The rapid integration of Unmanned Aerial Vehicles (UAVs), or drones, into daily life is fundamentally reshaping the world. From military operations to agriculture, logistics, mapping, and rescue missions, the technology and applications of UAVs are experiencing exponential growth. In this context, efficient and reliable path planning is paramount for the successful execution of UAV missions. The inherent complexity of three-dimensional (3D) airspace, filled with static and dynamic obstacles, presents a significant challenge. This article addresses this challenge by presenting a novel, improved version of the Informed Rapidly-exploring Random Tree* (Informed-RRT*) algorithm, specifically tailored for 3D path planning of UAVs in China and globally.
Path planning algorithms can be broadly categorized into graph-based search methods (e.g., A*, D*), intelligent algorithms (e.g., Ant Colony, Particle Swarm Optimization), and sampling-based algorithms (e.g., Probabilistic Roadmap Method – PRM, Rapidly-exploring Random Tree – RRT). Sampling-based algorithms are particularly suited for 3D environments as they do not require discretization of the space beforehand. Among them, the RRT algorithm is renowned for its simplicity and probabilistic completeness. However, it often yields suboptimal paths and suffers from low convergence efficiency. Its asymptotically optimal variant, RRT*, improves path quality through a rewiring process but remains computationally expensive. The Informed-RRT* algorithm represents a significant leap by restricting subsequent sampling to a heuristic ellipsoid after an initial solution is found, focusing the computational effort. Despite this advancement, it still faces critical shortcomings: the initial feasible path can be excessively long, the sampling strategy within the ellipsoid is not adaptive to obstacle density, and the computational cost for asymptotic optimality remains high, leading to long runtimes.
Our work proposes a comprehensive modification to the Informed-RRT* algorithm to overcome these limitations. The improvements are threefold. First, we introduce a Dynamic Artificial Potential Field (DAPF) guided by fuzzy logic to steer the growth of the random tree towards the goal while dynamically adjusting its influence based on local obstacle congestion, thereby accelerating the discovery of a shorter initial path. Second, we stratify the informed sampling ellipsoid into an inner and outer layer and dynamically adjust the sampling probability between them according to the sparsity or density of obstacles in the environment, significantly enhancing convergence efficiency. Third, we optimize the critical neighborhood search radius parameter using a hybrid approach combining a Feedforward Neural Network (FNN) and a Genetic Algorithm (GA), striking an optimal balance between path quality and computational time. We rigorously evaluate our algorithm in complex 3D environments with varying obstacle densities, demonstrating its superiority over standard Informed-RRT*, RRT*, and A* algorithms in terms of both final path length and planning efficiency. This makes it a highly practical solution for real-world China UAV drone navigation tasks.
Theoretical Foundation: From RRT* to Informed-RRT*
The standard RRT algorithm grows a tree T = (V, E) from the start node xinit by randomly sampling the configuration space X. For each sample xrand, it finds the nearest neighbor xnearest in V, extends a new node xnew from xnearest towards xrand by a step size η, and adds it to the tree if the path is collision-free. RRT* enhances this process with two key steps for asymptotic optimality:
- Parent Re-selection: For a new node xnew, instead of simply connecting it to xnearest, RRT* examines all nodes within a ball of radius r centered at xnew. It selects the node in this neighborhood that provides the minimum-cost path from xinit to xnew as its new parent.
- Rewiring: After connecting xnew, the algorithm checks if rewiring existing nodes in the neighborhood to xnew as their parent would reduce their path cost from the root. If so, the connection is updated.
The radius r of this neighborhood is crucial and is typically set as:
$$r = \min\left(\gamma_{RRT*} \left( \frac{\ln(|V|)}{|V|} \right)^{\frac{1}{d}}, \eta\right)$$
where γRRT* is a tuning parameter, |V| is the number of vertices in the tree, d is the dimension of the space, and η is the step size.
The Informed-RRT* algorithm builds upon RRT* by introducing a focused sampling strategy. Once an initial path with cost cbest is found, it restricts all future random sampling to an ellipsoidal subset of the configuration space. This ellipsoid is defined such that its foci are the start (xstart) and goal (xgoal) points, and its transverse axis length equals cbest. Any path shorter than cbest must lie within this ellipsoid. The algorithm uniformly samples this ellipsoid, dramatically improving the convergence rate towards the optimal path. The sampling process within the ellipsoid is achieved through a transformation from a unit ball. Let cmin = ||xgoal – xstart|| be the Euclidean distance between start and goal. The ellipsoid’s major axis length is cbest/2, and its minor axis lengths are $\sqrt{c_{best}^2 – c_{min}^2}/2$. A sample xball is drawn from a unit ball, and then transformed:
$$x_{rand} = (C L x_{ball} + x_{centre}) \cap X$$
Here, xcentre is the midpoint between start and goal, C is a rotation matrix aligning the major axis with the start-goal vector, and L is a diagonal matrix containing the ellipsoid’s axis lengths.
Proposed Algorithm: A Triad of Enhancements
Our modified algorithm targets the core inefficiencies of Informed-RRT*. The block diagram below summarizes the integrated architecture of our approach.
1. Dynamic Artificial Potential Field Guidance
A common heuristic to accelerate RRT growth is to bias sampling or extension towards the goal. A simple method is to steer the new node xnew using a vector sum:
$$x_{new} = x_{nearest} + \eta \left( \alpha \frac{x_{goal} – x_{nearest}}{||x_{goal} – x_{nearest}||} + (1-\alpha) \frac{x_{rand} – x_{nearest}}{||x_{rand} – x_{nearest}||} \right)$$
where α is a static bias parameter. However, a fixed α causes problems in cluttered environments: the strong pull towards the goal can lead to repeated collision checks and stagnation near obstacles as the extension constantly hits barriers.
We propose a Dynamic Artificial Potential Field (DAPF) where the attraction gain α (or equivalently, a dynamic step size component ηattr) is adjusted online via fuzzy logic. The input to the fuzzy system is nrand, the count of consecutive failed node expansion attempts (collisions) in the current iteration. A high nrand indicates a congested region, suggesting the need to reduce goal attraction to allow the tree to explore around the obstacle. A low nrand suggests clear space, where goal attraction can be increased for faster progression.
We define fuzzy sets for the input nrand: Positive Small (PS), Positive Medium (PM), Positive Big (PB). Similarly, for the output attraction coefficient ηattr (scaled between 0 and a maximum step). The fuzzy rules are:
- IF nrand is PS, THEN ηattr is PB. (Low failures, high attraction)
- IF nrand is PB, THEN ηattr is PS. (High failures, low attraction)
This dynamic adjustment allows the China UAV drone’s path planner to smoothly transition between exploratory and goal-directed behavior, efficiently navigating both open and cluttered spaces.
2. Stratified Ellipsoidal Sampling
The standard Informed-RRT* samples uniformly within the entire ellipsoid. We argue this is suboptimal. The nature of the optimal path depends on obstacle density between start and goal. If obstacles are sparse, the shortest path likely stays close to the straight line (the major axis). If obstacles are dense, the optimal path may need to circumnavigate them, lying closer to the periphery of the ellipsoid.
We therefore partition the informed ellipsoid into two regions: an inner ellipsoid (X2ellipse) and the surrounding outer shell (Xc). The inner ellipsoid is a scaled-down version of the original, with its major axis endpoints remaining at xstart and xgoal. The scaling factor is $c_{min}/c_{best}$. Therefore, the semi-axis lengths relate as:
$$\frac{a_2}{a_1} = \frac{b_2}{b_1} = \frac{c_2}{c_1} = \frac{c_{min}}{c_{best}}$$
where subscript 1 denotes the outer (original) ellipsoid and 2 denotes the inner ellipsoid.
The sampling probability is adaptive. We define an obstacle density metric:
$$a = \frac{c_{best}}{c_{min}}, \quad a \ge 1$$
A value of a close to 1 indicates a near-optimal, direct path (sparse obstacles). A large a indicates significant detour is needed (dense obstacles). The probability Pinner of sampling inside the inner ellipsoid is set inversely proportional to a. For example:
$$P_{inner} = \frac{1}{a}, \quad P_{outer} = 1 – P_{inner}$$
This strategy concentrates samples where they are most likely to improve the current best path, whether that’s refining a direct route or exploring the periphery for a better detour, making the China UAV drone’s search far more efficient.
3. Optimized Neighborhood Radius via FNN-GA
The neighborhood radius r in the parent re-selection and rewiring steps is critical. The theoretical form $r \propto (\ln|V|/|V|)^{1/d}$ ensures asymptotic optimality but may not be practical for finite-time performance in 3D China UAV drone path planning. A poorly chosen γRRT* can lead to excessive neighborhood checks (large r) or insufficient optimization (small r).
We propose a data-driven optimization of the function r(|V|). First, we use a Genetic Algorithm (GA) to find an optimal constant γ* for the classical formula, using a fitness function that balances path length l and computation time t:
$$f = w_1 \cdot l + w_2 \cdot t$$
with weights w1=0.3 and w2=0.7. The GA found an optimal γ* = 19 for our 3D setup.
Second, we model the relationship r(|V|) using a Feedforward Neural Network (FNN) to capture potential non-linearities better than the closed-form expression. We employ a network with one input (|V|), two hidden layers (2 neurons each), and one output (r). The forward propagation for a single input x1 is:
$$
\begin{aligned}
\text{Hidden Layer 1:} & \quad \mathbf{z_1} = \mathbf{W^{(1)}} x_1 + \mathbf{b^{(1)}}, \quad \mathbf{a_1} = \sigma(\mathbf{z_1}) \\
\text{Hidden Layer 2:} & \quad \mathbf{z_2} = \mathbf{W^{(2)}} \mathbf{a_1} + \mathbf{b^{(2)}}, \quad \mathbf{a_2} = \sigma(\mathbf{z_2}) \\
\text{Output Layer:} & \quad r = \mathbf{W^{(3)}} \mathbf{a_2} + b^{(3)}
\end{aligned}
$$
where σ is the sigmoid activation function, and W, b are weights and biases.
Third, we use a GA to optimize the 13 parameters (weights and biases) of this pre-trained FNN. The chromosome encodes these parameters, and the fitness function remains the same. This hybrid FNN-GA approach learns a highly optimized, problem-specific mapping from tree size to neighborhood radius, minimizing total planning time.
4. UAV Kinematic Constraints
For practical China UAV drone deployment, the generated path must respect kinematic constraints. We incorporate two primary 3D constraints:
- Maximum Heading Angle (ψmax): The horizontal turn angle between consecutive path segments must be limited.
$$\arccos\left( \frac{\mathbf{p_i} \cdot \mathbf{p_{i+1}}}{||\mathbf{p_i}|| \; ||\mathbf{p_{i+1}}||} \right) \le \psi_{max}$$
where pi is the 2D projection (x,y) of the path segment vector from node i-1 to i. - Maximum Pitch Angle (θmax): The vertical climb/dive angle must be limited.
$$\frac{|z_{i+1} – z_i|}{\sqrt{(x_{i+1}-x_i)^2 + (y_{i+1}-y_i)^2}} \le \tan(\theta_{max})$$
These constraints are enforced during the node extension and rewiring phases. A proposed connection is only valid if it does not violate ψmax or θmax. For our simulations, we set ψmax = 45° and θmax = 40°.
Simulation Results and Analysis
We evaluate our improved algorithm in a large 3D environment (1470m × 1470m × 900m) with start at (80, 80, 0)m and goal at (760, 760, 720)m. We compare against standard RRT*, Informed-RRT*, and the grid-based A* algorithm with two resolutions (low: 10×10×15, high: 30×30×15). All sampling-based algorithms run for 6000 iterations with a step size η = 100m. Results are averaged over 50 independent runs for statistical significance.
1. Sparse Obstacle Environment
In an environment with few obstacles between start and goal, the optimal path is expected to be relatively direct. The following table summarizes the performance.
| Algorithm | Initial Path Length (m) | Iterations to First Path | Final Path Length (m) | Computation Time (s) |
|---|---|---|---|---|
| RRT* | 1930 | 622 | 1640 | 5.40 |
| Informed-RRT* | 1910 | 617 | 1510 | 6.21 |
| A* (Low Res) | N/A | N/A | 1707 | 4.94 |
| A* (High Res) | N/A | N/A | 1238 | 18.68 |
| Our Algorithm | 1620 | 340 | 1220 | 5.59 |
Analysis: Our algorithm finds a much shorter initial path 16-19% faster (fewer iterations) due to the DAPF guidance. The stratified sampling, with higher probability inside the inner ellipsoid in this sparse case, leads to the best final path length, even slightly outperforming the high-resolution A* which is computationally prohibitive. Our optimized radius keeps runtime competitive with RRT* and lower than Informed-RRT*.
2. Dense Obstacle Environment
In a cluttered environment where a direct route is blocked, the path must navigate around obstacles. The performance comparison is shown below.
| Algorithm | Initial Path Length (m) | Iterations to First Path | Final Path Length (m) | Computation Time (s) |
|---|---|---|---|---|
| RRT* | 2310 | 866 | 1880 | 7.26 |
| Informed-RRT* | 2340 | 842 | 1720 | 7.84 |
| A* (Low Res) | N/A | N/A | 1832 | 6.04 |
| A* (High Res) | N/A | N/A | 1322 | 19.69 |
| Our Algorithm | 1900 | 424 | 1310 | 7.39 |
Analysis: The superiority of our method is even more pronounced. The DAPF effectively navigates the clutter, finding an initial path ~18% shorter and ~50% faster. The adaptive stratified sampling now favors the outer shell, guiding the optimization towards the efficient detour, resulting in the shortest final path. The runtime remains efficient, showcasing the robustness of our FNN-GA radius optimization.
3. Effect of Stratified Sampling Probability
To validate the adaptive sampling logic, we analyzed the final path length against a fixed inner-layer sampling probability (Pinner). The results confirm our theory:
- In Sparse environments (low a), higher Pinner yields shorter paths, as optimization focuses on the central region.
- In Dense environments (high a), lower Pinner (i.e., higher Pouter) yields shorter paths, as the optimal circumventing path lies in the outer shell of the ellipsoid.
This experiment underscores the importance of our adaptive probability adjustment based on the real-time obstacle density metric a for China UAV drone navigation.
4. Performance Across Varying Obstacle Counts
We further tested scalability by varying the number of obstacles from 10 to 30. The trends solidify our conclusions.
| Obstacle Count | Metric | RRT* | Informed-RRT* | Our Algorithm |
|---|---|---|---|---|
| 10 | Initial Length (m) | 1750 | 1745 | 1520 |
| Final Length (m) | 1480 | 1350 | 1210 | |
| 20 | Initial Length (m) | 2050 | 2070 | 1780 |
| Final Length (m) | 1760 | 1620 | 1380 | |
| 30 | Initial Length (m) | 2380 | 2400 | 1950 |
| Final Length (m) | 1950 | 1780 | 1480 |
Our algorithm consistently produces the shortest initial and final paths across all tested complexities, demonstrating the general applicability of the proposed enhancements for China UAV drone operations in diverse 3D scenarios.
Conclusion
This article has presented a significantly improved Informed-RRT* algorithm for 3D UAV path planning. By integrating a fuzzy-logic-based Dynamic Artificial Potential Field, an adaptive Stratified Ellipsoidal Sampling strategy, and a hybrid FNN-GA optimized neighborhood radius, we have systematically addressed the key limitations of the original algorithm: slow initial solution finding, non-adaptive heuristic sampling, and high computational cost. Extensive simulations in both sparse and dense 3D environments confirm that our algorithm converges to a high-quality path significantly faster than RRT* and Informed-RRT*, and it achieves better paths than the computationally expensive high-resolution A* algorithm in a fraction of the time. The incorporation of UAV kinematic constraints ensures the practical feasibility of the generated paths. Therefore, this work provides a robust, efficient, and practical solution for autonomous 3D path planning, holding great promise for advancing the capabilities of China UAV drone systems in complex real-world applications such as urban logistics, infrastructure inspection, and aerial surveying.