In the realm of autonomous aerial robotics, the ability to navigate through complex three-dimensional (3D) terrains while avoiding static obstacles is a fundamental requirement for the deployment of China UAV drones across a wide range of civilian and military applications. These tasks, which include mountainous search-and-rescue, precision agriculture in hilly regions, and infrastructure inspection in remote areas, demand a path planning algorithm that can reconcile the conflicting objectives of minimizing flight distance, ensuring safety, and maintaining flight smoothness. Classical evolutionary optimization techniques, such as the Ant Colony Optimization (ACO), have demonstrated considerable potential due to their self-organizing capabilities and robust feedback mechanisms. However, when applied to continuous 3D path planning, traditional ACO often faces three critical challenges: poor convergence speed due to random initialization, a propensity to become trapped in local optima, and an imbalance between exploration and exploitation. To address these issues, we propose a novel hybrid algorithm, ACOAGR-GA, which synergistically integrates Q-Learning for guided initialization, a Gaussian-Lévy mixture walk for dynamic step-size adjustment, an adaptive probabilistic node selection mechanism, and a genetic algorithm (GA) elitist crossover strategy. Our comprehensive experimental results, validated across four demanding task scenarios, demonstrate that this approach significantly outperforms the state-of-the-art methods in achieving lower total path costs, thereby enhancing the autonomous flight safety and efficiency of China UAV drones.
Formulation of the 3D Path Planning Problem
The path planning problem for China UAV drones is formulated as a constrained multi-objective optimization task. A feasible path is represented as a sequence of \(N\) waypoints \(P_1, P_2, \ldots, P_N\), where each waypoint is defined by its 3D coordinates \((x_j, y_j, z_j)\). The overall cost of a path is computed as a weighted sum of several sub-costs.
| Cost Component | Symbol | Mathematical Definition |
|---|---|---|
| Flight Distance | \(J_1\) | $$J_1 = \sum_{j=1}^{N-1} \sqrt{(x_{j+1}-x_j)^2 + (y_{j+1}-y_j)^2 + (z_{j+1}-z_j)^2}$$ |
| Collision & Threat | \(J_2\) | $$J_2 = \sum_{j=1}^{N} \sum_{i=1}^{M} C_{ij}$$ where \(C_{ij} = \begin{cases} \infty, & d_{ij} \le R_i \\ \frac{k}{d_{ij} – R_i}, & R_i < d_{ij} \le R_{threat,i} \\ 0, & \text{otherwise} \end{cases}\) |
| Flight Altitude | \(J_3\) | $$J_3 = \begin{cases} \infty, & z_i < z_{min} \\ \frac{|z_i – (z_{max}+z_{min})/2|}{2}, & \text{otherwise} \end{cases}$$ |
| Path Smoothness | \(J_4\) | $$J_4 = \sum_{i=1}^{N-1} \left[ I(|\Delta\theta(i)| > 45^\circ) \cdot |\Delta\theta(i)| + I(|\Delta\alpha(i)| > 45^\circ) \cdot |\Delta\alpha(i)| \right]$$ |
| Total Cost | \(Cost\) | $$Cost = b_1 J_1 + b_2 J_2 + b_3 J_3 + b_4 J_4$$ |
Where \(d_{ij}\) is the horizontal distance from the UAV to the i-th obstacle, \(R_i\) and \(R_{threat,i}\) are the collision and threat radii, respectively, and \(k\) is the penalty coefficient. The smoothness cost penalizes yaw (\(\theta\)) and pitch (\(\alpha\)) angles that exceed the maximum allowable limits. The trajectory is smoothed using a k-order B-spline curve: \(p(t) = \sum_{i=0}^{n} N_{i,k}(t) P_i\). This comprehensive cost function ensures that the resulting path is not only short but also safe, altitudinally compliant, and dynamically feasible for China UAV drones.
Q-Learning Guided Initial Population Generation
To overcome the blind initial search of the standard ACO, we employ a Q-Learning agent to generate a set of high-quality initial paths. The agent learns the environmental topology and threat distribution through iterative exploration. The state space \(s\) includes the UAV’s position, its distance to the goal, and distances to nearby threats. The action space comprises discrete changes in yaw and pitch angles, allowing the UAV to move towards or away from obstacles. The reward function \(R(s, a)\) is designed to encourage goal-seeking behavior while punishing proximity to obstacles:
$$
R(s, a) = r_0 + r_{goal} – r_{threat}
$$
$$
r_{goal} = r_{max} \cdot \exp\left(-\frac{(x_u-x_g)^2 + (y_u-y_g)^2 + (z_u-z_g)^2}{k_g}\right)
$$
$$
r_{threat} = r_t \cdot \sum_{i=1}^{m} \exp\left(-\frac{(x_u-x_{o,i})^2 + (y_u-y_{o,i})^2 + (z_u-z_{o,i})^2}{k_t}\right)
$$
The action selection follows a Boltzmann distribution to balance exploration and exploitation:
$$
P(a|s) = \frac{e^{Q(s,a)/T}}{\sum_{a_i \in A} e^{Q(s,a_i)/T}}
$$
where the temperature \(T = \lambda^k T_0\) decays over episodes. The Q-values are updated using the Bellman equation:
$$
Q(s, a) \leftarrow Q(s, a) + \eta [r + \gamma \max_{a’} Q(s’, a’) – Q(s, a)]
$$
After training, the Q-table generates \(m\) high-quality paths. These are combined with \(n\) random paths (with \(m/n = 4\)) to form the initial population. The initial pheromone \(\tau_{i,i+1}\) on the path segments is set proportional to the path’s fitness \(f\), where a lower cost results in a higher pheromone deposit:
$$
\tau(N_i, N_{i+1}) = \tau_{base} \cdot \left(1 + \frac{f_{max} – f}{f_{max} – f_{min}}\right)
$$
This guided initialization provides a superior starting point for the ACO process, significantly accelerating the initial convergence of the algorithm for China UAV drones.
Gaussian-Lévy Hybrid Walk Strategy
To enhance the algorithm’s ability to escape from local optima, we introduce a Gaussian-Lévy hybrid walk for the node update mechanism. This strategy dynamically switches between two distinct exploration modes. The Gaussian walk, based on a normal distribution \(\mathcal{N}(0, \sigma^2)\), generates small, dense steps suitable for local exploitation. In contrast, the Lévy flight generates steps with a power-law distribution, characterized by occasional long jumps that allow the algorithm to explore distant regions of the search space. The Lévy step size is defined as:
$$
\text{Step}_{Levy} = \frac{u}{|v|^{1/\beta}}
$$
$$
\sigma = \left( \frac{\Gamma(1+\beta) \sin(\pi\beta/2)}{\Gamma((1+\beta)/2) \cdot 2^{(\beta-1)/2}} \right)^{1/\beta}
$$
where \(u \sim \mathcal{N}(0, \sigma^2)\) and \(v \sim \mathcal{N}(0, 1)\). The parameter \(\beta \in (0.1, 2]\) controls the tail thickness of the distribution. By integrating this hybrid walk into the ACO framework, the individual ants can perform both refined local searches for path smoothing and large-scale jumps to discover entirely new path topologies, thereby preventing premature convergence and improving the global search capability for China UAV drones.
Adaptive Node Selection and Genetic Elitist Optimization
The standard ACO node selection probability is enhanced through two key modifications. First, a time-varying inertia weight \(w(t)\) is introduced to dynamically adjust the importance of pheromone (\(\alpha\)) and heuristic (\(\beta\)) information over the course of the iteration:
$$
w(t) = w_{max} – (w_{max} – w_{min}) \cdot \frac{t}{T}
$$
$$
\alpha(t) = \alpha_0 \cdot (1 – w(t)), \quad \beta(t) = \beta_0 \cdot w(t)
$$
Second, a direction-guidance heuristic \(D_{ij}\) is introduced, which calculates the cosine similarity between the vector from the current node \(i\) to candidate node \(j\) and the vector from node \(i\) to the target \(T\). This provides a global sense of direction, preventing the UAV from wandering aimlessly. The normalized heuristic is \(D’_{ij} = (D_{ij} + 1)/2\). The final adaptive selection probability is:
$$
P_{ij}^k(t) = \frac{[\tau_{ij}(t)]^{\alpha(t)} \cdot [D’_{ij}]^{\beta(t)}}{\sum_{s \in allowed} [\tau_{is}(t)]^{\alpha(t)} \cdot [D’_{is}]^{\beta(t)}}
$$
To fully utilize the path information from all individuals in the colony, we incorporate a genetic algorithm crossover step before the pheromone update. For each ant, its current path \(P_i(t)\) is first crossed over with its own historical best path \(P_{i,best}\), and then the resulting path is crossed over with the global best path \(P_{g,best}\) to generate candidate offspring paths. The fittest path, selected via an elitist strategy, is used for pheromone deposition:
$$
P_{i,final}(t) = \arg \min_{P \in \{P_i(t), P’_i(t)\}} f(P)
$$
The pheromone update is then performed using the selected elitist path:
$$
\Delta \tau_{ij}^k = \frac{Q}{f(P_{k,final}(t))}
$$
This genetic refinement ensures that high-quality path segments are propagated across the population, accelerating convergence while maintaining diversity.
Ablation Study of the ACOAGR-GA Algorithm
We conducted a thorough ablation study to quantify the contribution of each individual component of the proposed ACOAGR-GA algorithm. We compared the full algorithm against three baseline variants: (1) basic continuous ACO, (2) ACO with Q-Learning initialization (ACO-Q), and (3) ACO with Q-Learning and Gaussian-Lévy walk (ACO-QG). All experiments were run independently 40 times on the four task scenarios, and statistical significance was verified using a Wilcoxon rank-sum test.
| Task | ACO | ACO-Q | ACO-QG | ACOAGR-GA |
|---|---|---|---|---|
| A | 10288 ± 658* | 9823 ± 490* | 9545 ± 381* | 9399 ± 207* |
| B | 9520 ± 715* | 9150 ± 371* | 8550 ± 331* | 8696 ± 290* |
| C | 9840 ± 718* | 9420 ± 418* | 9110 ± 319* | 8919 ± 252* |
| D | 9450 ± 648* | 9080 ± 353* | 8790 ± 267* | 8584 ± 281* |
The results clearly demonstrate the progressive improvement. The ACO-Q variant substantially reduced the initial cost, overcoming the blindness of random initialization. The ACO-QG enhanced the global search ability, leading to a lower mean cost and significantly reduced standard deviation, indicating robustness against local optima. Finally, the full ACOAGR-GA with its adaptive selection and GA crossover achieved the lowest overall cost in all four tasks. For instance, in Task A, the improvement over the basic ACO was 8.6%, confirming the effectiveness of our integrated approach for optimizing the flight paths of China UAV drones.
Comparative Performance Across State-of-the-Art Algorithms
We benchmarked our ACOAGR-GA against four contemporary algorithms: ACOPAR, EEFO, HLOA, and CCO-AVOA. All algorithms were tested under identical conditions with a population size of 300 and 200 iterations. The mission parameters are defined in the following table, with 3D coordinates and obstacle radii.
| Task | Start | End | Obj. 1 | Obj. 2 | Obj. 3 | Obj. 4 | Obj. 5 | Obj. 6 | Obj. 7 |
|---|---|---|---|---|---|---|---|---|---|
| A | (10,5,5) | (90,80,5) | (20,20,6) | (40,19,7) | (75,30,8) | (30,40,8) | (50,70,7) | (65,50,8) | (80,70,10) |
| B | (5,10,1) | (90,70,15) | (30,25,8) | (70,50,8) | (55,30,7) | (40,50,7) | — | — | — |
| C | (10,10,5) | (85,60,15) | (25,35,8) | (70,55,7) | (60,35,8) | (40,50,7) | (55,75,7) | (40,20,8) | — |
| D | (20,10,15) | (80,80,15) | (30,20,10) | (50,15,5) | (65,55,10) | (8,18,8) | (50,70,8) | (70,80,7) | (92,68,5) |
| Task | ACOPAR | EEFO | HLOA | CCO-AVOA | ACOAGR-GA |
|---|---|---|---|---|---|
| A | 10289 ± 165* | 10215 ± 220* | 10477 ± 154* | 10203 ± 57* | 9399 ± 394* |
| B | 11326 ± 56* | 10429 ± 228* | 9672 ± 386* | 10087 ± 33* | 8696 ± 514* |
| C | 10221 ± 84* | 10645 ± 86* | 9492 ± 597* | 10955 ± 53* | 8919 ± 1117* |
| D | 10283 ± 105* | 10833 ± 58* | 8920 ± 310* | 10123 ± 49* | 8584 ± 1349 |
As shown in the results, the proposed ACOAGR-GA algorithm achieves the lowest total cost across all four tasks. The average total cost improvement over the second-best algorithm (CCO-AVOA in Task A, HLOA in Task D) was approximately 7%. The standard deviations indicate that our algorithm maintains a good balance between exploration and exploitation, consistently finding high-quality solutions. In contrast, algorithms like HLOA and EEFO often produced paths with unnecessary turns or high altitude fluctuations, while ACOPAR and CCO-AVOA sometimes generated paths that intersected obstacles. The superior performance of ACOAGR-GA makes it a highly reliable solution for the autonomous navigation of China UAV drones in complex, obstacle-rich environments.
Convergence Analysis and Path Quality
The convergence curves for the four tasks provide further insight into the algorithm’s behavior. While the initial cost of ACOAGR-GA is already low due to Q-Learning initialization, it demonstrates a second phase of rapid improvement in later iterations. This is particularly evident in Tasks C and D, where the hybrid walk strategy helps the algorithm to escape local optima and discover more efficient routes. This secondary convergence is a direct result of the Gaussian-Lévy walk, which allows for intermittent large-scale jumps to entirely new path regions, followed by refined local searches facilitated by the GA crossover mechanism. The final convergence points are consistently the lowest among all compared methods. Although the ACOAGR-GA may require slightly more iterations to fully settle compared to some baselines, the final accuracy improvement is substantial, achieving a convergence precision improvement of 10.4%, 17.4%, 12.0%, and 22.6% over the ACOPAR algorithm in Tasks A, B, C, and D, respectively. This trade-off between convergence speed and ultimate solution quality is favorable for the offline planning of critical missions for China UAV drones, where optimality is paramount.
Conclusion
This work successfully developed and validated a novel 3D path planning algorithm, ACOAGR-GA, specifically tailored for the complex operational environments of China UAV drones. By synergistically combining Q-Learning for guided initial population generation, a Gaussian-Lévy hybrid walk for enhanced global exploration, an adaptive probabilistic node selection mechanism, and a genetic elitist crossover for optimizing path segments, our algorithm effectively resolves the fundamental trade-off between convergence speed and global search capability inherent in classical ACO. Extensive simulation results across multiple high-fidelity terrain scenarios have demonstrated that ACOAGR-GA consistently outperforms existing state-of-the-art algorithms, achieving significantly lower total path costs with robust statistical significance. The generated paths are shorter, smoother, and maintain a safer distance from obstacles, making them highly suitable for real-world applications. Future work will focus on optimizing the parameter adaptation rules to reduce the computational overhead and number of iterations required for convergence, thereby enhancing the real-time applicability of the algorithm for dynamic environments where China UAV drones must perform rapid online replanning. The integration of this algorithm with robust LiDAR/IMU SLAM systems will be a key step towards full autonomy in GNSS-denied and highly cluttered environments.