Frontiers in Neurorobotics最新文献

Introduction: To address the challenges of cumulative errors, insufficient modeling of complex spatiotemporal features, and limitations in computational efficiency and generalization ability in 4D trajectory prediction, this paper proposes a high-precision, robust prediction method.

Methods: A hybrid model SVMD-DBO-RCBAM is constructed, integrating sequential variational modal decomposition (SVMD), the dung beetle optimization algorithm (DBO), and the ResNet-CBAM network. Innovations include frequency-domain feature decoupling, dynamic parameter optimization, and enhanced spatio-temporal feature focusing.

Results: Experiments show that the model achieves a low longitude MAE of 0.0377 in single-step prediction, a 38.5% reduction compared to the baseline model; in multi-step prediction, the longitude R2 reaches 0.9844, with a 72.9% reduction in cumulative error rate and an IQR of prediction errors less than 10% of traditional models, demonstrating high accuracy and stability.

Discussion: Experiments show that the model achieves a low longitude MAE of 0.0377 in single-step prediction, a 38.5% reduction compared to the baseline model; in multi-step prediction, the longitude R2 reaches 0.9844, with a 72.9% reduction in cumulative error rate and an IQR of prediction errors less than 10% of traditional models, demonstrating high accuracy and stability.

With the advancement of deep learning, road crack segmentation has become increasingly crucial for intelligent transportation safety. Despite notable progress, existing methods still face challenges in capturing fine-grained textures in small crack regions, handling blurred edges and significant width variations, and performing multi-class segmentation. Moreover, the high computational cost of training such models hinders their practical deployment. To tackle these limitations, we propose RSA-TransUNet, a novel model for road crack segmentation. At its core is the Axial-shift MLP Attention (ASMA) mechanism, which integrates axial perception with sparse contextual modeling. Through multi-path axial perturbations and an attention-guided structure, ASMA effectively captures long-range dependencies within row-column patterns, enabling detailed modeling of multi-scale crack features. To improve the model's adaptability to structural irregularities, we introduce the Adaptive Spline Linear Unit (ASLU), which enhances the model's capacity to represent nonlinear transformations. ASLU improves responsiveness to microstructural variations, morphological distortions, and local discontinuities, thereby boosting robustness across different domains. We further develop a Structure-aware Multi-stage Evolutionary Optimization (SMEO) strategy, which guides the training process through three phases: structural perception exploration, feature stability enhancement, and global perturbation. This strategy combines breadth sampling, convergence compression, and local escape mechanisms to improve convergence speed, global search efficiency, and generalization performance. Extensive evaluations on the Crack500, CFD, and DeepCrack datasets-including ablation studies and comparative experiments-demonstrate that RSA-TransUNet achieves superior segmentation accuracy and robustness in complex road environments, highlighting its potential for real-world applications.

Although deep learning methods have made significant strides in single image sand dust removal, the heterogeneous uncertainty induced by dusty environments poses a considerable challenge. In response, our research presents a novel framework known as the Hierarchical Interactive Uncertainty-aware Network (HIUNet). HIUNet leverages Bayesian neural networks for the extraction of robust shallow features, bolstered by pre-trained encoders for feature extraction and the agility of lightweight decoders for preliminary image reconstitution. Subsequently, a feature frequency selection mechanism is activated to enhance overall performance by strategically identifying and retaining valuable features while effectively suppressing redundant and irrelevant ones. Following this, a feature enhancement module is applied to the preliminary restoration. This intricate fusion culminates in the production of a restored image of superior quality. Our extensive experiments, using our proposed Sand11K dataset that exhibits various levels of degradation from dust and sand, confirm the effectiveness and soundness of our proposed method. By modeling uncertainty via Bayesian neural networks to extract robust shallow features and selecting valuable features through frequency selection, HIUNet can reconstruct high-quality clean images. For future work, we plan to extend our uncertainty-aware framework to handle extreme sand scenarios.

Introduction: Understanding group actions in real-world settings is essential for the advancement of applications in surveillance, robotics, and autonomous systems. Group activity recognition, particularly in sports scenarios, presents unique challenges due to dynamic interactions, occlusions, and varying viewpoints. To address these challenges, we develop a deep learning system that recognizes multi-person behaviors by integrating appearance-based features (HOG, LBP, SIFT), skeletal data (MediaPipe, MOCON), and motion features. Our approach employs a Dynamic Graph Neural Network (DGNN) and Bi-LSTM architecture, enabling robust recognition of group activities in diverse and dynamic environments. To further validate our framework's adaptability, we include evaluations on Volleyball and SoccerTrack UAV-recorded datasets, which offer unique perspectives and challenges.

Method: Our framework integrates YOLOv11 for object detection and SORT for tracking to extract multi-modal features-including HOG, LBP, SIFT, skeletal data (MediaPipe), and motion context (MOCON). These features are optimized using genetic algorithms and fused within a Dynamic Graph Neural Network (DGNN), which models players as nodes in a spatio-temporal graph, effectively capturing both spatial formations and temporal dynamics.

Results: We evaluated our framework on three datasets: a volleyball dataset, SoccerTrack UAV-based soccer dataset, and NBA basketball dataset. Our system achieved 94.5% accuracy on the volleyball dataset (mAP: 94.2%, MPCA: 93.8%) with an inference time of 0.18 s per frame. On the SoccerTrack UAV dataset, accuracy was 91.8% (mAP: 91.5%, MPCA: 90.5%) with 0.20 s inference, and on the NBA basketball dataset, it was 91.1% (mAP: 90.8%, MPCA: 89.8%) with the same 0.20 s per frame. These results highlight our framework's high performance and efficient computational efficiency across various sports and perspectives.

Discussion: Our approach demonstrates robust performance in recognizing multi-person actions across diverse conditions, highlighting its adaptability to both conventional and UAV-based video sources.

Robotic racket sports provide exceptional benchmarks for evaluating dynamic motion control capabilities in robots. Due to the highly non-linear dynamics of the shuttlecock, the stringent demands on robots' dynamic responses, and the convergence difficulties caused by sparse rewards in reinforcement learning, badminton strikes remain a formidable challenge for robot systems. To address these issues, this study proposes DTG-IRRL, a novel learning framework for badminton strikes that integrates imitation-relaxation reinforcement learning with dynamic trajectory generation. The framework demonstrates significantly improved training efficiency and performance, achieving faster convergence and twice the landing accuracy. Analysis of the reward function within a specific parameter space hyperplane intuitively reveals the convergence difficulties arising from the inherent sparsity of rewards in racket sports and demonstrates the framework's effectiveness in mitigating local and slow convergence. Implemented on hardware with zero-shot transfer, the framework achieves a 90% hitting rate and a 70% landing accuracy, enabling sustained humanrobot rallies. Cross-platform validation using the UR5 robot demonstrates the framework's generalizability while highlighting the requirement for high dynamic performance of robotic arms in racket sports.

Introduction: Wrist function impairment is common after stroke and heavily impacts the execution of daily tasks. Robotic therapy, and more specifically wearable exoskeletons, have the potential to boost training dose in context-relevant scenarios, promote voluntary effort through motor intent detection, and mitigate the effect of gravity. Portable exoskeletons are often non-backdrivable and it is challenging to make their control safe, reactive and stable. Admittance control is often used in this case, however, this type of control can become unstable when the supported biological joint stiffens. Variable admittance control adapts its parameters dynamically to allow free motion and stabilize the human-robot interaction.

Methods: In this study, we implemented a variable admittance control scheme on a one degree of freedom wearable wrist exoskeleton. The damping parameter of the admittance scheme is adjusted in real-time to cope with instabilities and varying wrist stiffness. In addition to the admittance control scheme, sEMG- and gravity-based controllers were implemented, characterized and optimized on ten healthy participants and tested on six stroke survivors.

Results: The results show that (1) the variable admittance control scheme could stabilize the interaction but at the cost of a decrease in transparency, and (2) when coupled with the variable admittance controller the sEMG-based control enhanced wrist functionality of stroke survivors in the most extreme angular positions.

Discussion: Our variable admittance control scheme with sEMG- and gravity-based support was most beneficial for patients with higher levels of impairment by improving range of motion and promoting voluntary effort. Future work could combine both controllers to customize and fine tune the stability of the support to a wider range of impairment levels and types.

Introduction: To address the challenges of current 4D trajectory prediction-specifically, limited multi-factor feature extraction and excessive computational cost-this study develops a lightweight prediction framework tailored for real-time air-traffic management.

Methods: We propose a hybrid RCBAM-TCN-LSTM architecture enhanced with a teacher-student knowledge distillation mechanism. The Residual Convolutional Block Attention Module (RCBAM) serves as the teacher network to extract high-dimensional spatial features via residual structures and channel-spatial attention. The student network adopts a Temporal Convolutional Network-LSTM (TCN-LSTM) design, integrating dilated causal convolutions and two LSTM layers for efficient temporal modeling. Historical ADS-B trajectory data from Zhuhai Jinwan Airport are preprocessed using cubic spline interpolation and a uniform-step sliding window to ensure data alignment and temporal consistency. In the distillation process, soft labels from the teacher and hard labels from actual observations jointly guide student training.

Results: In multi-step prediction experiments, the distilled RCBAM-TCN-LSTM model achieved average reductions of 40%-60% in MAE, RMSE, and MAPE compared with the original RCBAM and TCN-LSTM models, while improving R ^² by 4%-6%. The approach maintained high accuracy across different prediction horizons while reducing computational complexity.

Discussion: The proposed method effectively balances high-precision modeling of spatiotemporal dependencies with lightweight deployment requirements, enabling real-time air-traffic monitoring and early warning on standard CPUs and embedded devices. This framework offers a scalable solution for enhancing the operational safety and efficiency of modern air-traffic control systems.

Brain-computer interface (BCI) integration with virtual reality (VR) has progressed from single-limb control to multi-limb coordination, yet achieving intuitive tri-manual operation remains challenging. This study presents a consumer-grade hybrid BCI-VR framework enabling simultaneous control of two biological hands and a virtual third limb through integration of Tobii eye-tracking, NeuroSky single-channel EEG, and non-haptic controllers. The system employs e-Sense attention thresholds (>80% for 300 ms) to trigger virtual hand activation combined with gaze-driven targeting within 45° visual cones. A soft maximum weighted arbitration algorithm resolves spatiotemporal conflicts between manual and virtual inputs with 92.4% success rate. Experimental validation with eight participants across 160 trials demonstrated 87.5% virtual hand success rate and 41% spatial error reduction (σ = 0.23 mm vs. 0.39 mm) compared to traditional dual-hand control. The framework achieved 320 ms activation latency and 22% NASA-TLX workload reduction through adaptive cognitive load management. Time-frequency analysis revealed characteristic beta-band (15-20 Hz) energy modulations during successful virtual limb control, providing neurophysiological evidence for attention-mediated supernumerary limb embodiment. These findings demonstrate that sophisticated algorithmic approaches can compensate for consumer-grade hardware limitations, enabling laboratory-grade precision in accessible tri-manual VR applications for rehabilitation, training, and assistive technologies.

Fine-grained image classification tasks face challenges such as difficulty in labeling, scarcity of samples, and small category differences. To address this problem, this study proposes a novel fine-grained image classification method based on the MogaNet network and a multi-level gating mechanism. A feature extraction network based on MogaNet is constructed, and multi-scale feature fusion is combined to fully mine image information. The contextual information extractor is designed to align and filter more discriminative local features using the semantic context of the network, thereby strengthening the network's ability to capture detailed features. Meanwhile, a multi-level gating mechanism is introduced to obtain the saliency features of images. A feature elimination strategy is proposed to suppress the interference of fuzzy class features and background noise. A loss function is designed to constrain the elimination of fuzzy class features and classification prediction. Experimental results demonstrate that the new method can be applied to 5-shot tasks across four public datasets: Mini-ImageNet, CUB-200-2011, Stanford Dogs, and Stanford Cars. The accuracy rates reach 79.33, 87.58, 79.34, and 83.82%, respectively, which shows better performance than other state-of-the-art image classification methods.

Introduction: Accurate vehicle analysis from aerial imagery has become increasingly vital for emerging technologies and public service applications such as intelligent traffic management, urban planning, autonomous navigation, and military surveillance. However, analyzing UAV-captured video poses several inherent challenges, such as the small size of target vehicles, occlusions, cluttered urban backgrounds, motion blur, and fluctuating lighting conditions which hinder the accuracy and consistency of conventional perception systems. To address these complexities, our research proposes a fully end-to-end deep learning-driven perception pipeline specifically optimized for UAV-based traffic monitoring. The proposed framwork integrates multiple advanced modules: RetinexNet for preprocessing, segmentation using HRNet to preserve high-resolution semantic information, and vehicle detection using the YOLOv11 framework. Deep SORT is employed for efficient vehicle tracking, while CSRNet facilitates high-density vehicle counting. LSTM networks are integrated to predict vehicle trajectories based on temporal patterns, and a combination of DenseNet and SuperPoint is utilized for robust feature extraction. Finally, classification is performed using Vision Transformers (ViTs), leveraging attention mechanisms to ensure accurate recognition across diverse categories. The modular yet unified architecture is designed to handle spatiotemporal dynamics, making it suitable for real-time deployment in diverse UAV platforms.

Method: The framework suggests using today's best neural networks that are made to solve different problems in aerial vehicle analysis. RetinexNet is used in preprocessing to make the lighting of each input frame consistent. Using HRNet for semantic segmentation allows for accurate splitting between vehicles and their surroundings. YOLOv11 provides high precision and quick vehicle detection and Deep SORT allows reliable tracking without losing track of individual cars. CSRNet are used for vehicle counting that is unaffected by obstacles or traffic jams. LSTM models capture how a car moves in time to forecast future positions. Combining DenseNet and SuperPoint embeddings that were improved with an AutoEncoder is done during feature extraction. In the end, using an attention function, Vision Transformer-based models classify vehicles seen from above. Every part of the system is developed and included to give the improved performance when the UAV is being used in real life.

Results: Our proposed framework significantly improves the accuracy, reliability, and efficiency of vehicle analysis from UAV imagery. Our pipeline was rigorously evaluated on two famous datasets, AU-AIR and Roundabout. On the AU-AIR dataset, the system achieved a detection accuracy of 97.8%, a tracking accuracy of 96.5%, and a classification accuracy of 98.4%. Similarly, on the Roundabout dataset, it reached 96.9% det