Journal of Robotics, Networking and Artificial Life

Volume 11, Issue 4, September 2026

Preface
Preface
Takao Ito
Pages: 269
Download article (PDF)

Research Article
1. Mode-Scheduled RBF-PID Control for Attitude Stabilization of an Unmanned Aerial Manipulator Under Wall-Collision Disturbance
Qinlin Li, M. A. Ilyas, Nazhatul Hafizah Kamarudin, Deivasigamani.S, Chow LS
During high-rise building exterior wall crack inspection, an unmanned aerial manipulator (UAM) equipped with a 0-degree-of-freedom (0-DOF) manipulator suffers from attitude instability caused by transient wall collision disturbances. To address this problem, this study develops a dedicated flight control strategy. We first established a full rigid-body dynamic model with a non-diagonal inertia matrix for the asymmetric UAM platform. A mode-scheduled RBF neural network PID control strategy (RBFPID-MS) was then proposed, which integrates an improved double-momentum gradient descent algorithm to improve online convergence of the RBF network. Based on the impulse theorem, we constructed a transient collision disturbance model for wall impact scenarios and calibrated the −20° pitch angle disturbance threshold. Simulation experiments were conducted under steady translational flight and wall collision scenarios, with conventional PID as the benchmark controller, while response speed, attitude overshoot, and anti-disturbance robustness were used as core evaluation metrics. Simulation results indicate that the proposed RBFPID-MS controller improves disturbance rejection and reduces post-collision attitude overshoot relative to a conventional PID controller. This study provides a practical reference for controller design of aerial operation platforms subjected to transient collision disturbances. The proposed approach contributes to safer and more sustainable urban infrastructure inspection.
Pages: 270 - 280
Download article (PDF)

Research Article
2. The Impact of V2X Penetration Rates on Mixed-Traffic Coordination
Zhi-Yuan Su, Wei-Xiang Li, Jung-Shian Li, I-Hsien Liu, Kuan-Ting Lee
Navigating unsignalized mixed-traffic intersections presents a complex decentralized decision-making challenge. To address this, we propose a Dynamic Bayesian Game (DBG) framework that continuously recalculates optimal strategies under real-world V2X communication uncertainties. Our approach integrates a physical-layer model, utilizing AWGN and path loss-dependent SNR to evaluate realistic packet delivery. Crucially, the system relies on an autonomous belief-updating mechanism where vehicles construct probability distributions over opponents' hidden driving types. These beliefs are instantly refined upon receiving V2X packets, while safely reverting to a shared prior distribution during transmission dropouts. Comprehensive MATLAB-SUMO co-simulations confirm the architecture's ability to ensure robust vehicle safety and high macroscopic efficiency despite severe communication instabilities.
Pages: 281 - 285
Download article (PDF)

Research Article
3. The Dilemma of Self-Driving Cars: Evaluating Decision-Making Uncertainty at Unsignalized Intersections
I-Hsien Liu, Kuan-Ting Lee, Wei-Xiang Li, Chu-Fen Li
Interactions between autonomous vehicles (AVs) and human-driven vehicles (HVs) remain challenging in mixed-traffic environments, particularly at unsignalized intersections where right-of-way decisions rely on implicit negotiation. This study analyzes AV-HV encounters using a microscopic traffic simulation framework and temporal interaction metrics derived from vehicle trajectories. The results show that many encounters occur within a near-neutral timing region where small temporal differences may alter yielding outcomes. This phenomenon reveals the inherent ambiguity of mixed-traffic interactions and suggests that such uncertainty may be associated with the staleness of dynamic information. The findings provide insights into how information uncertainty can influence decision-making in autonomous driving systems.
Pages: 286 - 290
Download article (PDF)

Research Article
4. LLM-Supervised Adaptive Mutation Control for Evolutionary Behavior Tree Optimization in Multi-Objective Coastal Cleanup Robotics
Chi Jie Tan, M.A. Munjer, Pan Wei Zheng, Eiji Hayashi, Way Soong Lim
This study investigates the use of large language models (LLMs) as adaptive supervisory controllers for evolutionary optimization of behavior trees in multi-robot systems. Unlike conventional genetic programming approaches that rely on fixed mutation and crossover parameters, the proposed framework integrates an LLM as a high-level reasoning module that dynamically adjusts evolutionary parameters based on observed population dynamics during the search process. This enables the system to maintain a balance between exploration and exploitation without requiring manual tuning of evolutionary settings. A comparative study is conducted between the proposed LLM-guided evolutionary framework and a conventional genetic algorithm with static parameters in a coastal cleanup behavior tree optimization task. The results demonstrate that incorporating LLM-based supervision leads to consistently improved optimization performance and more effective search behavior. In particular, the proposed method achieves higher-quality solutions and faster convergence to high-fitness regions compared to the baseline approach, while maintaining greater behavioral diversity throughout the evolutionary process. Further analysis shows that LLM-guided parameter adaptation enhances the generation of diverse and specialized behavioral strategies, particularly in complex multi-objective scenarios involving task execution, resource management, and environmental interaction. These findings suggest that integrating language-model-based reasoning into evolutionary search can improve both the efficiency and robustness of behavior tree optimization, offering a promising direction for adaptive robotic control system design.
Pages: 291 - 298
Download article (PDF)

Research Article
5. Neural Dynamics of Inter-Brain Phase Synchronization in EEG During Coordinated Rhythmic Interaction
Rena Kato, Akio Wakata, Kosei Shibata, Shabbir Mahmood, Diunuge Buddhika Wijesinghe, Yide Yang, Masayuki Fujiwara, Laurent Bougrain, Kiyohisa Natsume, Hiroaki Wagatsuma
Investigating how neural synchronization emerges during joint movement is essential for understanding interpersonal coordination dynamics. This study investigated inter-brain phase synchronization as pairs of individuals performed rhythmic arm movements during various phases of a game. Electroencephalography (EEG) signals were simultaneously recorded from both individuals as they performed a rock-paper-scissors game initiated by a synchronized arm swing. Particular attention was given to transitions from uncoordinated to coordinated states. Temporal analysis techniques have been applied to capture dynamic changes in inter-brain connectivity linked with the emergence of coordinated behavior. The proposed approach provides a quantitative framework for evaluating neural coupling and may offer potential information for designing adaptive human-robot interaction systems.
Pages: 299 - 305
Download article (PDF)

Research Article
6. An Adaptive 2D–3D Approach for Urban Tree Pruning Using Computer Vision and Quantitative Structure Modeling
Mohammad Albaroudi, Abdullah Alraee, Hussam Alraie, Raji Alahmad, Kazuo Ishii
Urban trees provide significant ecological and aesthetic benefits in urban environments; however, ensuring their safe maintenance requires regular pruning to prevent conflicts with surrounding infrastructure. Traditional pruning practices rely heavily on manual labor and expert judgment, making them time-consuming, potentially hazardous, and highly dependent on experience. This study proposes a computer vision–based decision-support system for urban tree pruning that integrates two-dimensional (2D) perception with three-dimensional (3D) structural modeling. For trees with relatively simple crown structures, pruning decisions are derived from image-space coordinates. Tree trunks and branches are detected from RGB images using a YOLOv8-OBB model, enabling the extraction of branch positions to identify candidate branches for two common pruning strategies: crown raising and crown reduction, based on operator-defined height thresholds. For trees with dense or overlapping crowns, a Quantitative Structure Model (QSM) is constructed from point cloud data to provide a more complete representation of tree geometry. For the proposed 2D system, experimental results demonstrate reliable tree structure analysis using a large-scale dataset collected in this study, where the detection accuracy exceeds 90%. While pruning simulations for both the proposed 2D and 3D systems confirm stable and controllable crown raising and reduction under varying threshold conditions. These systems provide an efficient and scalable decision-support for safe, consistent, and interpretable urban tree management.
Pages: 306 - 315
Download article (PDF)

Research Article
7. Proposal of Coding Support Methodology Using Artificial Language and Generative AI to Address the Shortage of Software Developers
Masaya Nakahara, Kaito Andachi
In recent years, the shortage of IT professionals has become a critical issue globally, and particularly in Japan. While tools like GitHub Copilot improve efficiency for existing programmers, they often require a certain level of coding proficiency, leaving non-programmers or novice engineers unsupported. This study proposes a novel program generation method designed for novices with logical thinking skills but limited coding experience. The method utilizes a strictly defined "artificial language" (a constrained pseudo-flowchart) as a prompt for Generative AI, enabling users to generate code by focusing on logical flow rather than syntax. Empirical experiments compared the proposed method with manual coding by beginners. The results demonstrated that the proposed method achieved a 97% success rate compared to 41% for manual coding and significantly reduced average development time. These findings suggest that the proposed methodology can effectively convert non-programmers into active contributors in the development process.
Pages: 316 - 326
Download article (PDF)

Research Article
8. A Report on Conducting a Fundamental Mechanics Course as a Student-Led Community Contribution Activity
Kenji Kimura
In recent years, educational institutions have been actively engaging in robotics and STEAM education as part of their contributions to the local community. To address the issue of students losing interest in science, classes are being developed to stimulate early interest in mathematics among elementary school students. In this study, the students selected their own theme, developed teaching materials to convey the enjoyment of mechanics, and conducted workshops as a community outreach activity.
Pages: 327 - 330
Download article (PDF)

Research Article
9. Interpretive Gaps and Minimal Supplementation in Dialogue: A Category-Theoretic Perspective
Yukiko Furuya, Akinori Abe
Dialogue is susceptible to interpretive divergence. This paper introduces a category-theoretic framework for analyzing interpretive gaps in dialogue and characterizing the minimality of their repair. Building on our empirical study of improvisational acting, we distinguish two types of divergence: passive gaps, arising from cumulative drift of object representations across a semiotic chain, and active gaps, resulting from intentional abductive reconstruction of the interpretive frame. Within a unified categorical formalism, these correspond to object drift under a fixed analogical functor and functor switching, respectively. We characterize minimal supplementation via the universal property—a supplement is minimal if every other valid supplement factors uniquely through it—and prove its uniqueness up to isomorphism. Empirical grounding is provided through 106 inference episodes from five dyads, revealing significant associations between inference type and gap type (χ²(6) = 32.59, p < .01, Cramér's V = 0.39).
Pages: 331 - 338
Download article (PDF)

Research Article
10. Multi-Agent Simulation of Influenza Outbreaks Based on Small Data to Evaluate the Effect of Control Measures
Saori Iwanaga
It is difficult to predict where, when, and who will be affected by any given epidemic. An asymptomatic carrier of influenza can be infectious before the onset of symptoms, which makes it difficult to determine the infection status, such as when, where, and from whom they were infected. In this study, first, the spatiotemporal behavioral history and clinical information of all students at a Japanese boarding school during multiple influenza epidemic waves were collected and recorded as small data. Next, a spatiotemporal mathematical model (SEPIR) based on this data was proposed, and the parameters were calculated. It is confirmed that the infectivity rate should be divided into two phases and types of rooms. Finally, the effectiveness of several infection control measures against influenza outbreaks was evaluated using validated Multi-Agent Simulation (MAS).
Pages: 339 - 342
Download article (PDF)

Research Article
11. Interpretable Multimodal Data Fusion for Urban Hazard Forecasting and Constraint-Aware Decision Support in Smart Cities
Young-Im Cho, Doniyorjon Mukhtorov
This paper presents a multimodal framework for urban environmental intelligence that brings public and sensor data into a single pipeline for hazard detection and risk forecasting. Heterogeneous streams—weather, air quality, energy load, mobility, and IoT micro-sensing—are aligned in time and space and fused at the feature level so that cross-domain effects can be modelled directly. A Temporal Fusion Transformer produces multi-horizon forecasts, and a SHAP-based attribution layer reports how each input contributes to a given estimate, both overall and for individual alerts. A constraint-aware step then turns each risk estimate into a short list of interventions that respect operational, regulatory, and resource limits, keeping a human operator in the loop. On data from several Korean cities, the approach improved forecasting while keeping its outputs interpretable, with meteorological and air-quality variables emerging as the dominant short-term drivers. We also discuss how the framework would be deployed in a city operations setting, its limitations around transfer and data quality, and the directions we consider most useful for further work.
Pages: 343 - 349
Download article (PDF)

Research Article
12. Two-Stage YOLOv4-Tiny for Vascular Stenosis Severity Classification and Hemodynamic Correlation Analysis Using AVF Shunt Sound
Lucky Nindya Palupi, Hiroyuki Fukumoto, Yasuomi Tanaka, Kazuya Kibune, Hiroki Tamura
The two stage YOLOv4-Tiny framework for vascular stenosis using arteriovenous fistula (AVF) shunt sound spectrograms is introduced. AVF sounds were recorded using a digital stethoscope and converted into spectrogram images for classification. In the proposed hierarchical framework, the first stage classified samples into normal and non-normal categories, while the second stage further separated non-normal cases into gray and abnormal conditions. To improve training diversity, 1-second and 3-second spectrogram slice configurations were investigated, with the 1-second slices achieving the best file-level accuracy of 67.92% for the final three-class classification task. To further investigate the physiological relevance of the model outputs, confidence-derived features generated by the proposed framework were analyzed in relation to Doppler ultrasound parameters, including Flow Volume (FV), Pulsatility Index (PI), and Resistance Index (RI). Pearson correlation analysis revealed moderate relationships between several confidence-derived features and Doppler ultrasound parameters, with FV showing comparatively stronger correlation tendencies (r ≈ 0.55–0.60) than PI and RI. These findings suggest that the proposed framework can capture stenosis-related acoustic characteristics and may provide supplementary physiological information for non-invasive AVF monitoring vascular assessment.
Pages: 350 - 356
Download article (PDF)

Research Article
13. Comparative Evaluation of Transfer Learning CNN Architectures and Optimizers for Chinese Herbal Leaf Recognition
Lyu SiQian, Heshalini Rajagopal, Muhammad Amirul Aiman Asri, Zaris Izzati Mohd Yassin, Peng Lean Chong, Norrima Mokhtar
Accurate recognition of Chinese herbal leaves is essential in safeguarding the effectiveness of traditional medicinal practices, maintaining product quality standards, and preserving valuable plant biodiversity. Conventional identification approaches mainly depend on human expertise, which can be laborious, inconsistent, and vulnerable to classification errors. Furthermore, many computer-based recognition systems still face challenges in differentiating leaves with highly similar textures, venation patterns, and shapes. This study investigates an optimized training configuration for a ResNet-50 transfer learning model for automated Chinese herbal leaf recognition. The classification performance of five established convolutional neural network (CNN) architectures, namely ResNet-50, EfficientNet-B0, GoogLeNet, SqueezeNet, and DarkNet-53, was compared across 13 categories of Chinese medicinal leaves. The original dataset comprised 520 images, which were first divided into training and testing subsets using an 80:20 stratified split. Image augmentation was subsequently applied exclusively to the training subset to increase training diversity. The models were evaluated using accuracy, precision, recall and F1-score, while Adam, RMSprop, and stochastic gradient descent with momentum (SGDM) were compared using a common initial learning rate of 0.0001 and a batch size of 40. Experimental results showed that ResNet-50 trained using Adam achieved the best overall classification performance, with an accuracy of 0.911, precision of 0.901, recall of 0.910, and F1-score of 0.896. EfficientNet-B0 trained using RMSprop also demonstrated competitive performance, achieving an accuracy of 0.899 and the recall of 0.909. In contrast, SGDM generally produced lower performance across most evaluated architectures. These findings demonstrate that optimizer effectiveness is dependent on the CNN architecture and indicate that the ResNet-50 transfer learning model with Adam provides a promising approach for automated Chinese herbal leaf recognition.
Pages: 357 - 362
Download article (PDF)

Research Article
14. Classification of Edge and Side Contacts in Table Tennis Using High-Speed Camera Images
Kazuma Sakamoto, Reo Ishii, Iori Iwata, Riku Kaiba, Yoshihiro Ueda
This research investigates an image-based method for supporting edge and side classification in table tennis using a camera arrangement intended to reduce interference with play. Ball contacts generated under controlled conditions were recorded using two synchronized high-speed cameras. Images from one camera were used for classification, whereas the second view was used to confirm the contact position and assign the ground-truth label. A frame visually judged to be closest to the contact moment was selected for each contact. The ball region was estimated using YOLO11s-based instance segmentation, and the primary evaluation used a white-line region manually specified from the visible table geometry. Each contact was classified according to the proportion of the ball region overlapping the white-line region. Five repeated holdout evaluations compared the normalized overlap ratio with the raw number of overlapping pixels and examined the influence of alternative white-line definitions. The normalized overlap ratio achieved a mean accuracy of 84.2%, with edge and side recalls of 86.0% and 82.4%, respectively. These results indicate that the local spatial relationship between the ball and the white line can support edge and side classification when an appropriate contact frame is available. Automatic contact-frame extraction and a physically consistent automatic white-line reference remain necessary for practical deployment.
Pages: 363 - 372
Download article (PDF)

Research Article
15. An Automated Insect Hole Detection System for Wood Carvings Utilizing Enhanced YOLOv11
Jun-Lin Wu, Chung-Wen Hung
Wood-carved artifacts often develop tiny insect holes during long-term preservation. Traditional manual inspection is a response to such defects, but it is inefficient and highly dependent on human judgment. A lightweight vision-based framework for automatic insect hole detection in complex wood-grain scenes is discussed in this paper. The proposed system is developed from an enhanced YOLOv11 model and combines image super-resolution with feature refinement to improve the recognition of extremely small targets. To reconstruct higher-resolution inputs for clearer representation of micro-holes, an Efficient Transformer for Single Image Super-Resolution (ESRT) is employed, while an Edge-Gaussian Aggregation (EGA) module is introduced to strengthen shallow edge information and reduce interference in deeper features. In addition, Slicing Aided Hyper Inference (SAHI) is adopted to preserve local target details in high-resolution images. Experimental results show that the proposed framework increases the Mean Average Precision (mAP) from 0.796 to 0.859 and the F1-score from 0.779 to 0.829, while the memory usage of the model parameters only increases slightly from 2.6M to 2.7M. The results demonstrate that the proposed method improves the reliability of micro-hole detection in wood carvings while maintaining a compact model design.
Pages: 373 - 376
Download article (PDF)