Keywords: AI and Applications
Abstract: Vision–language models (VLMs) offer a promising interface for extracting package dimensions from electronic component drawings and ingesting them into printed circuit board (PCB) libraries. Package height is critical for downstream clearance checks, placement, and reuse; when a VLM returns an incorrect height value instead of abstaining, that value can be automatically ingested into library data. This paper studies a height-axis grounding failure in which a VLM reads visible dimensions and recognizes views but maps package height to the wrong physical axis when drawing layout semantics diverge from the PCB-grounded package baseline. We construct a controlled diagnostic set of 75 drawings from 15 package types under five layout conditions, with three runs per package–condition–provider case across OpenAI, Google, and Anthropic VLMs. Height extraction drops from 120/135 (88.9%) in Canonical+IDs to 61/135 (45.2%) in Rotated+IDs, while warning prompts and explicit seating-plane cues provide only limited recovery. To reduce the risk of automatic ingestion, we propose Top-View-Grounded Height Verification (TVGHV), a conservative verification layer that accepts direct height answers from VLMs only when they are supported by top-view-grounded drawing evidence. The proposed TVGHV can reduce the number of incorrect height answers from 14 to 2, while routing only 15/90 (16.7%) cases to human review.

Keywords: Intelligent Industrial Systems, Manufacturing Automation and Systems, AI and Applications
Abstract: Assembly tasks remain fundamental to modern manufacturing, where human–machine collaboration is essential for handling complex and variable scenarios. Among these tasks, snap-fit assembly requires precise control of insertion angle and applied force, making it susceptible to subtle failure modes. A critical yet often overlooked case is false success, in which components appear correctly assembled but are not securely engaged, resulting in latent quality risks. To address this issue, this paper proposes a multimodal sensing-based quality evaluation mechanism for snap-fit assembly. The proposed system integrates pressure, acoustic, and acceleration sensing to capture complementary spatiotemporal characteristics of the assembly process. Pressure signals characterize interaction dynamics, acceleration signals capture impact responses during contact, and acoustic signals capture transient pulse characteristics at the snapping moment. These heterogeneous signals are jointly processed for event triggering and fine-grained assembly behavior analysis. Based on the extracted multimodal features, a lightweight machine learning model is developed to classify assembly outcomes into successful assembly, failure, and false success. Experimental results show that the proposed multimodal configuration achieves 98.0% accuracy under 5-fold cross-validation. The results demonstrate the effectiveness of the proposed approach in distinguishing fine-grained assembly behaviors and support its potential for real-time human-centered assembly assistance. Future work will focus on deployment and validation in real-world production environments.

Keywords: Robotic Systems, Autonomous Vehicle, Optimization and Self-Organization Approaches
Abstract: This paper presents a gradient-based method for obstacle-avoidance trajectory optimization of unmanned aerial vehicles (UAVs) in architecture inspection missions. To improve computational efficiency, selective activation of collision-aware constraints is introduced into the optimization process. The target structure is first reconstructed from multi-view images using Structure-from-Motion (SfM), and the resulting point cloud is filtered by DBSCAN, downsampled, and converted into oriented bounding boxes (OBBs) for compact obstacle representation. A constrained optimization problem is then formulated to generate a collision-free UAV path, and the activated collision-aware constraints are incorporated into a Lagrangian formulation solved by the Steepest Descent Method with one-dimensional Newton-based line search. The proposed framework integrates scene reconstruction, point-cloud simplification, obstacle abstraction, and gradient-based trajectory optimization into a unified path planning process. The developed method supports safe and efficient UAV navigation in complex three-dimensional environments for applications such as building inspection, post-disaster assessment, and smart inspection tasks.

Keywords: Biometrics and Applications, Deep Learning, Biometric Systems and Bioinformatics
Abstract: Audio deepfake detection systems are increasingly deployed in high-stakes security applications, yet their fairness across demographic groups remains critically underexamined. Prior work measures gender disparity but does not investigate where it comes from or how to fix it systematically. We present the first diagnosis-first framework that identifies bias source before applying targeted mitigation, evaluated on two models, AASIST [5] and Wav2Vec2+ResNet18 [7], on ASVSpoof5 [3]. Our diagnosis shows that bias does not stem from imbalanced training data but from acoustic representation differences, gender leakage in learned features, and structural evaluation asymmetry. We test mitigation strategies across in- processing, post-processing and combined families, including novel methods introduced in this work. Adjusting the decision threshold separately per gender reduces unfairness by 54–75% at no cost to detection accuracy, and our new epoch-level fairness regularisation method outperforms existing per-batch approaches. Adversarial debiasing succeeds only when gender leakage is localised, and fails when it is diffuse, an outcome correctly predicted by our diagnosis before training. No single method fully closes the fairness gap, confirming that bias sources must be identified before fixes are applied and that fairer benchmark design is equally important.