Improved Particle Swarm Optimization and Its Fusion with Artificial Potential Fields for UAV Path Planning

Over the past decade, unmanned aerial vehicles, widely referred to as UAV drones, have gradually moved from specialized laboratories into everyday engineering practice. The rapid development of artificial intelligence, miniature sensing hardware, wireless communication and modern control theory has made UAV drones a flexible platform capable of replacing humans in dangerous, repetitive or inaccessible tasks. In military reconnaissance, disaster monitoring, agricultural surveying, urban inspection and logistics delivery, UAV drones have demonstrated remarkable operational value. Nevertheless, the usefulness of any UAV drones mission ultimately depends on whether the vehicle can travel from its launch point to its destination safely, efficiently and repeatedly. That requirement places path planning at the very center of autonomous flight.

Path planning for UAV drones is the process of determining a feasible flight trajectory that satisfies mission requirements and physical constraints while avoiding obstacles and threats. A good trajectory is not simply the shortest curve connecting two points; it must simultaneously respect terrain clearance, kinematic limits, energy budgets and threat exposure. In flat two-dimensional settings the problem is conceptually simple, but when altitude becomes a decision variable the difficulty grows rapidly. In a three-dimensional environment, UAV drones must cope with terrain undulation, vertical obstacles such as towers and buildings, restricted corridors and communication shadow zones. Consequently, three-dimensional path planning is a high-dimensional, non-linear, strongly constrained optimization problem whose solution space is enormous and whose constraints are mutually coupled.

In practical deployments, path planning for UAV drones faces four recurring tensions. The first is the conflict between global optimality and local feasibility: a planner should approach a globally attractive route while still reacting quickly to nearby hazards. The second is environmental complexity: obstacles in mountainous, forested or urban regions are irregular, and they may move. The third is the trade-off between computational cost and real-time responsiveness, since the onboard processors of small UAV drones are far weaker than desktop hardware. The fourth is path smoothness and flight feasibility: UAV drones cannot turn or climb arbitrarily, so routes containing sharp corners may be theoretically valid yet physically unrealizable. These tensions motivated the research presented in this article.

2. Research Background and Objectives

My research addresses the path planning problem for UAV drones operating in complex three-dimensional environments. I approach the problem from two complementary angles: complex terrain environments, in which elevation and threat fields dominate, and complex multi-obstacle environments, in which dense vertical obstacles dominate and local minima become the principal difficulty. Two improved planners are proposed and validated. The first is an adaptive hybrid particle swarm optimizer designed for terrain-constrained three-dimensional route generation. The second is a fused planner that combines particle swarm optimization with a Gaussian-modified artificial potential field, supported by a multi-index evaluation system and a trajectory executability verification framework.

The specific objectives of this work are summarized in Table 1.

Objective Challenge Addressed Proposed Countermeasure Validation Strategy
Improve global search accuracy in rough terrain Premature convergence of particle swarm optimization Adaptive inertia weight, mutation, employed-bee search, chaotic mapping Benchmark function suite plus three-dimensional terrain simulation
Eliminate local minima in dense obstacle fields Zero net force away from the goal Gaussian repulsion plus particle swarm refinement of key waypoints Static and dynamic obstacle scenarios
Balance path length against safety Single-criterion evaluation bias Weighted multi-index score derived analytically Cross-comparison of six planners
Confirm that planned routes can actually be flown Geometric feasibility does not imply control feasibility Closed-loop tracking simulation with multiple controllers Error, smoothness and robustness metrics

3. Related Work

Research on path planning for UAV drones can be grouped into four broad families: traditional planning methods, intelligent optimization methods, learning-based methods and hybrid methods.

3.1 Traditional Planning Methods

Traditional planners include graph search, geometric construction, sampling and potential field techniques. Graph search discretizes the environment; the A* algorithm extends Dijkstra’s method with a heuristic that combines accumulated cost and estimated remaining distance, which sharpens the search direction. In low-dimensional settings these algorithms perform well, but node counts explode as dimensionality rises. Geometric methods exploit spatial structure directly, most notably visibility graphs and Voronoi diagrams, which naturally maintain clearance from obstacles. Sampling methods such as the rapidly exploring random tree explore continuous spaces efficiently but often produce jagged routes with many inflection points, requiring post-processing smoothing.

Potential field methods, especially the artificial potential field, model the goal as an attractor and obstacles as repulsors so that the vehicle follows the gradient of a synthetic energy landscape. Their appeal lies in low computational cost and natural real-time avoidance. Their weakness is equally well known: spurious local minima, oscillation near the goal, and no guarantee of global optimality. Table 2 summarizes representative traditional approaches.

Family Representative Technique Strength Limitation for UAV Drones
Graph search A*, Dijkstra Complete and interpretable Combinatorial growth in three dimensions
Geometric Visibility graph, Voronoi diagram Explicit clearance reasoning Sensitive to obstacle count and dynamic change
Sampling Rapidly exploring random tree Handles high-dimensional spaces Jagged paths, stochastic quality
Potential field Artificial potential field Cheap, smooth, reactive Local minima and unreachable goals

3.2 Intelligent Optimization Methods

Swarm intelligence and evolutionary algorithms reformulate path planning as a numerical optimization problem. Particle swarm optimization is lightweight and has few parameters, which explains its popularity for UAV drones trajectory design, yet it is prone to stagnation in complex landscapes. Ant colony optimization uses pheromone feedback and performs well in planar problems but converges slowly and risks premature pheromone concentration in three dimensions. Genetic algorithms offer strong global exploration at the price of high computational cost and parameter sensitivity. Simulated annealing escapes local optima probabilistically but depends heavily on cooling schedules. Table 3 compares these families.

Algorithm Search Mechanism Primary Advantage Primary Weakness
Particle swarm optimization Velocity update guided by personal and global bests Simple, fast, easy to hybridize Early stagnation, loss of diversity
Ant colony optimization Pheromone accumulation with evaporation Positive feedback, good in 2D Slow convergence in 3D
Genetic algorithm Selection, crossover, mutation Strong global exploration Costly, parameter sensitive
Simulated annealing Probabilistic acceptance of worse solutions Escapes local optima Cooling schedule dependence
Artificial bee colony Employed, onlooker and scout bee phases Good balance of exploration and exploitation Slow neighbourhood search

3.3 Learning-Based Methods

Reinforcement learning lets UAV drones acquire decision policies through interaction. State variables describe vehicle position and nearby obstacles, actions correspond to flight commands, and rewards encode distance, energy and safety. Q-learning and its deep variants have been applied to grid-based avoidance and multi-agent coverage, while dynamic variants adapt the iteration budget to reduce redundant computation. Imitation learning instead imitates expert demonstrations, converting planning into supervised regression and producing smooth trajectories quickly. Nevertheless, reward design remains fragile, and both paradigms demand substantial training effort before deployment on real UAV drones.

3.4 Hybrid Methods

Because no single algorithm simultaneously guarantees global quality, local reactivity and real-time behaviour, hybridization has become the dominant research direction. Three combinations dominate the literature: traditional planners fused with optimizers, traditional planners fused with learning, and optimizer-to-optimizer fusion. Potential field guidance combined with swarm optimization is particularly attractive for UAV drones, because the potential field supplies directional bias while the swarm supplies global search. My own work follows this philosophy but strengthens it in two ways: the potential field itself is smoothed with a Gaussian modulation, and the swarm is embedded as a local refinement stage that acts on key waypoints rather than replacing the field entirely.

4. Three-Dimensional Environment Modeling

Before any planner can be evaluated, the environment must be represented in a form that is both realistic and computable. I adopt two complementary environment models, one for undulating terrain and one for dense vertical obstacles.

4.1 Mountain Terrain Model

Natural terrain is generated as a superposition of exponential peaks. The elevation at a horizontal position is

$$z(x,y)=\sum_{i=1}^{n} h_i \exp\!\left[-\left(\frac{x-x_i}{x_{s_i}}\right)^{2}-\left(\frac{y-y_i}{y_{s_i}}\right)^{2}\right],$$

where \(n\) is the number of peaks, \((x_i,y_i)\) is the centre of the \(i\)-th peak, \(h_i\) is its height, and \(x_{s_i}\) and \(y_{s_i}\) control the decay rate along each axis. The continuous surface is sampled over a rectangular domain and mapped onto a regular grid, producing a uniform three-dimensional coordinate matrix that the planner can query efficiently. This formulation allows me to vary terrain severity simply by changing peak heights and decay parameters, which is convenient for stress testing UAV drones planners.

4.2 Multi-Obstacle Model

Artificial obstacles such as buildings, towers and masts are approximated as vertical cylinders. The volume occupied by the \(i\)-th obstacle is

$$O_i=\left\{(x,y,z)\;\middle|\;(x-x_i)^{2}+(y-y_i)^{2}\le R_i^{2},\;0\le z\le H_i\right\},$$

and the total obstacle region is the union \(O=\bigcup_{i=1}^{m}O_i\). Each cylinder is therefore defined by a horizontal centre, a radius and a height. By distributing cylinders of different radii and heights I can construct sparse, moderate and dense obstacle fields, which I use to examine how different planners for UAV drones degrade as congestion increases.

4.3 Dynamic Environment

To move closer to realistic operating conditions, I also construct a dynamic scenario that superimposes three ingredients: a mountainous base surface, several clusters of static cylinders, and moving obstacles that travel along horizontal elliptical trajectories while oscillating vertically. This mixture forces UAV drones to combine long-horizon planning with short-horizon reactivity.

5. Algorithmic Foundations

5.1 Particle Swarm Optimization

Particle swarm optimization maintains a population of candidate solutions. Each particle carries a position and a velocity, and the velocity update is

$$v_k(t+1)=\omega\,v_k(t)+c_1 r_1\left(p_{\text{best},k}-x_k(t)\right)+c_2 r_2\left(g_{\text{best}}-x_k(t)\right),$$

followed by the position update

$$x_k(t+1)=x_k(t)+v_k(t+1).$$

Here \(\omega\) is the inertia weight, \(c_1\) and \(c_2\) are learning factors, and \(r_1, r_2 \sim U(0,1)\) are random coefficients. The inertia weight is commonly scheduled linearly,

$$\omega=\omega_{\max}-\frac{(\omega_{\max}-\omega_{\min})\,t}{T_{\max}},$$

but this rigid schedule cannot adapt to the changing landscape that UAV drones encounter in three dimensions.

5.2 Artificial Bee Colony

The artificial bee colony algorithm simulates the foraging behaviour of three bee castes. A candidate food source is a vector \(X_i=[x_{i1},\dots,x_{iD}]\). Employed bees generate neighbours through

$$x’_{ij}=x_{ij}+\phi_{ij}\left(x_{ij}-x_{kj}\right),\qquad j\in\{1,\dots,D\},$$

where \(k\neq i\) is a randomly chosen companion and \(\phi_{ij}\in[-1,1]\). Onlooker bees select sources using fitness-proportional selection after converting objective values according to

$$\text{fit}_i=\begin{cases}\dfrac{1}{1+f_i}, & f_i\ge 0,\\[6pt] 1+\lvert f_i\rvert, & f_i<0,\end{cases}\qquad p_i=\frac{\text{fit}_i}{\sum_{m=1}^{N}\text{fit}_m}.$$

Scout bees replace exhausted sources with random samples

$$x_{ij}=x_j^{\min}+r\left(x_j^{\max}-x_j^{\min}\right).$$

5.3 Artificial Potential Field

The artificial potential field composes an attractive term and a repulsive term,

$$U(q)=U_{\text{att}}(q)+U_{\text{rep}}(q),$$

with a quadratic attraction

$$U_{\text{att}}(q)=\tfrac{1}{2}k_{\text{att}}\lVert q-q_{\text{goal}}\rVert^{2},\qquad F_{\text{att}}(q)=-k_{\text{att}}\left(q-q_{\text{goal}}\right),$$

and a classical repulsion that activates only within a threshold distance \(\rho_0\):

$$U_{\text{rep}}(q)=\begin{cases}\dfrac{1}{2}k_{\text{rep}}\left(\dfrac{1}{\rho(q)}-\dfrac{1}{\rho_0}\right)^{2}, & \rho(q)\le \rho_0,\\[8pt] 0, & \rho(q)>\rho_0.\end{cases}$$

The resulting repulsive force is

$$F_{\text{rep}}(q)=\begin{cases}k_{\text{rep}}\left(\dfrac{1}{\rho(q)}-\dfrac{1}{\rho_0}\right)\dfrac{1}{\rho^{2}(q)}\nabla\rho(q), & \rho(q)\le \rho_0,\\[8pt] 0, & \rho(q)>\rho_0.\end{cases}$$

The total force acting on the UAV drones is

$$F_{\text{total}}=F_{\text{att}}+\sum_{i=1}^{n}F_{\text{rep}}^{(i)}.$$

Two pathological behaviours arise from this construction. A local minimum occurs at any point \(q^{*}\neq q_{\text{goal}}\) where \(\nabla U(q^{*})=0\) and \(U(q^{*})\) is lower than all neighbouring points, which happens whenever attraction and repulsion cancel. Goal unreachability occurs when the goal lies inside a repulsive zone and the repulsive force exceeds the attraction, so that \(\lim_{t\to\infty}\lVert s(t)-q_{\text{goal}}\rVert \neq 0\) along the vehicle trajectory \(s(t)\). Both defects are especially damaging for UAV drones flying through cluttered corridors.

6. Adaptive Hybrid Particle Swarm Optimization for Terrain-Constrained Planning

My first contribution is an adaptive hybrid particle swarm optimizer that augments the standard swarm with four cooperating mechanisms.

6.1 Cost Function Modeling

The planning problem is expressed as the minimization of a weighted sum of five cost terms. The path length cost accumulates Euclidean segment lengths:

$$d(i,i+1)=\sqrt{(X_{i+1}-X_i)^{2}+(Y_{i+1}-Y_i)^{2}+(Z_{i+1}-Z_i)^{2}},\qquad f_1=\sum_{i=1}^{n-1}d(i,i+1).$$

The collision cost penalizes insufficient terrain clearance, with \(\Delta d_i=z_i-z_c(x_i,y_i)\) denoting the vertical gap between the waypoint and the terrain:

$$f_2=\begin{cases}+\infty, & \exists\, i,\ \Delta d_i\le 0,\\[4pt] \displaystyle\sum_{i=1}^{n}\frac{c_{\text{punkt}}}{\Delta d_i}, & \Delta d_i>0.\end{cases}$$

The altitude cost keeps UAV drones near an ideal cruise band centred between the allowed minimum and maximum heights:

$$f_3=\sum_{i=1}^{n}p_{h,i},\qquad p_{h,i}=\begin{cases}h_{\text{punkt}}\left(\dfrac{z_i-z_{\text{mean}}}{z_{\max}-z_{\min}}\right)^{2}, & z_{\min}<z_i<z_{\max},\\[8pt] &="" 100\,h_{\text{punkt}},=""

with \(z_{\text{mean}}=(z_{\max}+z_{\min})/2\). The smoothness cost penalizes excessive turning and climbing:

$$f_4=\sum_{i=2}^{n-1}\left(p_{\text{turn},i}+p_{\text{climb},i}\right),$$

$$p_{\text{turn},i}=\left(\max\left(0,\lvert\theta_{\text{turn},i}\rvert-\theta_{\max}\right)\right)^{2}\lambda_{\text{turn}},\qquad p_{\text{climb},i}=\left(\max\left(0,\lvert\varphi_i-\varphi_{i-1}\rvert-\varphi_{\max}\right)\right)^{2}\lambda_{\text{climb}},$$

where the turning angle and climbing angle are computed as

$$\theta_{\text{turn},i}=\arctan\!\left(\frac{\lVert p_1\times p_2\rVert}{p_1\cdot p_2}\right),\qquad \varphi_i=\arctan\!\left(\frac{Z_{i+1}-Z_i}{\sqrt{(X_{i+1}-X_i)^{2}+(Y_{i+1}-Y_i)^{2}}}\right).$$

Finally, the radar threat cost discourages UAV drones from entering detection volumes, where \(d_{i,j}\) is the distance from waypoint \(i\) to radar \(j\) of radius \(r_j\), \(\Delta s\) is a safety buffer and \(\omega_{\text{radar}}\) scales the penalty:

$$f_5=\sum_{i=1}^{n}\sum_{j=1}^{m}\begin{cases}+\infty, & d_{i,j}\le r_j,\\[4pt] \dfrac{\omega_{\text{radar}}}{d_{i,j}-r_j}, & r_jr_j+\Delta s.\end{cases}$$

The aggregate fitness is

$$\text{Fit}=\sum_{i=1}^{5}b_i f_i,\qquad \sum_{i=1}^{5}b_i=1,$$

which converts three-dimensional planning for UAV drones into a multi-constraint optimization problem in which distance, safety, smoothness and threat exposure are considered jointly. The weight configuration used throughout my experiments is listed in Table 4.

Term Meaning Weight Behaviour When Violated
\(f_1\) Accumulated path length 0.1 Longer detours raise cost linearly
\(f_2\) Terrain collision penalty 0.3 Collision becomes infinite cost
\(f_3\) Altitude band deviation 0.3 Unsafe altitude heavily penalized
\(f_4\) Turning and climbing smoothness 0.2 Sharp manoeuvres accumulate penalty
\(f_5\) Radar threat exposure 0.1 Detection volume becomes infinite cost

6.2 Improved Mechanisms

Employed-bee integration. I embed the employed-bee neighbourhood rule directly into the swarm. Each particle produces a candidate by

$$x_{ij}^{\text{new}}=x_{ij}+\phi_{ij}\left(x_{ij}-x_{kj}\right),$$

which introduces a differential perturbation drawn from a randomly selected companion particle. This reinforces fine-grained exploitation around promising solutions, which is precisely where standard swarms lose resolution when planning for UAV drones.

Non-linear adaptive inertia. Instead of the conventional linear decay I adopt

$$\omega=\omega_1\exp\!\left(-\frac{\text{iter}}{N}\right),$$

where \(\omega_1\) is the initial inertia, \(\text{iter}\) is the current generation and \(N\) is the generation budget. The early phase favours exploration with a large inertia, while the late phase reduces the effective step length so that particles settle accurately near the global best. This schedule gives a smoother transition than linear decay and avoids an abrupt loss of mobility.

Mutation. To counteract clustering, a small fraction of particles is mutated with probability \(p_m\). The mutated velocity is

$$v_i^{\text{mut}}=\text{rand}(0,1)\left(v_{\max}-v_{\min}\right)+v_{\min},$$

which re-injects diversity and lets the population examine regions that would otherwise be abandoned.

Chaotic mapping. After seventy-five percent of the generation budget has elapsed, I apply a logistic chaotic map to the population:

$$c_v=m\cdot r\left(1-r\right),\qquad p_n=p_s+c_v\left(b_{\max}-b_{\min}\right),$$

where \(m\) is the control parameter, \(p_s\) is the pre-map position and \(p_n\) is the post-map position, clipped to the search bounds. The ergodicity and aperiodicity of the logistic sequence let the swarm redistribute itself over the decision space, which reduces the probability of permanent entrapment in a local basin. The cooperation of these four mechanisms across the search horizon is summarized in Table 5.

Mechanism Search Phase Principal Contribution Risk Mitigated
Adaptive inertia weight Whole run Smooth exploration-to-exploitation transition Overshooting and premature settling
Employed-bee search Whole run Neighbourhood refinement around good solutions Insufficient local resolution
Mutation Whole run Diversity injection Population collapse
Logistic chaotic mapping Late stage Ergodic redistribution Local optimum entrapment

6.3 Complexity Analysis

Let \(M\) denote population size, \(P\) the number of waypoints per particle, \(N\) the generation budget and \(f\) the cost of one fitness evaluation. The swarm update costs \(O(M(P+f))\) per generation, so the swarm phase totals \(O(NM(P+f))\). Mutation adds \(O(M)\), the employed-bee stage adds \(O(M(P+f))\) per generation, and chaotic mapping adds \(O(M)\). The dominant term therefore remains

$$O\!\left(NM(P+f)\right),$$

which confirms that the auxiliary mechanisms improve solution quality for UAV drones without altering the asymptotic cost of the underlying optimizer.

6.4 Benchmark Validation

I evaluated the proposed optimizer on a modern benchmark suite comprising unimodal, multimodal, hybrid and composition functions, always with thirty independent runs, a population of thirty and one thousand generations. To isolate the contribution of each mechanism I constructed progressive variants: one that adds only the adaptive inertia, one that adds the employed-bee stage, one that adds chaotic mapping, and the full algorithm that additionally includes mutation. Table 6 reports representative outcomes of this ablation study using mean and standard deviation as criteria.

Function Base Swarm Mean Adaptive Inertia Mean Plus Bee Stage Mean Plus Chaos Mean Full Algorithm Mean
Unimodal A 4.19E+08 5.04E+08 5.53E+08 5.48E+08 3.77E+03
Unimodal B 2.09E+04 2.47E+04 1.32E+03 3.17E+03 2.94E+03
Multimodal A 6.39E+02 6.20E+02 6.10E+02 6.10E+02 6.07E+02
Multimodal B 3.73E+03 2.18E+03 1.41E+03 1.41E+03 1.20E+03
Hybrid A 2.68E+03 2.57E+03 2.44E+03 2.44E+03 2.39E+03
Composition A 4.32E+03 4.14E+03 3.34E+03 3.34E+03 3.17E+03

The pattern is consistent across the suite. Adaptive inertia alone yields modest gains, since it mainly reshapes the tempo of the search. Adding the bee-inspired neighbourhood stage improves exploitation around good regions and benefits the late stages of optimization. Mutation produces the most visible improvement, particularly on hybrid and composition landscapes, because it restores the population variance that the swarm tends to lose. Chaotic mapping contributes a smaller but still measurable benefit, chiefly by diversifying the search routes and delaying stagnation.

I then compared the full method against four established metaheuristics: the standard particle swarm optimizer, the grey wolf optimizer, the dung beetle optimizer and the sparrow search algorithm, always with thirty particles and one thousand generations. Table 7 presents the average Friedman ranks obtained across the full benchmark set.

Algorithm Mean Friedman Rank Interpretation
Particle swarm optimization 4.2241 Weakest overall positioning
Grey wolf optimizer 2.9138 Moderate performance
Dung beetle optimizer 3.4310 Moderate but inconsistent
Sparrow search algorithm 3.2586 Competitive on some families
Proposed hybrid optimizer 1.1724 Best overall positioning

Because the benchmark suite includes unimodal, multimodal, hybrid and composition problems, a low average rank indicates that the advantage is not confined to a narrow class of landscape. The distribution of objective values across repeated runs is also narrower for the proposed method, which indicates that the routes it produces for UAV drones are consistent rather than occasionally lucky.

6.5 Three-Dimensional Terrain Results

Two terrain maps were created for the flight experiments. The first contains nine peaks and three radar threat volumes, while the second contains twelve peaks and four radar sites. Both share a cubic workspace of one thousand units per side, with a common start position and a goal positioned high above the far corner. Fifty particles were used with three hundred generations. Table 8 summarizes the statistical performance of the six planners compared in this study.

Map Statistic Swarm Grey Wolf Dung Beetle Chaotic Swarm Annealed Swarm Proposed
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 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

The proposed optimizer obtains the best worst-case, best-case and average values on both maps. The visual character of the routes also differs. The standard swarm produces wide lateral swings because its large early inertia keeps pushing the particles past the ideal corridor. The grey wolf optimizer generates acceptable detours but introduces pronounced bends and steep climbs near the tallest peak. The dung beetle optimizer achieves avoidance but with visible zigzag structure, which would be uncomfortable for real UAV drones. The chaotic variant converges slowly and produces long meandering sections. The annealed variant performs well in aggregate but still shows residual oscillation near the goal. In contrast, the proposed method returns routes that leave the start region promptly, thread between peaks with limited lateral deviation, and approach the goal along a nearly straight final segment. The convergence curves show the same ordering, with the proposed method dropping steeply in the first few dozen generations and remaining stable afterwards, while the grey wolf optimizer flattens and stalls midway.

7. Fusion of Particle Swarm Optimization and Artificial Potential Fields

My second contribution targets dense obstacle fields, where the potential field component of a planner is likely to become trapped. The fused planner keeps the potential field as the primary directional engine but adds a swarm-based local refinement stage and replaces the classical repulsion with a smoother alternative.

7.1 Gaussian Repulsion

The classical repulsive term is discontinuous at the boundary of the influence radius, which produces abrupt force changes and causes oscillation when UAV drones skim an obstacle surface. I therefore multiply the repulsive magnitude by a Gaussian envelope, giving

$$F_{\text{rep}}(q)=\begin{cases}K_{\text{rep}}\exp\!\left(-\dfrac{(\rho(q)-\varsigma)^{2}}{2\sigma^{2}}\right)\left(\dfrac{1}{\rho(q)}-\dfrac{1}{\varsigma}\right)\dfrac{1}{\rho^{2}(q)}\,\vec{d}, & \rho(q)\le \varsigma \ \text{and}\ z\le h,\\[10pt] 0, & \text{otherwise},\end{cases}$$

where \(\varsigma\) is the influence radius, \(\sigma\) is the Gaussian width that governs the decay rate, \(\vec{d}\) is the unit vector pointing from the obstacle towards the vehicle, \(z\) is the current altitude and \(h\) is the obstacle height. The envelope makes the force rise and fall gradually, so the transition near the obstacle boundary is smooth rather than abrupt. The altitude condition also ensures that an obstacle ceases to repel once the UAV drones have climbed above its top surface, which removes a large class of spurious repulsive contributions.

7.2 Fusion Mechanism

The fused planner proceeds in three stages at every waypoint. First, the potential field yields a guiding direction. The attractive component is linear,

$$F_{\text{att}}=K_{\text{att}}\left(x_{\text{goal}}-x_t\right),$$

and the total force is the sum of attraction and the Gaussian repulsion,

$$F_{\text{total}}=F_{\text{att}}+F_{\text{rep}},\qquad x_{\text{apf}}=x_t+F_{\text{total}}.$$

Second, a swarm is initialized in the neighbourhood of the guiding point,

$$x_i=x_{\text{apf}}+\delta_i,\qquad \delta_i\sim\mathcal{N}\!\left(0,\sigma^{2}I\right),$$

and each candidate is scored with a fitness that combines goal proximity and obstacle penalties,

$$f(x_i)=\lVert x_i-x_{\text{goal}}\rVert+\sum_{k=1}^{N}P_k(x_i).$$

Personal and global bests are maintained as

$$p_{\text{best}}^{(i)}=\arg\min_{t\in[1,T]}f\!\left(x_t^{(i)}\right),\qquad g_{\text{best}}=\arg\min_{i}f\!\left(p_{\text{best}}^{(i)}\right).$$

Third, the guiding point and the swarm best are blended linearly,

$$x_{t+1}=\alpha\,x_{\text{apf}}+(1-\alpha)\,g_{\text{best}},$$

where the mixing coefficient \(\alpha\) regulates the relative influence of field guidance and swarm refinement. A larger \(\alpha\) preserves the smooth directional bias of the potential field, while a smaller value allows the swarm to exploit local opportunities. Because the swarm can explore laterally around a force-balance point, it effectively breaks the symmetry that creates local minima, which is the principal failure mode for UAV drones flying through clustered obstacles. Table 9 lists the parameters used for the fused planner.

Parameter Symbol Value Role
Attractive gain \(K_{\text{att}}\) 0.04 Strength of goal attraction
Repulsive gain \(K_{\text{rep}}\) 0.10 Strength of obstacle repulsion
Inertia weight \(w\) 0.70 Velocity persistence of refinement swarm
Learning factors \(c_1, c_2\) 1.5 Cognitive and social attraction
Gaussian width \(\sigma\) 5 Decay rate of the repulsive envelope
Mixing coefficient \(\alpha\) 0.70 Balance between field guidance and swarm refinement

7.3 Multi-Index Evaluation System

Because a single scalar objective can hide important trade-offs, I evaluate routes with three complementary indices: total path length \(L\), average safety distance \(D_{\text{avg}}\) and minimum safety distance \(D_{\text{min}}\). The first measures efficiency, the second measures global clearance, and the third captures the most exposed segment of the trajectory, which is usually the decisive factor for the safety of UAV drones. I combine them through a weighted score in which length is normalized in reverse,

$$\text{score}=\omega_1\left(1-\frac{L}{L_{\max}}\right)+\omega_2\left(\frac{D_{\text{avg}}}{D_{\text{avg},\max}}\right)+\omega_3\left(\frac{D_{\text{min}}}{D_{\text{min},\max}}\right),$$

subject to \(\omega_1+\omega_2+\omega_3=1\). The weights are not chosen arbitrarily. I derive them analytically from pairwise comparisons using the standard analytical hierarchy procedure, in which the comparison matrix

$$A=\begin{bmatrix}1 & 4 & 7\\ 1/4 & 1 & 3\\ 1/7 & 1/3 & 1\end{bmatrix}$$

encodes the judgement that length matters more than average clearance, and that average clearance matters moderately more than minimum clearance. Solving the eigenproblem \(Av=\lambda_{\max}v\) and normalizing yields \(\lambda_{\max}=3.032\), the priority vector

$$v=\begin{bmatrix}0.9518 & 0.2847 & 0.1136\end{bmatrix}^{T},$$

and hence the normalized weights

$$\omega=\begin{bmatrix}0.705 & 0.211 & 0.084\end{bmatrix}^{T}.$$

Consistency is verified through

$$CI=\frac{\lambda_{\max}-n}{n-1}=\frac{3.032-3}{2}=0.016,\qquad CR=\frac{CI}{RI(3)}=\frac{0.016}{0.58}=0.028,$$

and since \(CR<0.10\) the comparison matrix is judged consistent. The resulting weights are rounded to \(0.7\), \(0.2\) and \(0.1\) for the experiments. Table 10 summarizes the hierarchy.

Criterion Role in Mission Derived Weight
Total path length Direct driver of flight time and energy consumption 0.705
Average safety distance Overall risk reduction along the whole trajectory 0.211
Minimum safety distance Feasibility guarantee at the tightest passage 0.084

7.4 Executability Verification Framework

Geometric feasibility does not guarantee that UAV drones can actually follow a planned route. Sharp corners may be geometrically valid yet require accelerations that violate actuator limits. I therefore built a closed-loop verification framework in which the planned waypoint sequence serves as a reference input to a simulated flight controller, and the resulting trajectory is compared against the reference. The vehicle is represented by a simplified translational model in which the acceleration follows from the commanded input plus disturbance, scaled by an uncertain mass factor:

$$a_k=\frac{u_k+d_k}{m_k},\qquad v_{k+1}=v_k+a_k\Delta t,\qquad p_{k+1}=p_k+v_{k+1}\Delta t.$$

Three controllers are compared. The proportional-integral-derivative controller generates commands from the position error,

$$u_k=K_p e_k+K_i\sum_{j=0}^{k}e_j\Delta t+K_d\frac{e_k-e_{k-1}}{\Delta t},\qquad e_k=x_k^{\text{ref}}-x_k.$$

The linear quadratic regulator uses state feedback on position and velocity,

$$X_{k+1}=AX_k+Bu_k,\qquad u_k=-K\left(X_k-X_k^{\text{ref}}\right).$$

The enhanced regulator adds a preview-based velocity feed-forward term with an error-dependent gain,

$$u_k=-K\left(X_k-X_k^{\text{ref}}\right)+\alpha\left(v_k^{\text{pred}}-v_k\right),\qquad \alpha=\alpha_0+\frac{\alpha_1}{1+\lVert X_k-X_k^{\text{ref}}\rVert},$$

$$v_k^{\text{pred}}=\frac{x_{k+L}^{\text{ref}}-x_k^{\text{ref}}}{L\Delta t}.$$

Sensor noise, mass variation and external wind-like disturbances are injected throughout the simulation so that the verification reflects realistic conditions for UAV drones. The recorded metrics are listed in Table 11.

Metric Definition Preferred Direction
Root mean square error Overall tracking accuracy across the run Lower
Maximum error Peak deviation from the reference path Lower
Final error Residual distance at the end of the trajectory Lower
Mean jerk Average rate of change of acceleration Lower
Maximum jerk Largest acceleration discontinuity Lower
Robustness Reciprocal of error variance Higher
Disturbance-to-error ratio Ratio of injected disturbance to resulting error Higher

7.5 Static Obstacle Results

Three static environments were constructed with eight, ten and twelve cylindrical obstacles respectively. Six planners were compared: the classical potential field, a tree-sampling planner, two improved potential field variants and a wolf-based potential field hybrid, all against the proposed fused planner. Table 12 reports the outcomes for the first environment.

Planner Success Path Length Average Safety Distance Minimum Safety Distance
Classical potential field Yes 363 14.06 2.61
Tree sampling Yes 293 25.93 0.10
Improved potential field A Yes 320 19.33 3.99
Improved potential field B Yes 377 7.96 0.02
Wolf-based hybrid Yes 351 12.81 1.56
Proposed fused planner Yes 236 12.42 2.38

The proposed planner returns the shortest route, reducing length by approximately thirty-five percent relative to the classical field, nineteen percent relative to tree sampling, twenty-six percent relative to the first improved field, thirty-seven percent relative to the second improved field and thirty-three percent relative to the wolf-based hybrid. In terms of safety, the tree-sampling planner maintains a large average clearance but its minimum clearance of 0.10 indicates segments that would be unacceptable for real UAV drones. The first improved potential field achieves the largest minimum clearance, but it does so by accepting a longer route, which illustrates how easily a single-index evaluation can be misleading. The proposed planner keeps its minimum clearance near 2.38 while preserving the shortest overall route, and it attains the highest weighted composite score.

The second environment produced similar conclusions, as shown in Table 13.

Planner Success Path Length Average Safety Distance Minimum Safety Distance
Classical potential field Yes 323 9.32 3.31
Tree sampling Yes 281 22.72 0.12
Improved potential field A Yes 289 19.03 4.05
Improved potential field B Yes 298 7.16 3.65
Wolf-based hybrid Yes 225 13.50 0.36
Proposed fused planner Yes 210 10.35 2.03

The fused planner again returns the shortest route at 210 units, with a minimum clearance that remains comfortably above the thresholds that would endanger UAV drones. The tree-sampling route fluctuates considerably and includes backtracking segments. The classical field is direct but rigid in congested regions. The first improved field expands and reorients frequently in locally complex areas. The wolf-based hybrid is competitive on length but its minimum clearance of 0.36 is marginal.

The third environment contains twelve obstacles arranged in a dense, narrow-corridor configuration, which is the scenario most likely to defeat field-based planners. Here three of the six planners fail entirely, as reported in Table 14.

Planner Success Path Length Average Safety Distance Minimum Safety Distance
Classical potential field No — — —
Tree sampling Yes 327 17.48 0.15
Improved potential field A No — — —
Improved potential field B No — — —
Wolf-based hybrid Yes 411 8.04 0.48
Proposed fused planner Yes 276 9.40 1.90

The classical field terminates after becoming trapped in a local minimum, the first improved field cannot escape the combined repulsion of the clustered cylinders, and the second improved field oscillates without producing a stable route. The two planners that do succeed generate visibly circuitous trajectories, and the tree-sampling planner approaches obstacles at a minimum distance of only 0.15 units. The fused planner reaches the goal with a continuous route and maintains a minimum clearance roughly twelve times larger, which demonstrates that the swarm refinement stage genuinely removes the symmetry that traps purely field-based planners. The improvement is directly attributable to the swarm’s ability to search laterally near force-balance points, so that the resulting motion of the UAV drones is no longer confined to the field gradient direction.

7.6 Dynamic Obstacle Results

The dynamic environment combines an undulating terrain base, forty-five static cylinders grouped into clusters, and three irregular moving obstacles that traverse horizontal ellipses with sinusoidal vertical oscillation. Table 15 reports the comparative outcomes.

Planner Success Path Length Average Safety Distance Minimum Safety Distance
Classical potential field Yes 202 6.70 0.10
Tree sampling Yes 269 9.42 0.99
Improved potential field A Yes 221 17.65 0.20
Improved potential field B Yes 204 18.28 0.01
Wolf-based hybrid Yes 282 17.87 0.05
Proposed fused planner Yes 224 16.54 4.45

The proposed planner achieves a minimum clearance of 4.45, which is an order of magnitude larger than any competing method. The classical field produces a compact route but passes dangerously close to obstacles at a minimum distance of 0.10, and the second improved field is even more aggressive at 0.01. Tree sampling and the wolf-based hybrid travel noticeably farther than necessary. The first improved field achieves good average clearance but still has a minimum clearance of 0.20, meaning that a small number of critical passages remain risky. Because the minimum clearance is the quantity that governs whether UAV drones can traverse a corridor without collision, this result is the most important safety indicator in the dynamic study. The proposed planner also attains the highest weighted composite score, balancing a moderate route length against substantially superior clearance.

7.7 Executability Verification Results

Two representative reference trajectories were tracked, one produced by the fused planner in the dense obstacle environment and one produced by the terrain-optimized hybrid swarm in the mountainous environment. In both cases the three controllers completed the trajectory, but their accuracy differed substantially. Table 16 reports the metrics for the obstacle-environment reference.

Metric PID Linear Quadratic Regulator Enhanced Regulator
Root mean square error 24.658 3.9034 3.5319
Maximum error 37.987 6.3940 5.8692
Final error 27.807 5.1163 4.8053
Mean jerk 19.520 37.795 35.651
Maximum jerk 32.876 108.35 114.67
Robustness 0.0079532 0.28789 0.33678
Disturbance-to-error ratio 0.00042734 0.015143 0.020346

Table 17 reports the corresponding metrics for the terrain-environment reference, where the spatial scale is larger and the altitude variation more pronounced.

Metric PID Linear Quadratic Regulator Enhanced Regulator
Root mean square error 184.60 28.722 27.078
Maximum error 258.92 53.619 51.023
Final error 189.06 33.264 31.447
Mean jerk 151.57 183.98 177.50
Maximum jerk 240.68 853.07 822.51
Robustness 0.00027275 0.0058088 0.0065847
Disturbance-to-error ratio 0.0000090055 0.00035242 0.00033593

Several conclusions follow. The proportional-integral-derivative controller performs worst in both environments. Its tracking error grows steadily during segments where the reference trajectory changes direction rapidly, and it is unable to recover quickly at the final approach. The state-feedback regulator reduces the root mean square error by roughly a factor of six in the obstacle environment and by more than a factor of six in the terrain environment, because the velocity feedback term allows it to anticipate rather than merely react. The enhanced regulator improves further by combining preview-based feed-forward with an adaptive gain, producing the smallest errors and the highest robustness values in both scenarios.

The trade-off between accuracy and jerk is visible and expected. The regulators achieve their accuracy through more aggressive control action, which increases jerk relative to the proportional-integral-derivative baseline. Nevertheless, the jerk distributions remain concentrated in the lower range, indicating that the controllers do not produce sustained violent motion. The obstacle-environment reference generates systematically smaller tracking errors than the terrain-environment reference, which follows directly from the differences in spatial scale and altitude variation: the terrain route traverses a much larger volume with steeper climbs, so the same controller authority translates into larger absolute deviations. Both planning methods therefore produce routes that are executable within the simplified translational model and under the injected uncertainty, which supports the conclusion that the planned trajectories for UAV drones are not merely geometrically valid but also dynamically trackable.

8. Discussion

The two contributions address complementary weaknesses. The adaptive hybrid swarm targets scenarios in which the dominant difficulty is the sheer size and ruggedness of the search space. Its improvements accumulate across the search horizon: the adaptive inertia controls tempo, the employed-bee stage sharpens local resolution, mutation maintains variance, and chaotic mapping provides a late-stage escape mechanism. Together they produce routes that are shorter, smoother and more consistent across repeated runs of UAV drones missions.

The fused planner targets scenarios in which the dominant difficulty is structural rather than dimensional. In dense obstacle clusters the potential field itself becomes the obstacle, because force cancellation can halt progress regardless of how much global search power is available. Replacing the discontinuous repulsion with a Gaussian envelope removes abrupt force transitions, and embedding a swarm refinement stage near each force-balance point breaks the symmetry that causes entrapment. The altitude-dependent activation condition further removes repulsive contributions from obstacles that the UAV drones have already overflown.

Table 18 consolidates the qualitative comparison between the two approaches.

Aspect Adaptive Hybrid Swarm Fused Swarm and Potential Field
Primary environment Rough terrain with threat volumes Dense static and dynamic obstacles
Dominant challenge Large search space and premature convergence Local minima and goal unreachability
Core mechanism Multi-strategy swarm with chaotic escape Field guidance blended with swarm refinement
Path quality emphasis Length, altitude band and smoothness Minimum clearance and reachability
Verification emphasis Benchmark accuracy and route statistics Weighted composite score and closed-loop tracking
Typical strength Robust convergence with low variance Reliable clearance in congested corridors

Together the two methods cover a broad range of operating conditions for UAV drones. The first is preferable when route optimality and convergence reliability dominate, while the second is preferable when obstacle congestion and tight clearances dominate. Neither is universally superior, and the selection between them should follow from the mission profile, the sensor suite and the available onboard computation.

9. Conclusions and Future Work

I have studied three-dimensional path planning for UAV drones in complex environments and proposed two improved planners supported by a comprehensive evaluation methodology. The first planner is an adaptive hybrid particle swarm optimizer that combines a non-linear inertia schedule, an employed-bee neighbourhood search, a mutation operator and a late-stage logistic chaotic map. Numerical experiments on a standard benchmark suite show that the proposed method achieves the best average rank among five metaheuristics and that each mechanism contributes measurably to the final performance. In three-dimensional terrain simulations the method attains the best worst-case, best-case and average cost on both test maps, confirming its value for terrain-constrained routing of UAV drones.

The second planner fuses particle swarm optimization with a Gaussian-modified artificial potential field. The Gaussian envelope smooths the repulsive transition at obstacle boundaries, and the swarm refinement stage breaks the force-balance symmetry that traps conventional field-based planners. Across three static environments of increasing obstacle density and one dynamic environment with moving obstacles, the fused planner consistently produced the shortest routes while maintaining minimum clearances far above those of competing methods, and it succeeded in the dense configuration where three alternative planners failed outright.

I further constructed an evaluation architecture that goes beyond geometric feasibility. The weighted composite score, derived from pairwise comparisons and validated through consistency analysis, balances route length against average and minimum clearance. Above this, the executability verification framework closes the loop by feeding planned reference trajectories into three flight controllers under injected noise, mass uncertainty and external disturbance. State-feedback control with preview-based feed-forward achieved the lowest tracking errors and the highest robustness in both the obstacle and terrain scenarios, which supports the conclusion that the planned routes for UAV drones are controllable as well as geometrically valid.

Several directions remain open. First, the validation is entirely simulation-based; deploying these planners on physical UAV drones with onboard sensing would provide stronger evidence of practical viability. Second, the environments studied are static or weakly dynamic, and the treatment of environmental uncertainty is limited. Incorporating online replanning triggered by newly sensed hazards would improve the responsiveness of UAV drones in genuinely dynamic settings. Third, this work considers only a single vehicle. Extending the framework to cooperative teams would require attention to communication constraints, conflict resolution and coordinated task allocation. Fourth, the vehicle model used for executability verification is deliberately simple and omits attitude dynamics, angular rate limits and aerodynamic effects; a richer model, potentially combined with adaptive parameter tuning, would improve the fidelity of the executability assessment. Addressing these gaps would move the proposed methods closer to routine operational use for UAV drones operating in complex three-dimensional environments.

</z_i

Scroll to Top