Mobile manipulators have become essential platforms for autonomous tasks that demand high-quality performance and efficient operational processes. This paper presents a complete grocery pick-and-pack system for a mobile manipulator, integrating a graphical user interface (GUI) with an end-to-end vision-based grasp detection pipeline designed for lightweight computation. The system is evaluated on the Grocery Pick-and-Pack Benchmark (Level-3), the most challenging level due to deformable objects, dimensional constraints, and strict grasp-point requirements. Experimental results demonstrate an average success rate of 92% across five item classes, with the deformable sweet bag the most challenging at 60% and an average execution time of 7.5 s on an edge device. The system achieves strong computational efficiency, reflected by a compute-to-speed ratio (CSR) of 0.008, with a total model size of only 30.9 MB. Performance is further validated across multiple hardware platforms and under real competition scenarios in the European Robotics League 2025. The findings highlight the practical impact of lightweight, vision-based mobile manipulation and provide insights into current challenges and future research directions for autonomous robotic applications.

Reliable weld seam perception remains challenging in industrial environments, where arc light, spatter, smoke, and varying seam geometries can seriously degrade visual sensing. These disturbances make it difficult to achieve a unified representation, accurate localization, and real-time inference at the same time. To address this problem, this paper presents an end-to-end lightweight framework for weld seam keypoint detection and tracking based on an improved SimCC. A unified three-keypoint formulation is introduced to represent different weld geometries by using one seam center point and two orientation reference points, thereby supporting a perception-to-control mapping in which position control and orientation control are decoupled. In addition, a lightweight C3k2-based backbone is designed, and a non-parametric log-domain quadratic peak-refinement decoder is proposed to alleviate the discretization-induced quantization error of SimCC classification distributions without adding model parameters. Experiments show that the proposed model contains only 1.4 M parameters, achieves 17.01 ms CPU inference latency, and obtains a detection accuracy of 1.89 px MAE. In curved weld seam tracking experiments with the integrated robotic system, it further achieves an average trajectory tracking error as low as 0.159 mm and an average orientation error of 3.738°, demonstrating its real-time accuracy and robustness for industrial welding applications.

Heterogeneous unmanned ground vehicle-unmanned aerial vehicle (UGV-UAV) collaborative systems offer clear advantages for field exploration. However, when tethered unmanned aerial vehicles (TUAVs) are introduced to extend mission capability, a major compatibility gap emerges for small and highly maneuverable UGVs: existing industrial tethered ground stations are generally too heavy and bulky to be carried by such platforms. In addition, on unstructured ground, residual station tilt can significantly complicate UAV launch and recovery. To address these issues, this paper develops an ultralight vehicle-mounted tethered ground station for micro unmanned aerial vehicles (micro-UAVs) that can be integrated directly with small UGVs. Through co-design of a 2-degree-of-freedom (2-DOF) self-leveling launch platform and a passive tether-assisted recovery scheme without visual fiducials, in which a customized UAV flight-control loop is coordinated with the state transitions of the ground tether-management system, the proposed system achieves practical tether-assisted recovery. Experiments show that the complete platform weighs only 4.1 kg and that the self-leveling mechanism compensates for ground inclinations over a total range of 24 degrees. Repeated passive-landing tests further demonstrate the feasibility of the proposed recovery scheme and its tolerance to moderate bay tilt and terminal off-axis activation. System-level flight validation confirms practical tether-assisted recovery without visual fiducials. In addition, we conduct a simplified exploratory simulation of tether-based ground-anchor localization under the proposed system architecture. Overall, these results establish a lightweight and low-cost hardware design and a practically viable recovery strategy for multimodal micro air-ground robotic systems.

The aim of this study is to propose an uncertainty-aware dynamic cascading framework based on spiking neural network (UDC-SNN) for multimodal emotion recognition, particularly to address the inherent trade-off between recognition performance and energy efficiency. An asymmetric dynamic routing mechanism was proposed to enable demand-driven activation of the high-power electroencephalogram (EEG) branch, coupled with preliminary inference on a low-power electrocardiogram (ECG) branch and uncertainty quantification via Shannon entropy. Meanwhile, a parameter-free log-linear aggregation strategy was developed to transform modality-specific entropy into dynamic Bayesian weights through an exponential decay function, effectively mitigating the negative transfer effects induced by unimodal noise. The UDC-SNN was evaluated on the multimodal affective dataset DREAMER, comprising 23 subjects (170,660 segments). The averaged recognition accuracy and energy consumption across the three dimensions of valence, arousal, and dominance were 90.75% and 4.62 J, respectively. The obtained results suggest that the proposed framework could potentially achieve a favorable balance between high emotion recognition and low energy consumption, thereby establishing its applicability for real-time monitoring in resource-constrained scenarios.

Wireless Sensor Networks (WSNs) are widely used in environmental monitoring, disaster early warning, and smart grids. However, sensor nodes face strict energy limitations. Unbalanced energy consumption and hotspots severely shorten the network lifetime. To address these problems, this paper proposes an optimized Spotted Hyena Optimization-Energy-Efficient Non-Uniform Clustering algorithm (SHOE) for cluster head selection and data transmission. The algorithm has three main innovations: combining a bio-inspired metaheuristic with an improved EEUC (Energy-Efficient Unequal Clustering) multi-hop relay and a Gaussian distribution model for non-uniform node deployment; designing a multi-dimensional fitness function considering energy, distance, and node location; and introducing empty cluster and isolated node repair mechanisms to balance exploration and exploitation. Specifically, the multi-dimensional fitness function guides the heuristic search process towards high-quality cluster head candidates, while the empty cluster and isolated node repair mechanisms dynamically rectify abnormal network structures, ensuring the robustness of the final architecture optimized by the bio-inspired framework. Simulations in MATLAB show that SHOE outperforms LEACH (Low-Energy Adaptive Clustering Hierarchy), PSOE (Particle Swarm Optimization with Evolutionary Strategy), PL-EBC (Probabilistic Localized Energy-Balanced Clustering), and CGWOA (Chaotic Grey Wolf Optimization Algorithm) in reducing node death, saving energy, and extending network lifetime. It improves adaptability to non-uniform distribution and optimizes energy balance, thus enhancing the efficiency and stability of WSNs.

Classification of vehicle engines using the chemical composition of the exhaust from these engines can be used to identify the engine’s design and verify compliance with environmental regulations through the vehicle’s emissions. This paper describes a method to identify the type of vehicles using machine learning (ML), where low-cost MQ series sensors measure the gases and particle emissions from a vehicle exhaust system, while simultaneously collecting and measuring the vehicle’s temperature and humidity levels. A custom-designed multi-sensor exhaust sensing module is employed to capture real-time exhaust emissions prior to entering the atmosphere. Exhaust samples are collected from vehicles representing three major engine categories: petrol, diesel, and compressed natural gas (CNG). In addition, fresh air samples are collected as a baseline environmental reference for comparison. All exhaust measurements are collected under controlled and consistent engine operating conditions to ensure comparable emission profiling across vehicle classes. To ensure consistent combustion-based emission profiling, this study focuses on conventional fuel-powered vehicles. MQ-series gas sensors are sensitive to combustion by-products emitted during engine operation, such as carbon monoxide (CO), hydrocarbons (HC), while also exhibiting cross-sensitivity to other gaseous components present in exhaust mixtures. Nevertheless, the proposed system performs pattern-based classification using relative sensor response signatures. Standardization of data is achieved through z-score normalization. The best models developed (based on three separate experimental designs) are trained and validated using six supervised machine learning algorithms such as Logistic Regression, Support Vector Machine (RBF), k-Nearest Neighbors, Random Forest, Gradient Boosting Decision Tree, and XGBoost and are compared against one another. Evaluation of the tested algorithms using various evaluation metrics demonstrated that ensemble models outperformed all other algorithms, achieving the highest accuracy of 99.5%. Furthermore, noise analysis confirms that the proposed solution maintains high classification accuracy (more than 89%) even under substantial sensor perturbations mimicking the real-world deployment. The solution proposed below illustrates how using gas sensors and advanced algorithms can provide accurate exhaust identification and identify engines in real-time.

Facial expression recognition (FER) for assistive and telepresence robotics remains challenging under resource-constrained conditions because landmark normalization is often unstable, many datasets have limited variability, and full facial landmark sets introduce redundancy. This paper proposes a lightweight, privacy-preserving FER framework for assistive healthcare robotics based on geometric facial landmarks rather than raw RGB images. The objective is to improve recognition robustness and deployment suitability on low-power edge devices through two complementary contributions: a revised nose-centered landmark normalization method and an optimized Facial Feature Mapping, FFM-L03. The proposed normalization replaces the expression-sensitive upper-lip reference with a geometrically stable nose-center anchor, while FFM-L03 combines FACS-guided anatomical priors with ANOVA F-score, LASSO, PCA, and t-SNE/UMAP to retain 60 informative landmarks. In addition, a heterogeneous Freepik dataset was constructed to increase variability in lighting, background, resolution, and subject appearance. Experimental evaluation across 15 landmark groups, four datasets, and four classifiers shows that the proposed method consistently improves performance over prior landmark configurations, achieving gains of up to 22.4 percentage points over the Ciraolo baseline and 22.1 percentage points over the full-landmark baseline in accuracy, precision, recall, and F1-score, while maintaining lightweight operation. These results demonstrate that principled normalization and targeted landmark selection can substantially improve FER for real-time, privacy-aware assistive robotic systems.

HighlightsWhat are the main findings?A microfluidic-based method enabled simultaneous quantification of blood viscosity and RBC aggregation index under continuous blood flow from a driving syringe.Hemorheological properties were strongly affected by experimental factors and thermal shock, which suppressed RBC aggregation and sedimentation.What are the implications of the main findings?The method allows for the reliable evaluation of blood properties under dynamic flow conditions, including syringe on–off operation.The method could be regarded as useful for assessing RBC dysfunction and abnormal hemorheological responses in microfluidic platforms.Accurate assessment of blood viscosity and red blood cell (RBC) aggregation under continuous flow is important for hemorheological analysis. However, simultaneous measurement remains challenging because both properties are influenced by flow conditions and RBC sedimentation. In this study, a microfluidic method is developed for the simultaneous measurement of blood viscosity and RBC aggregation index (AI) during continuous blood delivery from a driving syringe. The proposed device consists of a viscosity-sensing channel for viscosity measurement and aggregation-sensing channel for AI evaluation. The effects of flow rate, hematocrit, suspension medium, and syringe on–off operation are systematically investigated. Blood viscosity and AI are strongly affected by these factors, and transient flow interruption enhances RBC sedimentation in the syringe, thereby altering hemorheological properties. The proposed method is further used to evaluate thermally exposed RBCs, which reduce RBC aggregation and suppress RBC sedimentation when compared with control blood. At higher exposure temperatures and longer exposure times, blood viscosity and AI remain nearly constant over time, indicating minimal contribution of damaged RBCs to RBC sedimentation. These results demonstrate that the proposed method enables reliable simultaneous evaluation of blood viscosity and RBC aggregation and could be regarded as useful for detecting functional alterations of RBCs under continuous-flow conditions.

For large-scale embedded sensor-actuator networks, such as robotic swarms deployed over vast areas and other embedded intelligent devices, end-to-end message exchange is often impossible due to their limited communication range, power constraints, and device mobility. Devices, thus, rely on multi-hop relaying, exposing them to Man-in-the-Middle (MitM) attacks where compromised relays tamper with, forge, or inject false messages. The existing countermeasures, including end-to-end encryption or Byzantine consensus, involve high overhead while requiring global coordination and, thus, renders them impractical for time-sensitive message exchange in embedded intelligence. Security management on communication among embodied devices is highly desired. To address this challenge, we propose Reputation-Guided Dynamic Relay Selection (RDRS), a lightweight, distributed countermeasure against MitM attacks that leverages interactive feedback to evaluate reputation of embedded devices. Specifically, each device maintains reputation scores updated via recent interaction success rates with decay factors to counter dynamic adversaries. During exchanging messages, embedded devices select next-hop neighbors weighted by reputation scores, effectively bypassing malicious devices without explicit detection or in-path verification. Comprehensive simulations in embedded sensor-actuator networks demonstrate that RDRS reduces tampering success rate (TSR) by 80–95% compared to the baselines, martians request satisfaction rate (RSR) above 79% even at 40% malicious nodes, and achieves lower delay 64% with comparable overhead.

Accurate classification of seal wear is essential for condition-based and predictive maintenance of hydraulic cylinders, where seal degradation can cause fluid leakage and impair normal system operation. This study investigates the adaptation of a Transformer-based audio model for classifying seal wear conditions using acoustic emission (AE) signals. Specifically, we adapt the Audio Spectrogram Transformer (AST), a convolution-free, purely attention-based model that operates directly on audio spectrograms. The Transformer architecture enables the modeling of long-range dependencies, while the model learns discriminative representations directly from AE data without relying on manually engineered features. A selective fine-tuning strategy was implemented by adding layer-freezing functionality to the AST training pipeline, enabling different freezing configurations during fine-tuning. This allowed earlier pretrained representations to be preserved while adapting the later layers to the target AE signals, thereby reducing the risk of overfitting in the small-data setting. In addition, validation-driven early stopping was implemented to further improve generalization during fine-tuning. The model was initialized with ImageNet and AudioSet pretrained weights to exploit general-purpose representations learned from large-scale datasets. The AE data were acquired under varying pressure conditions on a hydraulic test rig designed to simulate hydraulic cylinder leakage. The datasets were partitioned into fine-tuning, validation, and evaluation subsets and labeled into three wear states: unworn, semi-worn, and worn. In addition, data augmentation techniques were applied to the fine-tuning data to increase diversity and mitigate class imbalance. The adapted model achieved 97.92% classification accuracy across all wear conditions and pressure settings, demonstrating its ability to learn discriminative wear-related patterns directly from AE data.Furthermore, the framework’s versatility was further assessed on a bearing strip dataset acquired from the same hydraulic test rig. Using the same fine-tuning configuration, the model achieved 95.65% accuracy and 100% recall for the worn state. These findings highlight the potential of transformer-based architectures for data-efficient, end-to-end AE-based diagnostics across hydraulic system components.