CBAM meets DropBlock: enhancing robot steering-angle prediction with hybrid attention and structured dropout Authors: Jing Niu, Guanghao Gao, Chuanyan Shen, Jiahao Zheng, Shifeng Liu, Yibo Wang, Jiapei Wei Published: 2026 ABSTRACT AbstractThe prediction of steering-angle for robots in complex scenarios is crucial in intelligent auto-navigation process. However, traditional angle prediction approach of end-to-end auto-driving system often fall short in insufficient prediction accuracy, interference of complex scene features in the steering. This paper mainly proposes a lightweight steering angle prediction network based on the cooperative optimization strategy of bi-domain attention fused region mask and de-fuzzy network. Firstly, this method constructs a three-stage progressive framework featured the multi-scale feature pyramid, the channel/spatial attention decoupling, and the dynamic region feature dropout. The dual-channel feature compression is carried out in the channel dimension by using global average pooling and maximum pooling. Then, the channel weight vectors are generated by a multilayer perceptron and the key semantic features, such as lane lines and obstacles, are autonomously strengthened. Finally, in the spatial dimension, the attention heat map is generated by gated convolutional kernel, and the probabilistic feature masking of non-critical regions is combined with the DropBlock module, so as to effectively suppress the interfering information such as road texture and sky background. The experimental results show that the proposed method performs well in the steering angle prediction, suggesting promising applications and broad prospects for future.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-54229-1.Keywords: Robot steering-angle prediction, CBAM, DropBlock, Deblurring network, Model lightweightingSubject terms: Engineering, Mathematics and computing Introduction Recently, with the rapid development of artificial intelligence and visual sensor technology, auto-navigation has been increasingly applied in the field of wheeled robots. Auto-navigation technology is one of the crucial technologies of robots. It determines how robots travel to their estimated positions and involves the collection and processing of environmental information, road detection, obstacle identification, optimal path selection, and auto-driving. End-to-end auto-driving technology is currently the mainstream method for implementing auto-navigation tasks for robots. It enables a direct mapping from perception input to vehicle control, bypassing complex intermediate steps. This direct mapping method greatly simplifies the architecture of autonomous driving systems1 and improves decision-making efficiency and response speed. End-to-end technology more closely mimics the way humans think when driving a vehicle. For familiar routes, people tend to use a fixed way of thinking, i.e., they are in a classic end-to-end driving state2,3. Meanwhile, NVIDIA’s deep convolutional neural network piotNet4 has been one of the most influential works in the field of end-to-end auto-driving, which uses a 9-layer CNN structure with 5 convolutional layers, 3 fully connected layers, and 1 normalization layer. The normalization layer is formed through hard coding, the convolutional layer uses a 5×5 kernel and a 2×2 stride for strided convolution, and the last two convolutional layers use non-strided convolution with a 3×3 kernel. This model could learn the mapping from images to steering angles automatically without manually design feature extraction. The PointNet5–7 uses a direct point cloud processing method, focusing primarily on the extraction of local features. It does not introduce an attention mechanism, and all channel features are treated equally, resulting in the neglect of key features. However, the data is large and complex, and no frame compensation is performed for occluded features leading to low prediction accuracy. However, in actual operation, the auto-driving behavior of robots is continuous. Auto-driving in dynamic real-time environments requires frame interpolation and enhancement capabilities. Therefore, comprehensive consideration of environmental information is crucial for precise motion control. Bai Liyun et al.8 proposed a model based on a deep cascade of convolutional neural networks (CNN) and long short-term memory (LSTM)9 networks to achieve end-to-end prediction from images to steering wheel angles. This method effectively utilizes the time information during the robot’s movement, significantly improving performance, but the model is relatively large and requires a large number of iterations during training10. Nowadays, auto-driving strategies based on deep learning have become a hot topic of research, but model performance is often limited by the scarcity of high-precision annotated data and insufficient temporal modeling capabilities in complex environments. Address to this challenge, Xu et al.11 at the University of California, proposed a hybrid FCN-LSTM architecture, which achieves spatio-temporal feature fusion in unstructured autonomous driving scenarios through the synergistic action of a fully convolutional network and a long short-term memory network. Unlike traditional models that rely on strictly calibrated data, this architecture utilizes large-scale uncalibrated network video data for the first time to perform self-supervised learning, which significantly reduces the data acquisition cost. Its core innovation lies in using semantic segmentation as an auxiliary learning task, enhancing the model’s understanding of road topology through a multi-task optimization strategy, and demonstrating a steering prediction accuracy of 18.6% better than that of a single-task model under complex working conditions such as intersections and night driving12. This result provides a new paradigm for driving behavior prediction in open road environments, breaking through the limitations of past methods in terms of data generalizability and environmental adaptability. Pomerleau proposed the Alvinn model13, which achieved good experimental results in a lane following task and was able to predict the steering angle using only a single frame of image as input. When humans are driving vehicles, they can quickly scan the road ahead and accurately obtain road conditions in key areas. That is, the focus of attention, and then more attentional resources are devoted to this region to acquire more target features to the exclusion of other redundant resources, a process called the biological visual attention mechanism14. With the continuous development of deep learning, the concept of visual attention mechanisms has been introduced into the field of autonomous driving15–17. The Google machine translation team introduced the self-attention mechanism into the Transformer architecture. This mechanism deeply mines the potential semantic associations of text by dynamically correlating the dependencies between different positions within a sequence, and has been successfully applied to large-scale text representation learning tasks18. When extracting RGB image features, CNN treats all information indiscriminately, resulting in a large amount of redundancy and increasing data processing time. Therefore, incorporating attention mechanisms into end-to-end learning autonomous driving models can optimize data processing methods, reduce the number of model parameters, and improve prediction accuracy. The ResNet architecture leverages residual connections to enable stable deep feature propagation, allowing the network to simultaneously capture low-level texture cues and high-level semantic structures, which provides strong representation capacity and generalization ability in complex visual environments. Building upon this backbone, the CBAM attention module is introduced to adaptively recalibrate feature responses along both channel and spatial dimensions, thereby enhancing the network’s sensitivity to motion-relevant regions such as road boundaries and trajectory trends. In addition, DropBlock regularization is incorporated to suppress over-reliance on localized salient responses by randomly masking contiguous spatial regions, which improves robustness and spatial generalization. Previous studies have primarily applied CBAM and DropBlock to static vision tasks such as image classification, object detection, and semantic segmentation, while their roles in continuous control prediction remain underexplored. Their integration with ResNet therefore enables a multi-scale spatial perception framework that jointly captures local obstacles and global path structures, which is particularly beneficial for robot navigation scenarios and introduces novelty in the context of vision-based turning-angle regression. Furthermore, a deblurring module, traditionally employed in image restoration, medical imaging enhancement, and remote sensing reconstruction, is incorporated to mitigate motion blur and sensor noise caused by camera vibration and environmental disturbances during robot operation. By improving the structural consistency and clarity of the visual input, this module provides more reliable perceptual cues for downstream feature extraction and regression-based motion prediction. To address the issues of low prediction accuracy, complex structure, and data redundancy in existing models, this paper proposes a lightweight wheeled robot steering angle prediction model based on a collaborative optimization strategy of cross-domain attention (CBAM) and region masking (DropBlock). By constructing a three-stage progressive framework featured the multi-scale feature pyramid, the channel/spatial attention decoupling, and the dynamic region feature dropout, channel-level feature compression is achieved using a dual-channel approach combining global average pooling and max pooling. This process generates channel weight vectors through a multi-layer perceptron, enabling autonomous reinforcement of key semantic features such as lane lines and obstacles. Additionally, the DropBlock module applies probabilistic feature masking to non-critical regions, effectively suppressing distracting information like road texture and sky background. As a result, the accuracy of robot steering-angle prediction has significantly improved. Materials and methods Task description The end-to-end auto-driving system directly establishes an end-to-end mapping relationship between multimodal perception inputs and robot control commands through deep neural networks. Among these, methods based on imitation learning demonstrate significant advantages by replicating human driving strategies. Existing studies typically use steering angle and speed as joint control targets for regression prediction. However, speed decisions are subject to multiple factors such as driver behavior differences and sensor sampling frequency limitations, which can lead to convergence instability issues during model training19. To address this challenge, this paper focuses on the steering-angle prediction task based on a staged control strategy and utilizes an improved channel feature fusion network to achieve high-precision steering control. Subsequently, based on this, a hybrid control architecture combining data-driven and model-driven approaches can be constructed by dynamically adjusting speed parameters according to the steering angle using a classic PID controller. This phased control strategy reduces the complexity of learning tasks while effectively avoiding the impact of speed label noise on model generalization performance, providing a new technical path for reliable autonomous driving in complex scenarios. The end-to-end robot steering-angle prediction network based on visual attention mechanism proposed in this paper is shown in Fig. 1. The model input is a sequence of images from the forward-facing onboard camera, and the network output is the currently predicted steering wheel angle. The prediction network uses ResNet18 as the backbone network, adds a CBAM module before the global pooling layer, changes the final fully connected layer to output an identity function, and adds a region mask module in the third layer. Attention-driven end-to-end steering angle prediction for wheeled robots using visual perception. A collaborative network integrating dual-domain attention, regional masking, and image deblurring ResNet18 base network ResNet20 solves the problems of gradient disappearance and gradient explosion during the training process of deep networks through the design of residual layers, and is widely used in image classification and computer vision tasks. ResNet generally consists of an initial feature extraction layer, a residual layer, and a fully connected layer. ResNet-18 is a lightweight version of ResNet, and its structure is shown in Fig. 2. ResNet18 network. In the ResNet18 network, the original RGB image is enhanced by horizontally flipping and randomly cropping it to increase the amount of training data and prevent overfitting. The image size is then converted to 224×224 before entering the Input stem layer, which contains a 7×7 convolution layer with a stride of 2 and a convolution kernel of 64. Then, through a 3×3 max pooling layer with a stride of 2, the image size is reduced by half to obtain a feature map of 56×56×64. This is followed by four residual layers, each of which contains two residual blocks as shown in Fig. 3. Each residual block consists of two 3×3 convolution layers and a skip connection. The residual block structure formula is: Basic block. In the formula, XL+1 and YL are the outputs of the L+1 and L residual blocks; h(XL)is the bottom layer mapping of the residual block L ; f is the relu activation function. Short connections directly transmit input information from the input layer to the output layer, preventing overfitting and enabling the network to learn residual maps more easily. After the residual layer, the ResNet-18 network uses a global average pooling layer to reduce the spatial dimension of the feature map. It reduces the height and width dimensions of each feature map to 1. Finally, the feature maps are mapped to specific categories through a fully connected layer. Improved ResNet-CBAM-DB network As shown in Fig. 4, compared with the basic ResNet18 network, the improved ResNet-CBAM-DB network adds a channel attention module CBAM (convolutional block attention module) before the global pooling layer of the ResNet18 network; adds a region mask module (Drop Block) after Layer3 of the network; uses the ReLU activation function in the intermediate layers of the network; and changes the final fully connected layer (fc) to an identity output function. Improved ResNet-CBAM-DB network. CBAM is a lightweight attention mechanism module, as shown in Fig. 5, which aims to enhance the neural network’s focus on important features. It improves the model’s expressiveness and performance by introducing attention mechanisms in both the channel and spatial dimensions21. CBAM consists of two main modules: the channel attention module (CAM), which highlights important channel features by weighting the channel dimension of feature maps; and the spatial attention module (SAM), which emphasizes key spatial regions by weighting the spatial dimension of feature maps. The two modules work in series to gradually enhance the expressive power of the feature map. In terms of feature extraction, CAM enhances the response to key channels for steering angle prediction (lane lines, obstacle edges) in the channel dimension and suppresses pixels unrelated to the channel (weeds, trees, sky). Channel attention module. Use the following formulas (3) and (4) to calculate the channel weights WC , suppress irrelevant feature channels (such as sky and vegetation), and enhance the response of road geometric features. In the formula, is the channel-level description vector of the input feature map. In the formula, MLP includes a dimension reduction layer (C \to C/r) and a dimension increase layer (C/r \to C), with a compression ratio of r = 16; σ is the sigmoid activation function, which maps the weights to the [0,1] interval. DropBlock is a regularization method that prevents overfitting in neural networks22. DropBlock has two main parameters: drop_prob and block_size. Drop_prob is the pixel dropout probability of DropBlock, which determines how many neurons in each region are discarded during each forward propagation. Block_size determines the size of the discarded region in the DropBlock module. Specifically, it is the side length of each discarded “block.” Since ResNet18 has a large number of convolutional layers and its features are spatially related, only a more structured DropBlock can better regularize the convolutional network. Compared with Dropout, DropBlock is structurally designed to remove consecutive regions from a layer of feature maps, effectively separating the relevant information between different features. This more effectively simulates local occlusion in real scenarios, enhances model robustness, and forces the network to learn redundant feature representations through random feature region discarding, avoiding overemphasis on a single local feature and enhancing the generalization ability of ResNet18. Traditional ResNet Networks are currently mostly applicable to classification tasks, where the output categories of the fully connected layers are multiple numerical values. Since the network described in this paper is used for the single-output regression task of predicting the steering angle for autonomous driving, the fully connected layer of the ResNet residual network is set as an identity function. The insertion locations and key hyperparameters of each module are summarized in Table 1. Table 1Insertion locations and key hyperparameters.ModuleDropBlockCBAMFCInsertion positionAfter layer3Before the global pooling layerN/AKey hyperparametersDrop = 0.2 Block = 7Reduction ratio = 16Identity function Insertion locations and key hyperparameters. MPRNet deblurring network Currently, most deblurring networks adopt a single-layer design, with their network architecture developed from advanced visual models such as residual learning image denoising, image deblurring, and image degraining23–25. In addition, encoders-decoders26,27 and expanded convolutional networks are also used to make multidimensional information more accessible. MPRNet is a three-layer network. The first two layers are based on encoders and decoders, which learn extensive information through large receptive fields. The last layer uses subnetworks to manipulate the resolution of the original input image. Insert an attention module between every two layers to pass the features from the previous layer to the input part of the next layer through supervised images, and scale the input. At the same time, a cross-stage feature fusion method was introduced, which utilizes multi-scale information from the previous layer network for consolidation, serving as intermediate features for the subsequent layer network. The input image is adjusted to a multi-layer patch hierarchy and divided into non-overlapping patches. Four patches are used for the first stage, two patches are used for the second stage, and the original image is used for the final stage. At any stage, the model does not directly predict the re-stored images Xs, but rather predicts the residual images Rs. The degraded input image I is added to the residual image to obtain: . End-to-end optimization of MPRNet using the loss function shown in Eq. (5): In the formula, Y represents the real image, and Lchar is the Charbonnier loss. In the formula, ε is set to 10–3, and Ledge is the edge loss, which is calculated by the following formula (7): In the equation, △ is the Laplace operator. Image restoration based on the U-Net network28 using an encoder-decoder subnetwork. First, a channel attention module is added to extract features at different scales. Second, features at different scales are also extracted for the skipped connections in U-Net. Finally, bilinear upscaling is used to reduce ghosting caused by transposed convolutions. To ensure that the image data is fine-grained at the output, an original resolution module (ORSNet) is introduced in the final stage. ORSNet does not perform downsampling operations and generates high-resolution spatial features. Insert a CSFF module between the encoder and decoder of the deblurring network, and also add a CSFF module between the encoder-decoder and ORSNe. The advantages of this approach are as follows: first, the CSFF module prevents information loss caused by the repeated use of upsampling and downsampling operations in the encoder-decoder, even as the network structure becomes more complex, without compromising inference speed; second, the capture of features at different scales in the previous stage enhances feature refinement in the subsequent stage; finally, unlike multi-stage deblurring networks [29-1.3130], MPRNet introduces a supervised attention module (SAM) between every two stages, significantly improving the network’s operational speed and stability. Data set and evaluation indicators Data set structure The experiment was conducted on the school track using a lightweight two-wheel drive robot [1.32] developed independently by our research team. The dataset obtained contains 10,000 image frames under different weather and operating conditions. The image resolution is 1920 × 1080 pixels, and the acquisition frequency is 10 Hz, as shown in Fig. 6. During testing, the dataset was divided into a training set of 10,000 images and a validation set of 2,000 images. Examples of images from the training set. Data set enhancement Firstly, Images are resized from 1920×1080 to a fixed dimension of 224×224. This standardization serves multiple purposes, it meets the fixed input requirement of the convolutional backbone network, substantially reduces the computational and memory overhead, and follows established practice in vision-based driving models. The reduction in resolution and alteration of aspect ratio also help suppress high-frequency background details, thereby directing the model’s attention towards salient, task-relevant road features. The steering angle labels, acquired from the vehicle controller as raw servo commands in the range [-1800, 1800], are normalized to [-π, π] radians. This normalization centers and scales the data, which is a standard practice to improve the convergence and numerical stability of gradient-based optimization.The dataset is partitioned using a sequential split to preserve the temporal order. For each driving sequence, the first 80% of the frames are used for training, and the remaining consecutive 20% are reserved for validation/testing. This ensures that the model is evaluated on future, unseen scenarios and prevents the leakage of near-duplicate frames across the splits, thereby providing a rigorous assessment of generalization, as shown in Table 2. Finally, to prevent model overfitting and improve model generalization, the data set is enhanced using methods such as random cropping, horizontal flipping, sharpness adjustment, contrast adjustment, and other methods to enhance the data set, as shown in Fig. 7. Table 2.Data enhancement.Processing step​Before processing​​​After processing​Image size1920 × 1080224 × 224Steering angle range[-1800,1800][-π, π]Image augmentation count10,00012,000​Pixel value range​0–2550–1 Data enhancement. Augmented samples. Evaluation indicators To objectively evaluate the performance of the prediction model, three complementary metrics are used for joint evaluation, namely RMSE (root mean square error), MAE(mean absolute error), and MSE (mean squared error). RMSE is used to quantify the standard deviation of errors, reflecting the degree of dispersion of data distribution. It balances global errors and local details. MAE is weighted by the absolute value of errors, is insensitive to outliers, and is more robust, making it less susceptible to distortion of the optimization direction by extreme values. It balances the restoration of local features. MSE is sensitive to outliers and reflects the severity of large deviations. Quantify the error between the model output and actual values using Eqs. (8) to (10). It is worth noting that smaller values of RMSE, MAE, and MSE indicate smaller model bias and lower rates of change in steering angle predictions. These three factors together constitute multiple requirements for model optimization. In the equation, Ytrue and Ypred represent the true steering angle and the model-predicted steering angle corresponding to the image of the i-th frame in the total dataset of N frames, respectively. Experiments and analysis Initialization of experimental parameters The experimental environment used in this paper is as follows: Windows 10 operating system, 13th Gen Intel (R) Core (TM) i5-13600F 3.50GHz processor, NVIDIA GeForce RTX4060Ti GPU, 16 GB memory, PyTorch 1.12.1 deep learning framework, Python programming language, and PyCharm code execution environment. Since the setting of hyperparameters affects the performance and training results of the model, StepLR is used as a stepwise learning rate scheduler33 in this paper. By periodically reducing the learning rate, the model parameters gradually converge, avoiding local optima. During the experiment, multiple sets of different step sizes and gamma values were compared. The experimental results are shown in Fig. 8. Based on the advantages and disadvantages of the three complementary indicators, the step size of the test model was finally determined to be 10, and the decay factor (gamma) was determined to be 0.9. The steering angle prediction curves under various step sizes. The code updates the learning rate at the end of each epoch, with a batch size of 128, 100 training iterations, and the Adam34 optimizer. The DropBlock regularization is applied after the third residual stage (layer3). Its two key hyperparameters block size​ and drop probability were determined through an empirical search on the validation set. We evaluated block sizes of {3, 5, 7} and drop probabilities of {0.1, 0.2, 0.3} in a grid-search manner. The combination of a block size of 7 and a drop probability of 0.2 was found to offer the optimal trade-off, consistently yielding the lowest validation loss while maintaining stable training dynamics. This configuration effectively encourages the model to learn robust, spatially distributed features without causing excessive information loss. All hyperparameter tuning was conducted exclusively on the validation set to ensure a completely unbiased evaluation on the held-out test set, as shown in Fig. 9. Error curves of different Dropblock configurations. Model performance comparison experiment To further verify the effectiveness of the proposed network, comparative experiments were conducted against several representative deep learning models, including AlexNet, MobileNetV3 and DenseNet-121. A five-fold cross-validation strategy was adopted during the evaluation stage, and the mean values and standard deviations of the evaluation metrics across the five runs were reported. The detailed results are presented in Table 3. Table 3.Experimental comparison results.Model​​​Test Set​RMSEMAEMSEHuman driving​N/AN/AN/AAlexNet0.3892 ± 0.030.2721 ± 0.140.0821 ± 0.052MobileNetV30.2438 ± 0.050.2123 ± 0.060.1335 ± 0.011DenseNet-1210.2312 ± 0.060.1742 ± 0.030.06525 ± 0.042ResNet-CBAM-DB0.2113 ± 0.020.1121 ± 0.050.0526 ± 0.013 Experimental comparison results. MobileNetV3 network MobileNetV3 is a lightweight convolutional neural network architecture specifically optimized for mobile and embedded platforms. Building upon the inverted residual structure of MobileNetV2, the model integrates neural architecture search with manual architectural refinements to achieve an improved balance between computational efficiency and representational capacity. In addition, it incorporates a lightweight Squeeze-and-Excitation attention mechanism (as shown in Fig. 10) and the computationally efficient h-swish activation function, thereby enhancing feature modeling capability while maintaining low computational cost and parameter complexity35. MobileNetV3 network. DenseNet-121 network DenseNet-121 is a representative architecture within the Dense Convolutional Network family, comprising 121 learnable layers, as shown in Fig. 11. Its core principle lies in the dense connectivity mechanism, whereby the feature maps of each layer are concatenated with those of all subsequent layers, establishing direct cross-layer information pathways. This design substantially enhances feature reuse and gradient propagation efficiency, effectively alleviating the vanishing gradient problem in deep networks while improving the continuity and stability of feature representation. Meanwhile, DenseNet-121 regulates the number of newly generated feature channels through a relatively small growth rate, thereby maintaining strong representational capacity while controlling model parameter size and computational complexity36. DenseNet-121 network. AlexNet network AlexNet shares a similar basic convolutional network structure, as shown in Fig. 12. The first layer of the convolutional layer uses a large 11×11 convolutional kernel with a stride of 4 to extract low-level features. Subsequent convolutional kernels become smaller, progressively capturing more complex features. The key features of the AlexNet network include: the first use of the ReLU activation function, the incorporation of dropout regularization to randomly mask 50% of neurons in the fully connected layer to prevent overfitting, and local response normalization, normalizing the outputs of adjacent neurons to enhance local contrast; and data augmentation, which involves randomly cropping, horizontally flipping, and color dithering images to improve the model’s generalization ability37,38. AlexNet network. Experimental results The three classic networks mentioned above and the improved network described in this paper were tested on the test set, and the test results are shown in Table 3. The results show that the improved ResNet-CBAM-DB network has significantly improved stability and accuracy on the test set. It can be seen that when CBAM and DropBlock are integrated into ResNet18, the model’s attention allocation becomes more focused, redundant information is filtered out, feature extraction capabilities are enhanced, and the ability to prevent overfitting is greatly improved. Compared to other convolutional neural network models, the ResNet-CBAM-DB network utilizes channel attention and region dropout modules to improve the model’s prediction accuracy and robustness. Network visualization To gain a deeper understanding of how decisions are made within end-to-end autonomous driving models, Grad-CAM technology39 was used to visualize the network’s operational process. Grad-CAM is a visualization method based on gradient information, designed to explain the decision-making basis of convolutional neural networks. Traditional class activation mapping (CAM) generates visualization results using the fully connected weights after global average pooling layers, but this method is only applicable to specific network architectures (such as fully convolutional tail structures). The introduction of Grad-CAM broke through this limitation by applying backpropagation gradient information to weight feature maps, making it applicable to a variety of complex architectures, including residual networks (ResNet) and visual transformers (ViT), and significantly improving the generality of model interpretation. By combining guided backpropagation, high-resolution, fine-grained visualization results (guided grad-CAM) can be generated. Using the gradient information of the last layer of convolutional feature maps for the target category, the importance weights of the feature maps are calculated, and finally, a heat map is generated through weighted superposition. Visualization of the AlexNet, MobileNetV3, DenseNet-121, and ResNet-CBAM-DB networks using the Grad-CAM algorithm, with heatmap displays as shown in Fig. 13, revealed the following: (1) Regardless of whether it was the DenseNet-121, AlexNet, MobileNetV3, or ResNet-CBAM-DB network, there was a clear focus on obstacles ahead and lane edges. This demonstrates that the neural network has mastered the technique of autonomously extracting useful road features when trained solely on images and steering angles; (2) Comparing the DenseNet-121, AlexNet, MobileNetV3, and ResNet-CBAM-DB networks, the ResNet18 network with channel attention was found to be more effective at quickly capturing edges. they have mastered the technique of autonomously extracting useful road features7. Visualiztion masks for models. Real-world deployment experiment To further validate the feasibility of the ResNet-CBAM-DB network and more accurately analyze the network’s accuracy and rapid response in end-to-end steering angle prediction on wheeled robots, the model trained by the network was deployed to a wheeled robot ROS system with stereo depth camera functionality to conduct end-to-end steering angle prediction experiments in a real-world environment. The experimental process is as follows: activate the motion library and sensor acquisition library functions of the wheeled robot, define the initial speed of the wheeled robot as 0.2 m/s and the initial steering angle as 0°, and set the model path and data processing environment. During the wheeled robot’s autonomous driving process, video streams are obtained via a stereo camera for frame preprocessing. The steering angle is output using an end-to-end approach, and the motor executes the corresponding steering angle. Based on the analysis of the experimental paths generated by different networks, the average steering angle deviations for each network are shown in Table 4. Table 4Mean steering angle deviation of different models.ModelAlexNetMobileNetV3DenseNet-121ResNet-CBAM-DBMSAD/(°)1.231.120.520.31 Mean steering angle deviation of different models. As shown in the table, the paths inferred by the AlexNet and MobileNetV3 networks are similar, both deviating from the training set paths, with an average steering angle deviation greater than 1°. The paths inferred by the DenseNet-121 are superior to those of MobileNetV3 and AlexNet, with an average steering angle deviation of 0.52°. The steering angles inferred by the ResNet-CBAM-DB model align with the training set paths and outperform other models, with an average steering angle deviation of only 0.31°. This validates the effectiveness of the model. The model not only demonstrates improvements in inference speed and stability but also achieves angle prediction accuracy within the acceptable deviation range for practical applications, enabling accurate prediction of the steering angle for wheeled robots. This result indicates that the model possesses excellent adaptability and robustness in complex scenarios, providing robust technical support for steering angle prediction tasks. Ablation test To validate that channel attention and random region loss can improve the steering angle prediction accuracy of wheeled robots in dynamic environments, this section focuses on ablation tests for the model’s channel attention feature extraction network and region loss: (1) Ablation tests for the MPRNet deblurring network to discuss whether deblurring improves accuracy and fast response in feature extraction; (2) Ablation tests for channel attention to verify the performance of the model’s feature extraction with the introduction of attention; (3) Ablation tests for region loss to determine whether the model’s overfitting ability is improved. In Table 5, it was presented that the experimental comparison results of ResNet18 under the conditions of adding a deblurring network, with CBAM, and adding regional loss respectively to the ResNet-CBAM-DB network. The results indicate that incorporating a deblurring network significantly improves prediction accuracy compared to models without a deblurring network, demonstrating that models with deblurring networks are powerful visual representations that are easily extracted by ResNet networks. This proves that the ResNet-CBAM-DB network is an effective method for utilizing information enhancement in end-to-end autonomous driving. Table 5Ablation study results.CBAMDrop blockMPRNetMAEMSERMSE × × × 0.24020.08750.3683 × × √0.23820.07350.3433√ × √0.23420.06720.3022√√√0.14410.05380.2319 Ablation study results. Meanwhile, it could be found that incorporating the channel attention module significantly improves prediction accuracy compared to the model without CBAM, demonstrating that the CBAM module significantly enhances feature extraction in both the channel dimension and spatial dimension. This proves that the channel attention module is an effective method for enhancing prediction accuracy by utilizing multi-dimensional information. The bolded parts are the models obtained by ResNet-CBAM-DB. The results show that the overfitting risk is reduced in the model with regional loss added, thereby improving the generalization ability of complex models. Based on the above ablation test, it was found that information from the channel dimension and spatial dimension can achieve precise prediction of the steering angle of a wheeled robot. However, it is worth noting that for the model proposed in this paper, more layers do not necessarily mean better performance. As the number of layers increases, the stability of the steering signal is further improved, but the prediction accuracy decreases. The experiments in this paper show that channel attention significantly improves feature extraction, region loss reduces the risk of model overfitting, and deblurring enhances prediction accuracy. In addition, Table 6 shows a comparison of the number of parameters and computing power between the ResNet-CBAM-DB model and the other three neural network models. The ResNet-CBAM-DB model has the smallest number of parameters and the fastest computing speed. Table 6.Performance comparison between the improved netword models and other models.​​Model​​Size (MB)Training time (h)​AlexNet2383.021MobileNetV321.42.543DenseNet-121302.421ResNet-CBAM-DB11.72.053Significant values are in bold Performance comparison between the improved netword models and other models. Significant values are in bold Conclusions To address the issues of low prediction accuracy, redundant model parameters, and inability to accurately extract effective features in complex scenarios within existing end-to-end auto-driving frameworks, this paper proposes a lightweight collaborative prediction network that integrates image feature deblurring with dual-domain attention and region masks. The network reduces the number of parameters by 34.6% compared to ResNet18 and by 29.7% compared to traditional CNN methods, and 14.2% fewer than a single attention model. The synergistic effect of dual-domain attention and region masks reduces feature map redundancy by 58.3%, improves inference speed by 41%, and reduces robot lateral control error by 62% compared to the baseline model. Specifically, the prediction network achieves significant performance improvements in the following aspects: The MPRNet deblurring network has been integrated into the three-tier progressive network framework of “multi-scale feature pyramid-channel/spatial attention decoupling-dynamic region feature discard,” enabling the network to quickly capture effective features even when image pixel quality is low. The ResNet18 model with channel attention modules can significantly improve the accuracy of steering angle prediction for wheeled robots. However, the model with only the channel attention module is insufficient for analyzing data inference in complex environments. Therefore, a region loss module is added to prevent the network from extracting features that do not match the labels in complex environments. Experimental results show that the proposed model outperforms existing models in end-to-end autonomous driving steer angle prediction. ResNet-CBAM-DB not only has fewer parameters and lower computational complexity but also achieves faster computation speed and higher accuracy. The MPRNet deblurring network has been integrated into the three-tier progressive network framework of “multi-scale feature pyramid-channel/spatial attention decoupling-dynamic region feature discard,” enabling the network to quickly capture effective features even when image pixel quality is low. The ResNet18 model with channel attention modules can significantly improve the accuracy of steering angle prediction for wheeled robots. However, the model with only the channel attention module is insufficient for analyzing data inference in complex environments. Therefore, a region loss module is added to prevent the network from extracting features that do not match the labels in complex environments. Experimental results show that the proposed model outperforms existing models in end-to-end autonomous driving steer angle prediction. ResNet-CBAM-DB not only has fewer parameters and lower computational complexity but also achieves faster computation speed and higher accuracy. Grad-CAM technology was used to visualize the network operation process. Different models were compared with ResNet-CBAM-DB through experimental tests. Based on the generated heat maps, the differences in feature recognition among different models were inferred. Additionally, group ablation tests were conducted. The results showed that the deblurring network enhanced the model’s robustness, the dual-domain attention module enabled feature reinforcement and removal of redundant features, and the region mask collaboration module prevented overfitting, thereby increasing model stability. Grad-CAM technology was used to visualize the network operation process. Different models were compared with ResNet-CBAM-DB through experimental tests. Based on the generated heat maps, the differences in feature recognition among different models were inferred. Additionally, group ablation tests were conducted. The results showed that the deblurring network enhanced the model’s robustness, the dual-domain attention module enabled feature reinforcement and removal of redundant features, and the region mask collaboration module prevented overfitting, thereby increasing model stability. The method described in this paper only considers robots operating in static scenes, which are relatively simple and do not involve complex situations such as dynamic obstacles. Future research will focus on complex dynamic scenes with many obstacles, further optimizing the prediction network in terms of improving model prediction accuracy and edge deployment capabilities, and laying the foundation for improving the robot’s global and local path planning in complex scenes. The method described in this paper only considers robots operating in static scenes, which are relatively simple and do not involve complex situations such as dynamic obstacles. Future research will focus on complex dynamic scenes with many obstacles, further optimizing the prediction network in terms of improving model prediction accuracy and edge deployment capabilities, and laying the foundation for improving the robot’s global and local path planning in complex scenes. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Information 1. Supplementary Information 2 Supplementary Information 3. Author contributions J.N. contributed to the investigation, formal analysis, experiment, validation, funding acquisition, and review. G.G. contributed to the methodology, formal analysis, experiment, validation, writing, and review. C.S. and J.Z. contributed to the conceptualization, methodology, funding acquisition, and review. J.W.; Y.W. and S. L. contributed to the experiment, methodology, experiment, and validation. All authors have read and agreed to the published version of the manuscript. Funding This work was supported by the Key R&D Program of Gansu Province No. 23YFFE0001, Gansu Province Higher Education Industry Support Program Project No. 2025CYZC-054, Tianshui Normal University Industry Support and Guidance Project No. CYZ2023-05, Tianshui Normal University Innovation and Entrepreneurship Project No. CXCYJG-JGXM202304JD, Tianshui Normal University Graduate Innovation Project No. 0309–202501030184. Data availability All data generated or analysed during this study are included in this published article (and its Supplementary Information files). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Jing Niu and Guanghao Gao: These authors contributed equally to this work. References Associated Data Supplementary Materials Supplementary Information 1. Supplementary Information 2 Supplementary Information 3. Data Availability Statement All data generated or analysed during this study are included in this published article (and its Supplementary Information files).