Enhanced A*–Fuzzy DWA Hybrid Algorithm for AGV Path Planning in Confined Spaces

Abstract

Addressing the challenges of inefficient prolonged trajectory resolution and unreliable dynamic obstacle avoidance for intelligent vehicles in complex confined environments, this study proposes an innovative hybrid path planning method. Its core novelty is the deep integration of an enhanced A* algorithm for smooth global planning with a fuzzy logic-controlled Dynamic Window Approach (DWA). The enhanced A* generates efficient and smooth global paths, while fuzzy control significantly improves DWA’s robustness in dynamic, uncertain scenarios. This hybrid strategy achieves efficient synergy between global optimality and local reactive obstacle avoidance. Simulations demonstrate its superiority over conventional A* or DWA in path length, planning efficiency, and obstacle avoidance success rate. Experimental validation on a physical platform in simulated complex scenarios shows an average trajectory deviation ≤ 7.14%. The work provides an effective theoretical and methodological framework for navigation in constrained spaces, offering significant value for practical applications like logistics and automated parking.

1. Introduction

With the rapid development of Industry 4.0 and intelligent logistics systems, Automated Guided Vehicles (AGVs) have become core equipment for modern manufacturing, warehouse logistics, and port automation. Over the years, AGV technological capabilities have been continuously improving. Their operational scenarios are expanding from structured warehouses to dynamic open environments, presenting challenges which traditional path planning algorithms struggle to adapt to. Compared with path planning in static environments, Automatic Guided Vehicles in dynamic industrial scenarios must tackle complex constraints such as moving obstacles, unstructured terrains, and other dynamic uncertainties, which significantly exacerbate the complexity of path planning. Zhang, J. et al. [1] improved efficiency by dynamically adjusting parameters with a density correlation factor and integrating five-neighborhood with the Floyd algorithm, while enhancing safety through a smoothing coefficient and local point selection strategy. Gu C et al. [2] introduced a target-weighted heuristic function and optimized paths by integrating the Floyd algorithm with B-spline curves. Zhang W et al. [3] refined the neighborhood traversal rules of A* algorithm and adjusted the evaluation function of the Dynamic Window Approach. Chen X et al. [4] proposed the A-DDQN method by integrating the A* algorithm with Double Deep Q-Network (DDQN), and determined the optimal navigation strategy via a reward function. Nhouchi A et al. [5] proposed a method for integrating the A* algorithm into CAD platforms. This approach enables path planning and collision detection during the design phase, thereby optimizing the design process. Zhou Z et al. [6] improved the algorithm through chaotic initialization, segmented search, and hard constraint mutation optimization, while enhancing the Dynamic Window Approach (DWA) with sector potential fields, dynamic weighting, and an obstacle-free path attraction strategy. Xu X et al. [7] replaced traditional neighborhoods with obstacle-free rectangular boundaries and designed a novel operator by integrating bidirectional search and Euclidean distance estimation to reduce algorithm complexity. Additionally, the path safety and smoothness were enhanced through an adaptive cost function and the Slide-Rail corner adjustment strategy. Xu D et al. [8] designed a cost function to ensure that collision avoidance strategies comply with the International Regulations for Preventing Collisions at Sea (COLREG) and introduced an evaluation function to avoid local optimality. Zhang D et al. [9] reduced unnecessary turns by dynamically adjusting the weight of the cost function and optimized the layout of turning points by integrating an artificial potential field method to design a heuristic function with obstacle penalty terms. Han C et al. [10] proposed a five-neighborhood dynamic weighting A* algorithm and optimized the path by integrating second-order Bézier curves. Zammit C et al. [11] conducted a comparative analysis of graph-based A* algorithm and sampling-based RRT algorithm along with their variants. Sang H et al. [12] divided sub-target sequences by integrating the global planning results of the improved A* algorithm, thereby avoiding local minimum issues and ensuring compliance with the dynamic constraints of Unmanned Surface Vehicles (USV). Xiao J et al. [13] employed the Dynamic Window Approach (DWA) with an extended evaluation function for local path optimization. Xu L et al. [14] proposed the MSF-MTPO algorithm, which employs adaptive neighborhood A* (with bidirectional search) for path planning, combines corner correction and local tangent circle smoothing optimization, and uses global paths to guide the artificial potential field method for dynamic obstacle avoidance. Tang G et al. [15] proposed the geometric A-star algorithm, which achieves path smoothing through grid modeling, node filtering, and cubic B-spline curve fitting. B. Fu et al. [16] proposed an improved A-star algorithm, which shortens paths by determining whether obstacles exist between the current waypoint and target point. However, this method suffers from high computational costs and an inability to smooth turning paths. Chang L et al. [17] proposed an improved Dynamic Window Approach (DWA) based on Q-learning to address the issues of incomplete evaluation function and dependence on global references in unknown environments. The algorithm extends the evaluation function and uses Q-learning to adaptively learn parameters, significantly enhancing navigation efficiency and success rate in complex unknown environments. Zhang J et al. [18] optimized the evaluation mechanism by integrating adaptive parameters and a Vector Field Histogram (VFH), evaluated the performance by combining a real-world dynamic model, and designed a velocity controller to form a hierarchical control scheme. Zhang B et al. [19] obtained data based on Geant4 modeling. The results show that in small-scale scenarios, the improved A* algorithm reduces turning points, shortens calculation time, and has slightly lower dosage compared with the traditional A*. In complex multi-γ-source environments, compared with both the traditional A* and PRM algorithms, the improved A* achieves comparable path dosage while significantly reducing the number of turns and calculation time. Wu B et al. [20] proposed an improved hybrid algorithm combining A* and DWA. The enhanced A* is used to plan adaptive global paths, while the improved DWA integrates a rolling window to optimize local paths and perform obstacle avoidance. In a study conducted by Liu W et al. [21], to address the issues of path redundancy in the traditional A* algorithm and the tendency of the artificial potential field (APF) algorithm to become stuck in local optima, an APF-A* fusion algorithm is proposed.

Stavrinidis S et al. [22] resolve NP-hard multi-goal navigation in dynamic environments via convex decomposition-based path planning and a Genetic Algorithm for optimal target sequencing, innovating a dual fuzzy controller system enabling adaptive collision-free navigation. The integrated TSP–fuzzy framework validated in CoppeliaSim demonstrates robust obstacle avoidance, establishing a new paradigm for real-time multi-objective navigation.

Yin Y et al. [23] propose an SCF framework (curriculum-prioritized SAC with fuzzy logic) where the CEP mechanism enhances sample efficiency/safety and the fuzzy controller remedies dynamic avoidance limits, resulting in superior performance in dynamic environments. Qin Z et al. [24] propose a composite fuzzy distributed control law for cooperative control of fuzzy multi-agent systems with a nonlinear leader via distributed observers, making three contributions: extending the small-gain theorem to fuzzy systems (proposing IOS and OL–IOS fuzzy small-gain theorems), presenting the OL–IOS small-gain theorem for multi-loop fuzzy systems, and constructing the control law integrating a pure fuzzy distributed controller and a distributed observer for the leader. Jin, Z et al. [25] presented a distributed optimization problem (DOP) of T–S fuzzy cyber–physical systems under weight-balanced graphs and quasi-strongly connected frameworks, aiming to drive all agents’ outputs to the global objective function’s optimal solution via partial local information. It uses distributed optimal coordinators (DOCs) to generate converging local optimal solutions and designs fuzzy reference-tracking controllers for tracking, proposes two Lyapunov-based fuzzy ISS small-gain theorems, applies the fuzzy ISS cyclic small-gain theorem for stability analysis, solves the DOP, and verifies effectiveness via a numerical example. H Li et al. [26], in their work, propose a collaborative source-seeking algorithm using communication-constrained sensor-equipped quadrotors, combining relevant controllers and algorithms to maintain formation while driving the average position to the scalar field’s optimal point, with its convergence and robustness proven via analysis and verified through experiments and simulations.

Although existing studies have improved path planning algorithms for Automated Guided Vehicles (AGV), most optimized paths still suffer from suboptimal solutions and frequently fail to avoid obstacles in complex and narrow environments, thereby reducing system safety. To address these issues, this paper proposes a hybrid path planning algorithm that integrates the improved A* algorithm with the fuzzy-controlled Dynamic Window Approach (DWA). Leveraging the advantages of fuzzy-controlled DWA in local obstacle avoidance and dynamic real-time adjustment, the algorithm meets the navigation requirements of narrow spaces. The improved A* and fuzzy DWA are, respectively, responsible for global path planning and local path optimization, jointly generating the final path. Compared with previous research, the contributions of this study are as follows:

• The improved A* algorithm incorporates dynamic weight factors, integrating actual costs with estimated costs to significantly enhance the flexibility of the evaluation function;
• Cubic quasi-uniform B-spline curves are applied to effectively ensure local path smoothness and substantially reduce unnecessary turns;
• The path generated by the improved A* algorithm serves as the global reference for the Dynamic Window Approach (DWA);
• Sub-target points are selected globally through a rolling window mechanism, guiding the DWA to avoid local minima and maintain real-time navigation performance;
• A fuzzy controller is introduced to dynamically adjust the weight parameters of heading angle, obstacle avoidance strategy, and speed control in real time, thereby reducing the number of path inflection points in narrow environments.

2. Improved A* Algorithm

2.1. Traditional A* Algorithm

The A* algorithm is a heuristic algorithm designed to find the shortest path. It performs a directed search using known maps and selects paths through a cost estimation function. The cost estimation function of the A* algorithm is expressed as

$$f\left(n\right)=g\left(n\right)+h\left(n\right)$$

(1)

$$g\left(n\right)=\sqrt{{\left(x_i-x_j\right)}^2+{\left(y_i-y_j\right)}^2}$$

(2)

$$h\left(n\right)=\sqrt{{\left(x_j-x_k\right)}^2+{\left(y_j-y_k\right)}^2}$$

(3)

In the pathfinding process, the evaluation function plays a crucial role, where $f(n)$ denotes the cumulative estimated total cost from the start node to the end node via node N, $g(n)$ represents the Euclidean distance from the start node to node n, and $h\left(n\right)$ is the Euclidean distance from node n to the end node.

2.2. Improved A* Algorithm

In complex and narrow environments, traditional heuristic functions are often affected by dynamic obstacles and complex environmental factors, leading to the generation of suboptimal paths. To enhance the algorithm’s search efficiency, an appropriate heuristic function is selected. Therefore, the heuristic function is adjusted by introducing a weight factor $w$, expressed as follows:

$$w=1+a·\left({\displaystyle \frac{R}{r}}\right)$$

(4)

Here, R represents the distance from the start point to the target point, r denotes the distance from the current node to the target point, and a is a fixed weight value. The improved evaluation function is expressed as

$$f\left(n\right)=g\left(n\right)+w·h\left(n\right)$$

(5)

The dynamic adjustment of the weight factor $\mathrm{w}$ significantly enhances the estimation accuracy of the heuristic function for the actual path cost from the current node to the target node. This adaptive adjustment is particularly critical in environments with obstacles or terrain variations, effectively mitigating the overestimation or underestimation of path lengths. Consequently, the optimized heuristic function can more efficiently guide the search process to avoid obstacles. By introducing the ratio of the traversed distance to the remaining Euclidean distance, this weighting mechanism reduces excessive reliance on straight-line distances and minimizes their misleading effects on the search path.

2.3. Path Smoothing Optimization Based on Cubic Quasi-Uniform β-Spline

To address the issue of redundant inflection points in paths generated by the A* algorithm in grid environments, and to enhance path smoothness while reducing collision risks during movement, a local path optimization method using cubic quasi-uniform β-spline curves is proposed.

Given $\left(\mathrm{k}+1\right)$ control points, the degree k β-Spline curve is defined as

$$P_k\left(x\right)=\sum _{j=0}^k{P}_j{J}_{j,m}\left(x\right) \left(0\le x\le 1\right)$$

(6)

where ${\mathrm{P}}_{\mathrm{j}}$ are the control points, and the basis function ${\mathrm{J}}_{\mathrm{j},\mathrm{m}}\left(\mathrm{x}\right)$ satisfies

$$J_{j,m}\left(x\right)={\displaystyle \frac{1}{m!}}\sum _{i=0}^m{\left(-1\right)}^i\left({\displaystyle \genfrac{}{}{0pt}{}{m}{i}}\right){\left(x+m-j-i\right)}^m$$

(7)

For degree m = 3, the analytical expression of the cubic β-Spline basis function is

$$\left\{\begin{array}{c}{J}_{\mathrm{0,3}}\left(x\right)=-{\displaystyle \frac{1}{6}}{x}^3+{\displaystyle \frac{1}{2}}{x}^2-{\displaystyle \frac{1}{2}}x+{\displaystyle \frac{1}{6}}\\ J_{\mathrm{1,3}}\left(x\right)={\displaystyle \frac{1}{2}}{x}^3-x^2+{\displaystyle \frac{2}{3}}\\ J_{\mathrm{2,3}}\left(x\right)=-{\displaystyle \frac{1}{2}}{x}^3+{\displaystyle \frac{1}{2}}{x}^2+{\displaystyle \frac{1}{2}}x+{\displaystyle \frac{1}{6}}\\ J_{\mathrm{3,3}}\left(x\right)={\displaystyle \frac{1}{6}}{x}^3\end{array}\right.$$

(8)

Uniform and non-uniform B-spline curves fail to cover the start and end points of the path during optimization. The same issue exists in quadratic B-spline curves, while piecewise Bézier curves exhibit smoothness within control points but suffer from discontinuity problems. Therefore, cubic quasi-uniform B-spline curves with continuity characteristics are selected for path optimization to generate a smooth path from the start point to the end point.

3. Hybrid Improved A* and Fuzzy-Controlled Dynamic Window Approach

3.1. AGV Kinematic Model

In the traditional Dynamic Window Approach (DWA) algorithm, it is necessary to first establish the kinematic model before path planning. The position of the Automated Guided Vehicle (AGV) at the initial time t₁ is X_t1 = $\left({\mathrm{x}}_{\mathrm{t}1},{\mathrm{y}}_{\mathrm{t}1},{\mathsf{\theta }}_{\mathrm{t}1}\right)$, and its position state at time t₂ is Xt₂ = $\left({\mathrm{x}}_{\mathrm{t}2},{\mathrm{y}}_{\mathrm{t}2},{\mathsf{\theta }}_{\mathrm{t}2}\right)$, as shown in Figure 1.

The coordinates of the instantaneous center O during AGV turning are (x₀,y₀), where a is the wheelbase, R is the arc radius, α is the arc rotation angle, and β is the heading deflection angle. Let v_t and v_r denote the rotational speeds of the left and right drive wheels, respectively, and v and ω represent the linear and angular velocities at the vehicle center. Since the position of the smart cart at time t is known, the coordinates of center O can be obtained by the following equation:

$$x_0=x_{t1}-{\displaystyle \frac{v}{w}}sin{\theta }_{t1}$$

(9)

$$y_0=y_{t1}-{\displaystyle \frac{v}{w}}cos{\theta }_{t1}$$

(10)

The kinematic model is expressed as

$$\left[\begin{array}{c}{x}_{t_2}\\ y_{t_2}\\ {\theta }_{t_2}\end{array}\right]=\left[\begin{array}{c}{x}_{t_1}\\ y_{t_1}\\ {\theta }_{t_1}\end{array}\right]+\left[\begin{array}{c}-{\displaystyle \frac{v}{\omega }}sin{\theta }_{t_1}+{\displaystyle \frac{v}{\omega }}sin\left({\theta }_{t_1}+\omega \Delta t\right)\\ {\displaystyle \frac{v}{\omega }}cos{\theta }_{t_1}-{\displaystyle \frac{v}{\omega }}cos\left({\theta }_{t_1}+\omega \Delta t\right)\\ \omega \Delta t\end{array}\right]$$

(11)

3.2. Speed Sampling and Trajectory Prediction

The Dynamic Window Approach (DWA) determines feasible speed ranges in real-time based on the local environmental information of the AGV’s navigation area and its own kinematic constraints. Within this dynamic window, speed combinations are discretely sampled according to predefined linear velocity (v) and angular velocity (ω) resolutions. Subsequently, these sampled speeds are used to simulate the AGV’s motion trajectories over a short time interval. The simulated trajectories are quantitatively evaluated by a trajectory assessment function, and the speed combination corresponding to the highest-scoring trajectory is ultimately selected as the optimal control speed for the AGV, as shown in Figure 2.

In the figure, Trajectories 1 and 5 are generated by maximum linear deceleration and angular acceleration, exhibiting the shortest length and maximum curvature. Trajectories 2 and 4 are formed under maximum linear acceleration and angular acceleration, showing the longest length and moderate curvature. Trajectory 3, generated at constant linear and angular velocities, is a longer approximate straight line.

To filter out feasible simulated trajectories, an evaluation function is constructed to obtain the optimal trajectory. The evaluation function is expressed as

$$G\left(\nu ,\omega \right)=\sigma \left[\alpha \cdot H\left(\nu ,\omega \right)+\beta \cdot D\left(\nu ,\omega \right)+\gamma \cdot V\left(\nu ,\omega \right)\right]$$

(12)

Here, $H\left(\nu ,\omega \right)$ denotes the heading angle evaluation function, $D\left(\nu ,\omega \right)$ represents the obstacle avoidance safety evaluation function, and $V\left(\nu ,\omega \right)$ signifies the velocity evaluation function. $\sigma$ is the smoothing coefficient, while $\alpha$, $\beta$, and $\gamma$ are the weighting coefficients.

(1)

Heading Angle Evaluation Function $H\left(\nu ,\omega \right)$

$\mathrm{H}\left(\mathsf{\nu },\mathsf{\omega }\right)$ is used to evaluate the alignment degree between the heading angle at the terminal position of the simulated trajectory and the direction towards the target point. The specific expression is

$$H\left(\nu ,\omega \right)={180}^{\circ }-\theta$$

(13)

(2)

Obstacle Avoidance Safety Evaluation Function $D\left(\nu ,\omega \right)$

$\mathrm{D}\left(\mathsf{\nu },\mathsf{\omega }\right)$ is used to evaluate the minimum distance between points on the simulated trajectory and the nearest obstacle, quantifying the safety of the trajectory. The specific expression is

$$D\left(\nu ,\omega \right)=min\left(d_{min},d_{set}\right)$$

(14)

Here, ${\mathrm{d}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$ denotes the minimum distance between points on the simulated trajectory and the nearest obstacle and ${\mathrm{d}}_{\mathrm{s}\mathrm{e}\mathrm{t}}$ represents the predefined maximum distance value.

(3)

Velocity Evaluation Function $V\left(\nu ,\omega \right)$

$\mathrm{V}\left(\mathsf{\nu },\mathsf{\omega }\right)$ is used to evaluate the mobility efficiency of the mobile robot along the simulated trajectory, reflecting the absolute value of its linear velocity. The specific expression is

$$V\left(\nu ,\omega \right)=\left|v_l\right|$$

(15)

Here, ${\mathrm{v}}_{\mathrm{l}}$ denotes the linear velocity of the mobile robot along the simulated trajectory. The greater the linear velocity, the higher the mobility efficiency of the mobile robot, enabling it to reach the target point more quickly.

3.3. Fuzzy Control

The fixed weight strategy of the traditional Dynamic Window Approach (DWA) tends to fall into local optima in complex dynamic environments, leading to reduced path planning efficiency, trajectory oscillation, and even stagnation. Its limitations include the following: ① fixed weights lack adaptability to scenarios such as narrow passages and dense obstacles, where obstacle avoidance weights may cause the robot to decelerate conservatively or approach obstacles aggressively; ② static weights struggle to respond to sudden environmental changes like moving obstacles, weakening real-time path optimization capabilities. To address these bottlenecks, a fuzzy logic controller is introduced to achieve dynamic adaptive adjustment of evaluation function weights. Taking obstacle distance and target distance as input variables, the controller outputs corresponding heading weight, obstacle avoidance weight, velocity weight, and distance correction weight, constructing a mapping between environmental states and weight configurations through 16 fuzzy rules to enable real-time strategy optimization in complex scenarios. The fuzzy control rules are shown in

Table 1. Fuzzy control rules.

Number	Input		Output
	D_obs	D_goal	α	β	γ	λ
1	VS	VS	S	S	M	S
2	VS	S	S	M	M	S
3	VS	M	S	M	M	S
4	VS	B	S	B	B	S
5	S	VS	M	VS	M	M
6	S	S	M	S	M	M
7	S	M	S	M	M	M
8	S	B	S	B	B	M
9	M	VS	M	VS	M	VS
10	M	S	M	VS	M	S
11	M	M	M	VS	M	M
12	M	B	M	VS	B	B
13	B	VS	B	VS	S	VS
14	B	S	B	VS	S	VS
15	B	M	B	VS	M	S
16	B	B	B	VS	B	B

.

The fuzzy sets for input variables in the table are {VS, S, M, B}, corresponding to “very close, close, moderate, far” in distance; those for output variables are {VS, S, M, B}, representing “very small, small, moderate, large” in weight values. Taking Rule 3 as an example, when the robot is very close to obstacles and the target point is moderately distant, it should moderately abandon tracking the global path and target to prioritize obstacle avoidance for safety. According to the defined output rules, the heading weight α is assigned a very small value, the obstacle avoidance weight β a maximum value, the velocity weight γ a moderate value, and the distance correction weight λ a very small value. This configuration ensures the robot prioritizes obstacle avoidance at a moderate speed. The parameters α, β, γ, and λ outputs are shown in Figure 3, Figure 4, Figure 5 and Figure 6. The membership function graphs are shown in Figure 7.

The proposed integrated algorithm combines the global optimality of the A* algorithm with the dynamic adjustment capability of the Dynamic Window Approach (DWA). This fusion enables the system to effectively respond to dynamic environmental changes and perform real-time obstacle avoidance, allowing AGVs to plan safe, efficient, and stable paths in complex and narrow environments. The flow chart of the integrated algorithm is shown in Figure 8.

4. Simulation Experiment and Analysis

4.1. Simulation

To validate the feasibility of the proposed improved A* algorithm, comparative experiments were conducted against the traditional A* algorithm on grid maps of varying scales (20 × 20 and 30 × 30). To further evaluate the hybrid algorithms, the RRT-DWA and PSO-DWA approaches were benchmarked using MATLAB R2024b for AGV path planning simulations. The simulation environment was configured on a workstation equipped with Windows 11, an Intel Core i9 processor (2.0 GHz), and 8 GB RAM. A grid-based map representation was adopted, where black cells denote obstacles, white cells represent traversable paths, the green circle marks the starting point, and the blue triangle indicates the destination. Comparative simulations of different algorithms in a 20 × 20 environment are shown in Figure 9, Figure 10 and Figure 11 and comparative simulations of different algorithms in a 30 × 30 environment are shown in Figure 12, Figure 13 and Figure 14.

Across both 20 × 20 and 30 × 30 map sizes, the path length planned by the improved A* smoothing algorithm is slightly longer than that of the traditional A* algorithm, suggesting that the traditional A* algorithm may have a slight edge in finding the theoretical shortest path. However, the runtime of the improved A* smoothing algorithm is significantly shorter than that of the traditional A* algorithm in both scenarios. For the 20 × 20 map, the improved algorithm achieves a runtime of 0.028 s, marking a ~34.9% reduction compared to the traditional method. For the 30 × 30 map, the improved algorithm’s runtime is 0.08 s, representing a ~55.6% increase in speed. Notably, the time efficiency advantage of the improved algorithm becomes more pronounced as the map scale increases.

The data in

Table 2. Comparison of the performances of different algorithms.

Map Specification	Algorithm	Path Length	Running Time/s
20 × 20	Traditional A* Algorithm	27.87	0.043
	Improved A* Smoothing Algorithm	28.31	0.028
30 × 30	Traditional A* Algorithm	54.62	0.18
	Improved A* Smoothing Algorithm	54.92	0.08

demonstrate that the improved A* smoothing algorithm exhibits a remarkable advantage in operational efficiency, particularly for larger-scale (30 × 30) path planning, where its runtime is less than half of the traditional A* algorithm. Although the path length increases slightly, the magnitude of this increase is relatively small. Thus, in scenarios demanding rapid response or handling complex maps, the improved A* smoothing algorithm emerges as a superior choice. By introducing strategies such as smoothing processing, it achieves a substantial boost in execution speed at the acceptable cost of a minor increase in path length.

Figure 15, Figure 16 and Figure 17 illustrate the path comparison in terms of length, smoothness, and obstacle avoidance among the proposed fused algorithm, the hybrid RRT-DWA algorithm, and the hybrid PSO-DWA algorithm. The experimental data in Table 3 clearly demonstrate that the fusion-enhanced A* and fuzzy DWA algorithm exhibits significant advantages across multiple core metrics, outperforming the comparative hybrid RRT-DWA and hybrid PSO-DWA algorithms. These performance improvements are primarily manifested in path quality, planning efficiency, and motion smoothness. The fusion-enhanced A* and fuzzy DWA achieved the shortest path length. Compared to hybrid RRT-DWA, it was reduced by approximately 5.3%. Similarly, it was shortened by approximately 5.1% relative to hybrid PSO-DWA. The fusion-enhanced A* and fuzzy DWA exhibited the fastest computation time. Its runtime was improved by approximately 3.7% compared to hybrid RRT-DWA and by a more substantial 11.6% compared to hybrid PSO-DWA. Paths generated by the fusion-enhanced A* and fuzzy DWA algorithm possessed the fewest inflection points. The number was reduced by 7.4% compared to hybrid RRT-DWA and by 3.8% compared to hybrid PSO-DWA algorithms. Although the cumulative turning angle for the fusion-enhanced A* and fuzzy DWA algorithm was slightly higher than those of hybrid RRT-DWA and hybrid PSO-DWA, its significantly fewer inflection points suggest that its paths are generally more direct. This implies a reduction in frequent small-angle adjustments; while individual turns might be slightly larger or occur in specific complex segments, the overall number of steering maneuvers is likely lower. Consequently, the overall path smoothness and executability remain competitive. Figure 18, Figure 19, Figure 20, Figure 21 and Figure 22 illustrate the variations in the AGV’s linear velocity, angular velocity, and heading direction throughout the process. The linear velocity profile reveals that the AGV first decelerates upon encountering obstacles, then stabilizes, and finally reaches the destination safely.

Table 3. Comparison of the performances of different algorithms.

Algorithm	Path Length	Running Time/s	Inflection Points/No.	Turning Angle/Deg
Fusion-Enhanced A* and Fuzzy DWA	26.313	101.0338	75	95
Hybrid RRT-DWA Algorithm	27.782	104.826	81	80.2
Hybrid PSO-DWA Algorithm	27.714	114.3049	78	77.2

Table 3. Comparison of the performances of different algorithms.

Algorithm	Path Length	Running Time/s	Inflection Points/No.	Turning Angle/Deg
Fusion-Enhanced A* and Fuzzy DWA	26.313	101.0338	75	95
Hybrid RRT-DWA Algorithm	27.782	104.826	81	80.2
Hybrid PSO-DWA Algorithm	27.714	114.3049	78	77.2

4.2. Experiment

To validate the feasibility of the improved fusion algorithm in practical scenarios, simulations and experiments were conducted within the Robot Operating System (ROS) framework. ROS, originally developed by Willow Garage, is a flexible software—ubuntu18.04 framework tailored for robotic applications. Rviz, a 3D visualization tool integrated into ROS, enables the visualization of sensor data, thereby facilitating intuitive development and validation processes. For the proposed algorithm, comparative evaluations between simulation and experimental results were performed on the Rviz platform to comprehensively assess its performance.

To validate the algorithm’s feasibility in real-world environments, simulations and experiments were conducted using a ROS-based experimental platform. First, path simulations of the proposed improved algorithm were performed in Gazebo, as shown in Figure 23. Subsequently, real-world environment emulations were carried out via the experimental platform,. The architecture of the experimental platform, depicted in Figure 24, utilizes LiDAR for environmental mapping, with the constructed map visualized in Figure 25. A real-world environment simulation map is shown in Figure 26.

From

Table 4. Comparison of the performances of different situation.

Situation	Path Length	Running Time/s	Inflection Points/No.
Gazebo Robotics Simulator	2.6	0.6	6
experiment	2.8	0.75	7

, it can be inferred that the deviations between the experimental and simulated results are approximately 7.14% for path length, 20% for running time, and 14.3% for motion trajectory smoothness. This suggests that the improved algorithm is well-adapted to complex environments with stringent requirements for real-time capability, path efficiency, and motion continuity.

5. Conclusions

This study addresses the path planning challenge of Automated Guided Vehicles in complex and narrow environments. To enhance global path optimality and mitigate inefficient obstacle avoidance, a hybrid path planning method integrating an improved A* algorithm with a fuzzy-controlled Dynamic Window Approach is proposed. The improved A* algorithm incorporates several modifications: (1) a dynamic weight factor is introduced to fuse actual and heuristic costs, enabling a more flexible evaluation function; (2) cubic quasi-uniform B-spline curves are adopted to ensure local path smoothness and reduce redundant inflections; (3) the path generated by the improved A* serves as the reference trajectory for DWA; (4) a rolling window mechanism is employed to select sub-goals globally, guiding DWA to avoid local traps while maintaining real-time performance; (5) a fuzzy controller is integrated to dynamically adjust the weights of heading, obstacle avoidance, and velocity, reducing path inflections in narrow scenarios. Experimental results demonstrate that the proposed algorithm shortens path length and planning time, while significantly enhancing path smoothness and obstacle avoidance capability. By combining the A* and DWA algorithms, this study provides a feasible solution for AGV path planning in complex and narrow environments, with simulations and experiments verifying the stability and reliability of the planned paths.