Obstacle avoidance for three-dimensional environmental robotic arms based on SGBM binocular stereo matching

In complex industrial operation environments, robotic arms need to achieve precise environmental perception and safe obstacle avoidance under the coexistence of occlusion, weak texture and dynamic obstacles. To enhance the real-time performance and stability of three-dimensional spatial recognition, obstacle location, and path planning of robotic arms, a method for obstacle avoidance of three-dimensional environment robotic arms based on binocular stereo matching of semi-global block matching algorithm is constructed. Based on the 11th version of the real-time object detection network, the Convolutional Kolmogorov-Arnold Networks, Semantic and Detail Interaction, Bidirectional Feature Pyramid Network are introduced to enhance the nonlinear feature expression of obstacles, multi-scale fusion, and boundary positioning capabilities. Subsequently, the semi-global block matching algorithm is adopted to restore the binocular parallax and depth information, and the obstacle avoidance path planning is completed by combining the threedimensional point cloud of obstacles, the kinematic model of the robotic arm and the safety distance constraint. The average accuracy rate of the environmental detection model reached 98.21%, the precision rate reached 98.37%, the recall rate reached 97.42%, the positioning error was reduced to 1.26 cm, which was 1.58 cm less than that of the original detection model, and the frame rate still remained at 60.7, meeting the real-time detection requirements. In the three-dimensional obstacle avoidance experiment, the average path length of the proposed method was 1.83 m, the collision rate was reduced to 1.84%, the minimum safe distance was increased to 0.21 m, the average response time was only 48.3 m, the dynamic obstacle avoidance success rate reached 98.27%, and the task completion rate reached 98.64%. The research results show that this method can effectively improve the perception accuracy, path safety and dynamic obstacle avoidance stability of the robotic arm in complex three-dimensional environments.

In complex industrial scenarios, robotic arms are expanding from repetitive execution at fixed workstations to tasks such as flexible manufacturing, intelligent assembly, warehousing and handling, and human-machine collaboration. The types, spatial positions, and movement states of obstacles in the working environment are showing greater uncertainty (Cao et al., 2023). Traditional robotic arms mostly rely on preset trajectories, fixed safety boundaries or manually calibrated environments to complete motion control. They can maintain high execution efficiency in structured scenarios, but when facing situations such as workpiece stacking, equipment occlusion, weak texture backgrounds, dynamic personnel entry and changes in obstacle positions, they are prone to insufficient environmental perception, lagging path planning and local collision risks. Especially in the three-dimensional working space, the robotic arm not only needs to identify the category and two-dimensional position of obstacles, but also accurately obtain the depth distance, spatial profile and distribution of passable areas between the obstacles and the end effector (Yueli et al., 2023).

Kim M et al. proposed a hybrid path planning strategy based on the combination of GRU surrogate dynamics and cross-entropy method and enhanced by automatic differential in response to the challenge of efficient path planning for spacecraft in non-convex constraint environments. This method converged faster and cost less than the standard cross-entropy method. Moreover, under similar computational budgets, it outperformed the SCP and linearized cross-entropy method dynamics MPC, and could generate more accurate, fuel-efficient and collision-free trajectories (Kim et al., 2026). Chen Y et al. proposed a fast exploration random tree algorithm guided by random geometric graphs for the path planning of efficient and safe inspection of Unmanned Aerial Vehicles (UAVs) in the three-dimensional environment of complex substations. 3D occupation grid graph construction, dynamic target bias sampling, and trajectory smoothing optimization strategies were integrated. The results showed that this method outperformed the mainstream algorithms in path length, planning time consumption, and trajectory smoothness, significantly improving the quality of inspection trajectories and operational safety (Chen et al., 2026). He S et al. designed a hybrid path planning method that integrated multi-strategy potential field ant colony optimization and elastic dynamic window method to enhance the path planning ability of wheeled mobile robots in complex environments. Experiments confirmed that this method outperformed traditional algorithms in convergence speed, safety distance, path redundancy, and environmental adaptability, demonstrating significant practical application value (He et al., 2026). To sum up, the above research is prone to problems such as insufficient response speed, numerous path turns and low safety margin in dynamic obstacles and complex occlusion environments. Li J et al. developed a machine vision-based obstacle avoidance optimization method for dual-arm cooperative robots, where binocular vision was employed for obstacle perception and trajectory optimization. However, the method mainly focused on path generation after obstacle recognition and lacked deeper integration between semantic perception, depth reconstruction and motion decision (Li et al., 2025). Furthermore, Scholz C et al. summarized sensor-enabled safety systems for human-robot collaboration and pointed out that reliable environmental perception and rapid response remain challenging under dynamic and uncertain industrial conditions. These studies indicate that although vision-based robotic safety control has achieved significant progress, constructing a unified framework integrating high-precision obstacle recognition, binocular depth recovery and three-dimensional avoidance planning is still required (Scholz et al., 2025). Xie et al. developed a binocular vision-based road condition detection system in an Internet-of-Things environment, where disparity information and lightweight deep learning models were combined to achieve reliable road structure perception and passable area detection. Another study proposed an intelligent binocular vision sensor for automotive perception systems, which utilized stereo cameras to obtain point cloud information and perform obstacle perception and environmental localization. These methods provide valuable solutions for real-time stereo perception in large-scale outdoor environments (Xie et al., 2023).

Therefore, the research aims to construct the Convolutional Kolmogorov-Arnold Networks-Semantic and Detail Interaction-Bidirectional Feature Pyramid Network-You Only Look Once version 11 (KAN-SDI-BiFPN-YOLOv11) for robotic arm environment detection and localization. Combined with the Semi-Global Block Matching (SGB) algorithm, 3D environment reconstruction and obstacle avoidance path planning are achieved. This method enhances the nonlinear feature expression of complex obstacles through KAN, strengthens the interaction between shallow edge details and deep semantic information through SDI, and improves the feature fusion efficiency of multi-scale obstacles through BiFPN. The disparity map and depth map are generated using SGBM, mapping the two-dimensional detection results to the three-dimensional space. Then, the path generation and optimization are completed by combining the kinematic constraints of the robotic arm, safety distance evaluation and dynamic feedback mechanism. The research aims to provide technical support for the autonomous obstacle avoidance control of industrial robotic arms in scenarios such as assembly, handling, sorting, and human-machine collaboration. Different from the conventional sequential combination of independent perception, depth estimation and path planning modules, the proposed framework establishes a task-level closed-loop interaction mechanism among environmental detection, binocular depth reconstruction and obstacle avoidance planning. Specifically, the detection module provides obstacle semantic information and adaptive matching regions for SGBM depth recovery, while the reconstructed three-dimensional spatial information is further introduced into safety distance evaluation and trajectory optimization. During robotic arm execution, environmental changes and trajectory feedback are used to update obstacle perception and planning decisions. Although the proposed framework does not adopt end-to-end joint network training, it achieves cross-module information interaction through shared spatial constraints and feedback-driven optimization, thereby improving the consistency between perception accuracy and motion safety in complex three-dimensional environments.

Construction of environmental detection and positioning model for robotic arms

In the trajectory planning process, the kinematic characteristics and physical constraints of the robotic arm are considered to ensure the feasibility of the generated obstacle avoidance trajectory. The experimental platform adopts a six-degree-of-freedom industrial robotic manipulator, including three positioning joints and three wrist rotation joints, which can complete spatial translation and attitude adjustment of the end effector. The Denavit-Hartenberg parameter model is used to describe the relationship between adjacent joints, and the forward kinematic relationship between the base coordinate system and the end-effector coordinate system is established according to the joint configuration. During trajectory optimization, the joint rotation range, workspace boundary, maximum joint velocity and acceleration limitations are incorporated as motion constraints to prevent unreasonable or unstable movements caused by excessive joint changes. Furthermore, collision constraints are considered from both external obstacle avoidance and internal self-collision prevention perspectives. The geometric structure of the robotic arm links is modeled as collision bodies, and the minimum distance between the manipulator links and surrounding obstacles is continuously evaluated during path generation. The trajectory is regarded as feasible only when the distance between the robotic arm and obstacles satisfies the predefined safety margin. Meanwhile, interference between adjacent and non-adjacent links is checked to avoid self-collision during complex spatial movements. Through the integration of kinematic constraints and collision constraints, the generated trajectory can simultaneously satisfy environmental safety requirements and the mechanical execution capability of the robotic arm.

In complex three-dimensional working environments, the primary task of obstacle avoidance for robotic arms is to accurately identify surrounding obstacles, target workpieces, and passable spaces, providing reliable perception input for subsequent binocular depth estimation and path planning. In scenarios where occlusion, weak texture, edge blurring and multi-scale obstacles coexist, traditional object detection models are prone to problems such as missed detection of small targets, boundary positioning deviation and insufficient feature expression, which in turn affect the safety and real-time performance of obstacle avoidance decisions of robotic arms. Therefore, the research aims to construct an environmental detection model for robotic arms based on YOLOv11. This model is based on the YOLOv11 detection framework and introduces the Convolutional Kolmogorov-Arnold Network (KAN) to enhance the expression ability of nonlinear features. The Semantic and Detail Interaction (SDI) is utilized to fuse shallow texture and deep semantic information, and the bidirectional feature pyramid network is combined to improve the feature transmission efficiency of multi-scale obstacles. The structures of KAN and BiFPN are shown in Figure 1.

From Figure 1, the feature mapping process of the KAN module is shown in Equation 1.

In Equation 1, Y represents the output of the KAN node. x i represents the i -th input feature. ϕ i represents a learnable nonlinear mapping function. n represents the input feature dimension. After the combined processing of convolution and KAN, the expression of the feature map is shown in Equation 2 (Nemade and Pujeri, 2023).

In Equation 2, P l represents the output feature of the l -th layer. P l · represents the feature extraction function of the corresponding layer. l represent the image of the working environment of the robotic arm. To further enhance the expressive power of small target obstacles and occlusion boundaries, BiFPN performs weighted fusion on features of different scales, and its expression is shown in Equation 3.

In Equation 3, P ^ l represents the feature of the l -th layer after fusion. P k represents the k -th scale feature involved in the fusion. w k represents the learnable weight. K represents the number of fused features. ε represents the stability coefficient (Zhang et al., 2025). The final output feature is shown in Equation 4.

In Equation 4, P l o u t represents the final output feature. P l t d represents the top-down fusion feature. D o w n · represents the downsampling operation. ⊕ stands for feature fusion. o n v · stands for convolution operation. Through the nonlinear expression of KAN and the bidirectional multi-scale fusion of BiFPN, the model can enhance the ability to identify obstacle categories, contour positioning and scale adaptability in complex backgrounds, providing stable visual features for three-dimensional obstacle avoidance of robotic arms (Zhao et al., 2024; Sunil and Afzal, 2023). Although KAN and BiFPN can enhance the expression of nonlinear features and the ability of multi-scale fusion, in the complex operation scenarios of robotic arms, the insufficient coupling between shallow edge textures and deep semantic information still easily occurs. Especially in small-scale obstacles, occluded targets and weak texture backgrounds, detailed features may gradually decay during multi-layer transmission, leading to unstable target boundary positioning. The SDI module is introduced, and its structure is shown in Figure 2.

As shown in Figure 2, the input image first passes through the encoder to extract feature information at different levels. The shallow features mainly include edge, texture and local contour information, while the deep features contain target category and spatial semantic information. Subsequently, the features of each layer are input into the SDI module, and the collaborative fusion of shallow details and deep semantics is achieved through multi-scale feature interaction. SDI internally combines upsampling, downsampling and convolutional mapping operations to uniformly represent features of different resolutions. It introduces spatial attention and channel attention mechanisms to strengthen key obstacle regions and important feature channels, thereby enhancing the expression ability of target boundaries and local structures in complex scenarios. The fused features are further input into the decoder for layer-by-layer recovery and output supervision, enabling the model to maintain high detection stability and environmental perception accuracy under obstacle occlusion, small target detection, and complex background conditions, providing a reliable visual feature basis for subsequent binocular stereo matching and three-dimensional obstacle avoidance of robotic arms. The structure of the KAN-SDI-BiFPN-YOLOv11 robotic arm environmental detection and positioning model based on YOLOv11 is shown in Figure 3.

From Figure 3, the data transmission process of the proposed KAN-SDI-BiFPN-YOLOv11 model can be divided into backbone feature extraction, semantic-detail interaction, multi-scale feature fusion and detection output stages. The input robotic arm environmental image is first transmitted into the YOLOv11 backbone network for hierarchical feature extraction. In this process, the original convolution operation in selected backbone layers is supplemented with KAN convolution rather than completely replacing the convolutional structure. The KAN module is embedded after the feature extraction layers to enhance the nonlinear mapping capability of intermediate feature representations, allowing the network to better characterize complex obstacle textures and boundaries. After backbone feature extraction, the multi-level feature maps containing different semantic information are transmitted to the SDI module. Unlike independent feature enhancement operations, SDI establishes interaction between shallow spatial details and deep semantic representations through cross-scale feature exchange. The output features from SDI are then delivered to the BiFPN structure, where bidirectional top-down and bottom-up fusion are performed to aggregate obstacle information at different scales. This process enables the network to simultaneously retain fine-grained boundary information of small obstacles and high-level semantic features of large obstacles. Finally, the fused multi-scale features are transmitted to the detection head to generate obstacle categories, bounding boxes and confidence scores. Therefore, the integration of KAN, SDI and BiFPN follows a hierarchical feature enhancement strategy rather than a simple parallel combination. KAN improves feature representation at the backbone stage, SDI enhances semantic-detail interaction between different levels, and BiFPN performs adaptive multi-scale aggregation before detection output, forming a complete perception pipeline for complex robotic arm environments.

Three-dimensional obstacle avoidance path planning for robotic arms based on SGBM binocular stereo matching

Based on the detection of obstacles, target workpieces and key areas in the working environment of the robotic arm, obtaining only the category and boundary information in the two-dimensional image is still difficult to meet the requirements of three-dimensional obstacle avoidance. The actual movement process of the robotic arm requires further clarifying the spatial distance, depth distribution and passable area position between the obstacle and the end effector (Dabestani et al., 2024; Pascariu et al., 2023). Therefore, the study introduces the Semi-Global Block Matching (SGBM) algorithm to restore the environmental depth using the disparity information between the left and right views of the binocular camera. The binocular ranging flow chart is shown in Figure 4.

From Figure 4, single-target determination is first performed on the left and right cameras, respectively. By obtaining the internal parameters, distortion coefficients, and imaging characteristics of the cameras, the influence of lens distortion on depth calculation is eliminated. Subsequently, the system jointly calibrates the binocular cameras, establishes the spatial geometric relationship between the left and right cameras, and calculates the rotation matrix, translation matrix, and binocular external parameters, thereby completing the unification of the binocular vision coordinate system (Zhou et al., 2023). After the parameter solution is completed, the left and right views enter the stereo matching stage. By searching and matching the pixels with the same name, the pixel displacement difference between the left and right images is calculated. The system further generates parallax maps based on parallax information, in which different gray-scale regions correspond to different spatial depths between the target and the camera. Obstacles that are relatively close usually have larger parallax values, while distant targets correspond to smaller parallax values. Ultimately, the parallax map can restore the three-dimensional spatial structure in the working environment of the robotic arm, providing a reliable depth information basis for subsequent obstacle distance measurement, spatial positioning, and three-dimensional obstacle avoidance path planning of the robotic arm. The binocular vision system used in this study consists of a calibrated stereo camera module installed on the robotic arm platform. The image acquisition resolution is set to 1280 × 720 pixels, and the camera pair is configured with a fixed baseline distance to ensure stable depth measurement during robotic motion. Before stereo matching, A standard checkerboard calibration target with a fixed grid arrangement is used for parameter estimation, and multiple image pairs are collected from different orientations and distances to improve calibration robustness. The calibration process is evaluated using reprojection error, and the average reprojection error remains below 0.15 pixels, indicating that the estimated camera parameters provide sufficient accuracy for subsequent stereo reconstruction. After calibration, stereo rectification is performed to align the epipolar lines of the left and right images, reducing vertical disparity and improving the matching reliability of SGBM under complex industrial environments. The model structure of SGBM is shown in Figure 5.

From Figure 5, the system first performs grayscale conversion, denoising and normalization on the left and right views to reduce the influence of lighting changes and noise interference on the accuracy of stereo matching. Subsequently, the matching cost between the left and right images is established through the Census transformation and pixel block matching, and its initial matching cost is shown in Equation 5.

In Equation 5, C x, y, d represents the matching cost of pixel x, y under parallax d. W represents the matching window area. I L and I R respectively represent the gray values of the left image and the right image. After the cost calculation is completed, SGBM uses a multi-path dynamic programming method to perform semi-global cost aggregation in eight directions, and the aggregated cost is shown in Equation 6.

In Equation 6, L r p, d represents the cumulative cost of pixel p on path r under parallax d. P 1 represents the penalty term for small parallax variation. P 2 represents the penalty term for large parallax variation. Subsequently, the costs of all paths are accumulated to obtain the final aggregated cost, as shown in Equation 7 (Ali et al., 2024; Wa et al., 2025).

In Equation 7, P 1 represents the final aggregation cost of pixel p under parallax d. v represents different aggregation directions. The system further adopts a winner-takes-all strategy to obtain the optimal parallax, and its expression is shown in Equation 8 (Alazzawi and Alsamer, 2024; Di Natale et al., 2024).

In Equation 8, d * represents the optimal parallax value. Finally, a complete disparity map is generated through left-right consistency detection, median filtering and hole filling. Based on the relationship between disparity and depth, the three-dimensional spatial structure is restored to achieve obstacle distance measurement and spatial positioning in the working environment of the robotic arm (Wang, 2025). Considering that the robotic arm operates under dynamic conditions, the standard SGBM process is further adapted to reduce the influence of motion-related disturbances on stereo matching. Before disparity computation, the collected binocular images are processed through distortion correction, illumination normalization and noise suppression to reduce image quality degradation caused by camera vibration and robotic arm movement. For short-term motion blur caused by rapid manipulator motion, the system utilizes consecutive image frames and temporal consistency between adjacent observations to eliminate unstable matching regions. During stereo matching, the Census transformation is adopted because it is less sensitive to illumination changes compared with direct gray-level matching, which improves the robustness of correspondence estimation in industrial environments with variable lighting. After semi-global cost aggregation, the generated disparity map is further processed through left-right consistency checking, median filtering and invalid disparity removal. Regions with large disparity fluctuations or unreliable matching confidence are identified as noise points and excluded from three-dimensional reconstruction. The final constructed model structure is shown in Figure 6.

From Figure 6, the working environment images of the robotic arm are first synchronously collected by the left and right binocular cameras, and the preprocessing operations such as grayscale conversion, distortion correction, noise reduction and image scaling are completed to improve the subsequent stereo matching accuracy. Subsequently, the system uses the SGBM algorithm to match the left and right images. Through matching cost calculation, semi-global cost aggregation, and disparity optimization, disparity maps and depth maps are generated to reconstruct three-dimensional spatial information of the environment. On this basis, the obtained obstacle detection results are mapped to the depth space to generate the three-dimensional point cloud of the obstacle and the spatial surrounding area, thereby obtaining the distance, position and spatial distribution information of the obstacle. The system further combines the DH parameters of the robotic arm, the boundary of the workspace, and the pose information of the target to establish a motion model of the robotic arm, and constructs a passable area through voxelized environment modeling. Subsequently, the obstacle avoidance trajectory is generated using global path planning and local path optimization algorithms, and the path is verified by combining collision detection and safety distance assessment simultaneously. Ultimately, the system outputs the joint movement trajectory of the robotic arm and conducts dynamic feedback and path re-planning based on real-time environmental changes, thereby achieving safe and stable three-dimensional obstacle avoidance control of the robotic arm in complex environments.

It should be noted that the proposed framework is not a simple sequential pipeline in which each module operates independently. Instead, a task-level interaction mechanism is constructed between perception, depth reconstruction and motion planning. The detection results generated by KAN-SDI-BiFPN-YOLOv11 are used to constrain the effective matching area and obstacle regions in SGBM stereo reconstruction, reducing unnecessary matching operations caused by irrelevant background regions. Meanwhile, the obtained depth information and three-dimensional obstacle distribution are introduced into safety distance calculation and trajectory optimization, enabling the planning module to adaptively adjust the motion strategy. Furthermore, the execution feedback of the robotic arm updates the environmental perception region, forming a closed-loop optimization process from visual perception to spatial reconstruction and motion control. Therefore, the contribution of this study is not the independent introduction of existing modules, but the construction of an integrated perception-depth-planning interaction framework for three-dimensional obstacle avoidance.

Performance of environmental detection and positioning model for robotic arm based on KAN-SDI-BiFPN-YOLOv11

The experimental platform adopts a Windows 11 workstation, with an Intel Core i9-13900K processor, 64 GB of memory and an NVIDIA RTX 4090 graphics card. To verify the target detection and spatial positioning capabilities of KAN-SDI-BiFPN-YOLOv11 in complex robotic arm operation environments, the study selected industrial assembly scenarios, dynamic obstacle scenarios, and complex occlusion scenarios as test environments. During model training, all detection networks were trained under the same experimental configuration to ensure a fair comparison. The input images were uniformly resized to 640 × 640 pixels before being fed into the network. The batch size was set to 16, and the AdamW optimizer was adopted for parameter updating. The initial learning rate was set to 0.001 and gradually adjusted using a cosine annealing strategy during training to improve convergence stability. A weight decay coefficient of 0.0005 was introduced to reduce overfitting. The training process adopted pretrained weights obtained from the original YOLOv11 model, and all compared models were fine-tuned under the same initialization strategy. To enhance the robustness of obstacle perception under complex industrial environments, data augmentation strategies including random scaling, horizontal flipping, brightness adjustment, and mosaic augmentation were applied during training. The validation set performance was monitored after each training epoch, and early stopping was enabled when the detection performance did not improve for 50 consecutive epochs. The final model was selected according to the highest validation mAP value and then evaluated on the independent test set. Comparative experiments were conducted with YOLOv11, BiFPN-YOLOv11, and KAN-BiFPN-YOLOv11. A binocular robotic arm environmental dataset was constructed for model training and evaluation. The dataset was collected from typical industrial operation scenarios, including assembly workspaces, material handling areas, equipment storage regions and dynamic obstacle environments. The image acquisition system consisted of a binocular camera mounted on the robotic arm platform, and the collected image pairs contained diverse obstacle conditions such as target occlusion, weak texture regions, different illumination conditions and multi-scale objects. A total of 12,600 binocular image pairs were collected, and the detection targets mainly included robotic obstacles, workpieces, shelves, tools and surrounding environmental objects. To ensure the diversity and reliability of the dataset, the collected samples were manually annotated according to obstacle categories and bounding box locations. The dataset was randomly divided into training, validation and test subsets with a ratio of 7:2:1. To avoid data leakage caused by highly similar images, images from the same continuous acquisition sequence were assigned to only one subset, ensuring that the test set contained independent environmental conditions that were not involved in model training. The training set was used for parameter optimization, the validation set was used for hyperparameter adjustment and model selection, and the test set was used for final performance evaluation. To analyze the target detection capability and training stability of the proposed model in a complex robotic arm environment, a comparative analysis was conducted on the Precision-Recall curves and the changing process of training losses of different models, as shown in Figure 7.

From Figure 7a, when the recall of KAN-SDI-BiFPN-YOLOv11 reached 0.8, the precision of YOLOv11 dropped to about 0.60, while KAN-SDI-BiFPN-YOLOv11 still remained above 0.86. When recall approached 0.95, YOLOv11 experienced a rapid decline, remaining at approximately 0.30, while the proposed model remained stable around 0.65. From Figure 7b, the loss values of each model gradually decreased with the increase of the number of training rounds, but there were significant differences in the convergence speed and final stable values among different models. The loss of YOLOv11 was still close to the order of 10 −2 at 200 rounds, while KAN-SDI-BiFPN-YOLOv11 rapidly dropped below 10 −2 after about 40 rounds and eventually stabilized around 3 × 10 −3, indicating that the proposed model has a faster parameter convergence ability and higher training stability. In conclusion, the KAN-SDI-BiFPN-YOLOv11 outperforms the comparison models in detection accuracy, training stability, and adaptability to complex scenarios, which is more suitable for three-dimensional environmental perception tasks of robotic arms. The performance of the model in actual situations was further analyzed, as illustrated in Figure 8.

From Figure 8a, the proposed model could relatively completely identify forklifts, shelves, distant obstacles and small targets on the right side. The detection box had a high degree of fit with the target contour, and no obvious missed detecments occurred. This indicates that the SDI module enhances the interaction between shallow edge details and deep semantic information, and improves the ability to locate small targets in complex backgrounds. From Figure 8b, KAN-BiFPN-YOLOv11 could identify the main obstacles, but the target box on the right side at a long distance was relatively small, and there was a shift in the local boundary, indicating that the detail recovery ability is insufficient after the lack of SDI. From Figure 8c, BiFPN-YOLOv11 showed target box expansion in the shelves and occluded areas, indicating that it is still difficult to accurately distinguish adjacent targets from the background merely by multi-scale fusion. From Figure 8d, YOLOv11 has missed detection of small targets and boundary offset. The reason lies in the limited feature extraction ability of the original network for weak textures, occlusions and distant targets. In conclusion, KAN-SDI-BiFPN-YOLOv11 performs more stably in complex environment detection of robotic arms. The ablation experiment was adopted to analyze each module of the model. The results are shown in Table 1.

As shown in Table 1, all results are reported as mean ± standard deviation over five independent runs. YOLOv11 achieved an mAP@0.5 of 91.87% ± 0.21% and a localization error of 2.84 ± 0.06 cm, indicating insufficient feature representation under complex obstacle conditions. After introducing BiFPN, the mAP@0.5 increased to 94.61% ± 0.17% and the localization error decreased to 2.21 ± 0.05 cm, demonstrating the effectiveness of multi-scale feature fusion. With the addition of KAN, the mAP@0.5 further increased to 96.23% ± 0.15%, while the localization error decreased to 1.73 ± 0.04 cm, indicating that nonlinear feature enhancement improves obstacle representation. The SDI module increased the recall to 96.24% ± 0.15%, showing better capability in preserving boundary details and reducing missed detection of occluded targets. The complete KAN-SDI-BiFPN-YOLOv11 achieved the best performance, with mAP@0.5 of 98.21% ± 0.12%, precision of 98.37% ± 0.11%, and localization error of 1.26 ± 0.03 cm. Compared with YOLOv11, the proposed model improved mAP@0.5 by 6.34 percentage points and reduced localization error by 1.58 cm. The low standard deviations and statistical significance results with p < 0.05 further verify the stability and effectiveness of the proposed feature enhancement strategy. Although the FPS decreased to 60.7, the model still satisfies real-time robotic arm perception requirements.

Performance of three-dimensional obstacle avoidance path planning for robotic arms based on SGBM binocular stereo matching

After detecting the working environment and obstacle location of the robotic arm, the research further verified the stability of SGBM binocular stereo matching in three-dimensional environment reconstruction and its supporting role in the obstacle avoidance path planning of the robotic arm. The experiments were conducted on the same six-degree-of-freedom industrial robotic manipulator described in the proposed framework. During trajectory execution, the controller simultaneously considered joint motion limitations, workspace boundaries, velocity and acceleration constraints, and self-collision avoidance requirements to ensure that the generated paths could be executed safely in the physical robotic system. The study selects A* and Rapidly-exploring Random Tree (RRT) path planning algorithms for comparison. The result is shown in Figure 9.

From Figure 9a, the path length of the A* algorithm was the highest overall. When the safety distance was approximately 0.10m, the path length generally remained above 3.5m, indicating that although it can maintain a relatively high safety margin, it is prone to problems such as longer paths and frequent turns. The path length of the RRT algorithm has decreased, with most points distributed between 1.5m and 2.8 m. However, the discreteness was relatively large, indicating that the random expansion mechanism led to insufficient path stability. In contrast, the path lengths of the proposed model were mainly concentrated between 0.8m and 2.2 m. It still maintained a relatively short path when the safety distance reached more than 0.30m, indicating that the research model can achieve a good balance between safety and path efficiency. From Figure 9b, after the response time of the A* algorithm reached 0.30 s, the collision rate rapidly rose above the order of 10^1, indicating that its real-time adaptability in dynamic environments is limited. The collision rate of the RRT algorithm was lower than that of A, but the overall fluctuation was relatively large. This is because the random search process is prone to cause instability of local trajectories. In contrast, the collision rate of the proposed model consistently remained around the order of 10–1. Even when the response time was close to 0.35 s, the collision rate was still significantly lower than that of the comparison model, indicating that the proposed model has a more stable dynamic obstacle avoidance ability. In conclusion, the proposed method outperforms the A and RRT algorithms in path efficiency, safe distance control, and dynamic obstacle avoidance stability, which is more suitable for complex three-dimensional obstacle avoidance scenarios of robotic arms. The depth error and point cloud reconstruction quality under different distances, occlusion ratios, and working scenarios were analyzed around the SGBM model, as shown in Figure 10.

From Figure 10a, with the continuous increase of the target distance and occlusion ratio, the overall Depth RMSE showed an upward trend. When the target distance was 0.6 m and there was no occlusion, the error was only 0.29 cm. However, when the distance increased to 3.4 m and the occlusion ratio reached 50%, the error rose to 1.11 cm, indicating that long-distance and highly occlusive environments will reduce the stability of binocular matching. Especially after 2.6m, the growth rate of error accelerated significantly. In Figure 10b, SGBM had the highest point cloud reconstruction quality in an open workspace, with a completeness rate close to 98%. Both boundary continuity and reconstruction stability remained above 96%. Under conditions of clear texture and less obstacle distribution, the model can obtain relatively stable depth information. In contrast, all the indicators in the occlusion scene and the dense obstacle scene have decreased, among which Hole suppression and Boundary continuity have decreased more significantly. In conclusion, SGBM can maintain relatively stable depth recovery and point cloud reconstruction performance in complex industrial environments, providing reliable spatial information for the three-dimensional obstacle avoidance path planning of robotic arms. The comprehensive performance of each model was further analyzed, as shown in Table 2.

From Table 2, the proposed model was generally superior to A* and RRT in the comprehensive performance of three-dimensional obstacle avoidance for robotic arms. The average path length of A* was 2.12 m, the collision rate was 6.25%, and the minimum safety distance was only 0.06 m. This indicates that although its path search is deterministic, the safety margin is insufficient in complex dynamic obstacle environments. The collision rate of RRT dropped to 3.57%, but the path smoothness index was 2.31, and the average response time reached 112.3 m, indicating that the random sampling mechanism would increase trajectory fluctuations and planning time consumption. In contrast, the average path length of the proposed model was shortened to 1.83 m, and the path smoothness index was reduced to 0.96, indicating that it can generate shorter and more stable motion trajectories. Meanwhile, the collision rate of the proposed model was only 1.84%, the minimum safe distance was increased to 0.21 m, and the dynamic obstacle avoidance success rate reached 98.27%, indicating that the SGBM three-dimensional environment reconstruction and real-time feedback mechanism can improve the spatial perception accuracy of obstacles. Its average response time was 48.3 m, and the task completion rate reached 98.64%, indicating that this model takes into account real-time performance, security and execution stability, and is more suitable for complex three-dimensional obstacle avoidance tasks of robotic arms.

Aiming at the challenges such as severe obstacle occlusion, unstable target boundary positioning, insufficient three-dimensional spatial distance perception and lagging dynamic obstacle avoidance response of robotic arms in complex industrial environments, a three-dimensional obstacle avoidance method integrating KAN-SDI-BiFPN-YOLOv11 environmental detection and positioning and SGBM binocular stereo matching was constructed. The experimental results showed that the mAP@0.5 of complete detection model reached 98.21%, which was 6.34% higher than 91.87% of YOLOv11. The precision reached 98.37%, an increase of 6.23% compared to YOLOv11. The recall reached 97.42%, an increase of 6.79% compared to YOLOv11. The F1-score reached 97.89%, an increase of 6.51% compared with YOLOv11. The positioning error decreased from 2.84 cm of YOLOv11 to 1.26 cm, with a reduction rate of 55.63%. This indicates that KAN, SDI and BiFPN can effectively improve the boundary positioning and small target recognition capabilities in complex backgrounds. In terms of the comprehensive obstacle avoidance performance, the average path length of the proposed method was 1.83 m, which was 13.68% shorter than 2.12 m of A* and 18.67% shorter than 2.25 m of RRT. The collision rate was only 1.84%, 4.41% lower than that of A* and 1.73% lower than that of RRT. The minimum safety distance was increased to 0.21m, which was significantly higher than 0.06 m for A* and 0.07 m for RRT. The average response time was 48.3 m, which was 44.29% shorter than that of A* and 56.99% shorter than that of RRT. In conclusion, the proposed method can achieve superior performance in detection accuracy, depth recovery, path smoothness, obstacle avoidance safety and real-time response, and can provide effective support for the autonomous obstacle avoidance of robotic arms in complex industrial scenarios. However, the research still mainly focuses on typical industrial assembly and occlusion scenarios. There is insufficient verification of highly reflective, transparent objects, high-speed dynamic obstacles, and multi-robotic arm collaborative environments. Subsequently, the generalization ability and engineering application stability can be further enhanced by combining multi-sensor fusion, lightweight deployment and online trajectory optimization mechanisms. In particular, for embedded robotic platforms with limited computing resources, model compression techniques including pruning, quantization and knowledge distillation can be explored to achieve efficient edge deployment. Unlike automotive binocular vision systems that mainly optimize large-scale environmental perception, future lightweight robotic perception models should maintain high-precision obstacle localization and depth-aware collision avoidance under resource-constrained industrial conditions.