Frontiers in Neurorobotics最新文献

Under the influence of Masked Language Modeling (MLM), Masked Image Modeling (MIM) employs an attention mechanism to perform masked training on images. However, processing a single image requires numerous iterations and substantial computational resources to reconstruct the masked regions, resulting in high computational complexity and significant time costs. To address this issue, we propose an Effective and Efficient self-supervised Masked model based on Mixed feature training (EESMM). First, we stack two images for encoding and input the fused features into the network, which not only reduces computational complexity but also enables the learning of more features. Second, during decoding, we obtain the decoding features corresponding to the original images based on the decoding features of the two input original images and the mixed images, and then construct a corresponding loss function to enhance feature representation. EESMM significantly reduces pre-training time without sacrificing accuracy, achieving 83% accuracy on ImageNet in just 363 h using four V100 GPUs-only one-tenth of the training time required by SimMIM. This validates that the method can substantially accelerate the pre-training process without noticeable performance degradation.

During turning maneuvers in the galloping gait of quadruped animals, a strong relationship exists between the turning direction and the sequence in which the forelimbs make ground contact: the outer forelimb acts as the "trailing limb" while the inner forelimb serves as the "leading limb." However, the control mechanisms underlying this behavior remain largely unclear. Understanding these mechanisms could deepen biological knowledge and assist in developing more agile robots. To address this issue, we hypothesized that decentralized interlimb coordination mechanism and trunk movement are essential for the emergence of an inside leading limb in a galloping turn. To test the hypothesis, we developed a quasi-quadruped robot with simplified wheeled hind limbs and variable trunk roll and yaw angles. For forelimb coordination, we implemented a simple decentralized control based on local load-dependent sensory feedback, utilizing trunk roll inclination and yaw bending as turning methods. Our experimental results confirmed that in addition to the decentralized control from previous studies which reproduces animal locomotion in a straight line, adjusting the trunk roll angle spontaneously generates a ground contact sequence similar to gallop turning in quadruped animals. Furthermore, roll inclination showed a greater influence than yaw bending on differentiating the leading and trailing limbs. This study suggests that physical interactions serve as a universal mechanism of locomotor control in both forward and turning movements of quadrupedal animals.

Accurate pedestrian tracking and behavior recognition are essential for intelligent surveillance, smart transportation, and human-computer interaction systems. This paper introduces TSLNet, a Hierarchical Multi-Head Attention-Enabled Two-Stream LSTM Network, designed to overcome challenges such as environmental variability, high-density crowds, and diverse pedestrian movements in real-world video data. TSLNet combines a Two-Stream Convolutional Neural Network (Two-Stream CNN) with Long Short-Term Memory (LSTM) networks to effectively capture spatial and temporal features. The addition of a Multi-Head Attention mechanism allows the model to focus on relevant features in complex environments, while Hierarchical Classifiers within a Multi-Task Learning framework enable the simultaneous recognition of basic and complex behaviors. Experimental results on multiple public and proprietary datasets demonstrate that TSLNet significantly outperforms existing baseline models, achieving higher Accuracy, Precision, Recall, F1-Score, and Mean Average Precision (mAP) in behavior recognition, as well as superior Multiple Object Tracking Accuracy (MOTA) and ID F1 Score (IDF1) in pedestrian tracking. These improvements highlight TSLNet's effectiveness in enhancing tracking and recognition performance.

Load imbalance is a major performance bottleneck in training mixture-of-experts (MoE) models, as unbalanced expert loads can lead to routing collapse. Most existing approaches address this issue by introducing auxiliary loss functions to balance the load; however, the hyperparameters within these loss functions often need to be tuned for different tasks. Furthermore, increasing the number of activated experts tends to exacerbate load imbalance, while fixing the activation count can reduce the model's confidence in handling difficult tasks. To address these challenges, this paper proposes a dynamically balanced routing strategy that employs a threshold-based dynamic routing algorithm. After each routing step, the method adjusts expert weights to influence the load distribution in the subsequent routing. Unlike loss-function-based balancing methods, our approach operates directly at the routing level, avoiding gradient perturbations that could degrade model quality, while dynamically routing to make more efficient use of computational resources. Experiments on Natural Language Understanding (NLU) benchmarks demonstrate that the proposed method achieves accuracy comparable to top-2 routing, while significantly reducing the load standard deviation (e.g., from 12.25 to 1.18 on MNLI). In addition, threshold-based dynamic expert activation reduces model parameters and provides a new perspective for mitigating load imbalance among experts.

Introduction: Traditional methods for farmland road extraction, such as U-Net, often struggle with complex noise and geometric features, leading to discontinuous extraction and insufficient sensitivity. To address these limitations, this study proposes a novel dual-phase generative adversarial network (GAN) named UHGAN, which integrates Hough-transform constraints.

Methods: We designed a cascaded U-Net generator within a two-stage GAN framework. The Stage 1 GAN combines a differentiable Hough transform loss with cross-entropy loss to generate initial road masks. Subsequently, the Stage 2 U-Net refines these masks by repairing breakpoints and suppressing isolated noise.

Results: When evaluated on the WHU RuR+rural road dataset, the proposed UHGAN method achieved an accuracy of 0.826, a recall of 0.750, and an F1-score of 0.789. This represents a significant improvement over the single-stage U-Net (F1 = 0.756) and ResNet (F1 = 0.762) baselines.

Discussion: The results demonstrate that our approach effectively mitigates the issues of discontinuous extraction caused by the complex geometric shapes and partial occlusion characteristic of farmland roads. The integration of Hough-transform loss, an technique that has received limited attention in prior studies, proves to be highly beneficial. This method shows considerable promise for practical applications in rural infrastructure planning and precision agriculture.

Introduction: Urban traffic congestion, environmental degradation, and road safety challenges necessitate intelligent aerial robotic systems capable of real-time adaptive decision-making. Unmanned Aerial Vehicles (UAVs), with their flexible deployment and high vantage point, offer a promising solution for large-scale traffic surveillance in complex urban environments. This study introduces a UAV-based neural framework that addresses challenges such as asymmetric vehicle motion, scale variations, and spatial inconsistencies in aerial imagery.

Methods: The proposed system integrates a multi-stage pipeline encompassing contrast enhancement and region-based clustering to optimize segmentation while maintaining computational efficiency for resource-constrained UAV platforms. Vehicle detection is carried out using a Recurrent Neural Network (RNN), optimized via a hybrid loss function combining cross-entropy and mean squared error to improve localization and confidence estimation. Upon detection, the system branches into two neural submodules: (i) a classification stream utilizing SURF and BRISK descriptors integrated with a Swin Transformer backbone for precise vehicle categorization, and (ii) a multi-object tracking stream employing DeepSORT, which fuses motion and appearance features within an affinity matrix for robust trajectory association.

Results: Comprehensive evaluation on three benchmark UAV datasets-AU-AIR, UAVDT, and VAID shows consistent and high performance. The model achieved detection precisions of 0.913, 0.930, and 0.920; tracking precisions of 0.901, 0.881, and 0.890; and classification accuracies of 92.14, 92.75, and 91.25%, respectively.

Discussion: These findings highlight the adaptability, robustness, and real-time viability of the proposed architecture in aerial traffic surveillance applications. By effectively integrating detection, classification, and tracking within a unified neural framework, the system contributes significant advancements to intelligent UAV-based traffic monitoring and supports future developments in smart city mobility and decision-making systems.

To enhance the obstacle avoidance performance and autonomous decision-making capabilities of robots in complex dynamic environments, this paper proposes an end-to-end intelligent obstacle avoidance method that integrates deep reinforcement learning, spatiotemporal attention mechanisms, and a Transformer-based architecture. Current mainstream robot obstacle avoidance methods often rely on system architectures with separated perception and decision-making modules, which suffer from issues such as fragmented feature transmission, insufficient environmental modeling, and weak policy generalization. To address these problems, this paper adopts Deep Q-Network (DQN) as the core of reinforcement learning, guiding the robot to autonomously learn optimal obstacle avoidance strategies through interaction with the environment, effectively handling continuous decision-making problems in dynamic and uncertain scenarios. To overcome the limitations of traditional perception mechanisms in modeling the temporal evolution of obstacles, a spatiotemporal attention mechanism is introduced, jointly modeling spatial positional relationships and historical motion trajectories to enhance the model's perception of critical obstacle areas and potential collision risks. Furthermore, an end-to-end Transformer-based perception-decision architecture is designed, utilizing multi-head self-attention to perform high-dimensional feature modeling on multi-modal input information (such as LiDAR and depth images), and generating action policies through a decoding module. This completely eliminates the need for manual feature engineering and intermediate state modeling, constructing an integrated learning process of perception and decision-making. Experiments conducted in several typical obstacle avoidance simulation environments demonstrate that the proposed method outperforms existing mainstream deep reinforcement learning approaches in terms of obstacle avoidance success rate, path optimization, and policy convergence speed. It exhibits good stability and generalization capabilities, showing broad application prospects for deployment in real-world complex environments.

Overcoming visual degradation in challenging imaging scenarios is essential for accurate scene understanding. Although deep learning methods have integrated various perceptual capabilities and achieved remarkable progress, their high computational cost limits practical deployment under resource-constrained conditions. Moreover, when confronted with diverse degradation types, existing methods often fail to effectively model the inconsistent attenuation across color channels and spatial regions. To tackle these challenges, we propose DWMamba, a degradation-aware and weight-efficient Mamba network for image quality enhancement. Specifically, DWMamba introduces an Adaptive State Space Module (ASSM) that employs a dual-stream channel monitoring mechanism and a soft fusion strategy to capture global dependencies. With linear computational complexity, ASSM strengthens the models ability to address non-uniform degradations. In addition, by leveraging explicit edge priors and region partitioning as guidance, we design a Structure-guided Residual Fusion (SGRF) module to selectively fuse shallow and deep features, thereby restoring degraded details and enhancing low-light textures. Extensive experiments demonstrate that the proposed network delivers superior qualitative and quantitative performance, with strong generalization to diverse extreme lighting conditions. The code is available at https://github.com/WindySprint/DWMamba.

Introduction: Abandonment rates for myoelectric upper limb prostheses can reach 44%, negatively affecting quality of life and increasing the risk of injury due to compensatory movements. Traditional myoelectric prostheses rely on conventional signal processing for the detection and classification of movement intentions, whereas machine learning offers more robust and complex control through pattern recognition. However, the non-stationary nature of surface electromyogram signals and their day-to-day variations significantly degrade the classification performance of machine learning algorithms. Although single-session classification accuracies exceeding 99% have been reported for 8-class datasets, multisession accuracies typically decrease by 23% between morning and afternoon sessions. Retraining or adaptation can mitigate this accuracy loss.

Methods: This study evaluates three paradigms for retraining a machine learning-based classifier: confidence scores, nearest neighbour window assessment, and a novel signal-to-noise ratio-based approach.

Results: The results show that all paradigms improve accuracy against no retraining, with the nearest neighbour and signal-to-noise ratio methods showing an average improvement 5% in accuracy over the confidence-based approach.

Discussion: The effectiveness of each paradigm is assessed based on intersession accuracy across 10 sessions recorded over 5 days using the NinaPro 6 dataset.

[This corrects the article DOI: 10.3389/fnbot.2025.1604453.].