{\text{ref}}}{L\Delta t}$$
This allows the controller to react earlier to sharp turns or altitude changes.
| Controller | Main Mechanism | Expected Advantage for UAV Drone |
|---|---|---|
| PID | Proportional, integral, derivative feedback | Simple and widely used |
| LQR | State feedback with optimal gain | Stable tracking |
| ELQR | LQR plus preview and feedforward | Better tracking on sharp paths |
8.2 Performance Metrics
I use RMSE, maximum error, final error, mean jerk, maximum jerk, robustness, and disturbance-to-error ratio to evaluate tracking performance. The definitions are summarized below.
| Metric | Meaning | Direction |
|---|---|---|
| RMSE | Root mean square tracking error | Lower is better |
| Max_err | Maximum tracking error | Lower is better |
| Final_err | Final distance to target | Lower is better |
| Mean_jerk | Average rate of acceleration change | Lower is better |
| Max_jerk | Maximum rate of acceleration change | Lower is better |
| Robustness | Inverse of error variance | Higher is better |
| DAR | Disturbance-to-error ratio | Higher is better |
9. Simulation Results and Discussion
I test FPAPF in four environments. Environments 1, 2, and 3 are static multi-obstacle environments with 8, 10, and 12 cylindrical obstacles. Environment 4 is a dynamic environment with mountainous terrain, 45 static cylinders, and three moving obstacles. The start and target points vary across environments.
9.1 Static Environment Results
In Environment 1, FPAPF obtains the shortest path length of 236, which is 35%, 19%, 26%, 37%, and 33% shorter than TAPF, RRT, IAPF, MAPF, and GWOAPF, respectively. Its average safety distance is 12.42 and its minimum safety distance is 2.38. In Environment 2, FPAPF again obtains the shortest path length of 210 and a minimum safety distance of 2.03. In Environment 3, TAPF, IAPF, and MAPF fail to reach the target, while FPAPF succeeds with a path length of 276 and a minimum safety distance of 1.90. RRT and GWOAPF succeed but have lower minimum safety distances of 0.15 and 0.48.
| Environment | Algorithm | Success | Path Length | Average Safety Distance | Minimum Safety Distance |
|---|---|---|---|---|---|
| Env 1 | TAPF | Yes | 363 | 14.06 | 2.61 |
| Env 1 | RRT | Yes | 293 | 25.93 | 0.10 |
| Env 1 | IAPF | Yes | 320 | 19.33 | 3.99 |
| Env 1 | MAPF | Yes | 377 | 7.96 | 0.02 |
| Env 1 | GWOAPF | Yes | 351 | 12.81 | 1.56 |
| Env 1 | FPAPF | Yes | 236 | 12.42 | 2.38 |
| Env 2 | TAPF | Yes | 323 | 9.32 | 3.31 |
| Env 2 | RRT | Yes | 281 | 22.72 | 0.12 |
| Env 2 | IAPF | Yes | 289 | 19.03 | 4.05 |
| Env 2 | MAPF | Yes | 298 | 7.16 | 3.65 |
| Env 2 | GWOAPF | Yes | 225 | 13.50 | 0.36 |
| Env 2 | FPAPF | Yes | 210 | 10.35 | 2.03 |
| Env 3 | TAPF | No | – | – | – |
| Env 3 | RRT | Yes | 327 | 17.48 | 0.15 |
| Env 3 | IAPF | No | – | – | – |
| Env 3 | MAPF | No | – | – | – |
| Env 3 | GWOAPF | Yes | 411 | 8.04 | 0.48 |
| Env 3 | FPAPF | Yes | 276 | 9.40 | 1.90 |
The weighted score results show that FPAPF obtains the highest comprehensive score in Environments 1, 2, and 3. This means it balances path length and safety distance better than the other methods. IAPF sometimes has a larger minimum safety distance, but it often requires a longer path. RRT has a large average safety distance but a very small minimum safety distance, which indicates risk in some segments. MAPF and TAPF are often too close to obstacles or produce longer paths.
9.2 Dynamic Environment Results
In Environment 4, FPAPF obtains a path length of 224 and a minimum safety distance of 4.45. TAPF and MAPF produce shorter paths of 202 and 204, but their minimum safety distances are only 0.10 and 0.01. RRT and GWOAPF produce longer paths of 269 and 282 with minimum safety distances of 0.99 and 0.05. IAPF obtains a good safety distance of 17.65 but a longer path of 221. Overall, FPAPF achieves a good balance between safety and efficiency in the dynamic environment.
| Algorithm | Success | Path Length | Average Safety Distance | Minimum Safety Distance |
|---|---|---|---|---|
| TAPF | Yes | 202 | 6.70 | 0.10 |
| RRT | Yes | 269 | 9.42 | 0.99 |
| IAPF | Yes | 221 | 17.65 | 0.20 |
| MAPF | Yes | 204 | 18.28 | 0.01 |
| GWOAPF | Yes | 282 | 17.87 | 0.05 |
| FPAPF | Yes | 224 | 16.54 | 4.45 |
9.3 Path Executability Results
I use FPAPF paths and AHAPSO paths as reference trajectories. I track them with PID, LQR, and ELQR controllers under observation noise and mass disturbance. The results show that PID has larger tracking errors, especially near sharp turns and target regions. LQR and ELQR reduce the tracking error substantially. ELQR performs best in terms of RMSE, maximum error, final error, robustness, and disturbance rejection. For the FPAPF path, ELQR achieves an RMSE of 3.5319, while LQR achieves 3.9034 and PID achieves 24.658. For the AHAPSO path, ELQR achieves an RMSE of 27.078, while LQR achieves 28.722 and PID achieves 184.60.
| Path | Controller | RMSE | Max Error | Final Error | Mean Jerk | Max Jerk | Robustness | DAR |
|---|---|---|---|---|---|---|---|---|
| FPAPF | PID | 24.658 | 37.987 | 27.807 | 19.520 | 32.876 | 0.0079532 | 0.00042734 |
| FPAPF | LQR | 3.9034 | 6.3940 | 5.1163 | 37.795 | 108.35 | 0.28789 | 0.015143 |
| FPAPF | ELQR | 3.5319 | 5.8692 | 4.8053 | 35.651 | 114.67 | 0.33678 | 0.020346 |
| AHAPSO | PID | 184.60 | 258.92 | 189.06 | 151.57 | 240.68 | 0.00027275 | 0.0000090055 |
| AHAPSO | LQR | 28.722 | 53.619 | 33.264 | 183.98 | 853.07 | 0.0058088 | 0.00035242 |
| AHAPSO | ELQR | 27.078 | 51.023 | 31.447 | 177.50 | 822.51 | 0.0065847 | 0.00033593 |
The jerk distributions show that most jerk values are concentrated in low ranges, indicating that the controllers can suppress aggressive acceleration changes. The AHAPSO path has a wider jerk distribution because it is generated in a larger terrain environment with greater altitude changes. The FPAPF path has a narrower jerk distribution because the obstacle environment is smaller and the path is relatively smooth. In both cases, ELQR provides better tracking accuracy than PID and LQR, which indicates that preview and feedforward compensation are useful for complex three-dimensional UAV drone trajectories.
10. Conclusions and Future Work
In my research, I studied UAV drone path planning in complex three-dimensional environments. I proposed an adaptive hybrid particle swarm optimization algorithm for complex terrain and a fusion of particle swarm optimization and artificial potential field for complex multi-obstacle environments. The first method improves the search accuracy and convergence stability of PSO by using adaptive inertia weight, mutation, employed bee search, and logistic chaotic mapping. The second method improves the artificial potential field by using a Gaussian repulsion function and uses PSO to optimize key path points, which helps avoid local minima and target unreachability.
I also constructed a multi-index evaluation system with path length, average safety distance, and minimum safety distance. I used AHP to determine the weights and verified consistency. Furthermore, I designed a path executability verification framework using PID, LQR, and ELQR controllers. The results show that the paths generated by my methods can be tracked with acceptable accuracy and smoothness, and ELQR performs best in tracking complex trajectories.
For future work, I plan to deploy the algorithms on real UAV drone platforms and test them with onboard sensors and flight controllers. I will also consider dynamic obstacles and online replanning for real-time adaptation. Another direction is multi-UAV drone cooperative path planning, where information sharing and conflict resolution are important. Finally, I will introduce more detailed dynamic constraints and learning-based parameter adaptation to further improve the executability and robustness of planned paths for UAV drones in complex environments.
| Research Direction | Current Limitation | Future Improvement |
|---|---|---|
| Real-world deployment | Simulation only | Onboard flight tests with real UAV drone |
| Dynamic adaptation | Limited dynamic scenarios | Online replanning and moving obstacle prediction |
| Multi-UAV cooperation | Single UAV drone focus | Cooperative planning and conflict resolution |
| Dynamics modeling | Simplified point-mass model | Full dynamic model and learning-based adaptation |
Overall, my study shows that improved particle swarm optimization and its fusion with artificial potential field can provide effective solutions for UAV drone path planning in complex three-dimensional environments. The combination of intelligent optimization, safety evaluation, and controller-based executability verification offers a practical direction for future autonomous UAV drone navigation.
</z_i
