TL;DR
This study proposes “Vanguard AI,” an end-to-end cloud-edge architecture utilizing an experimentally validated YOLOv11m detection framework, delivering the first unified, production-viable architecture for context-aware compliance auditing and dynamic risk assessment in industrial environments.
This paper develops a BIM-integrated, explainable, and deployment-aware decision-support framework for real-time personal protective equipment (PPE) monitoring in Yemen’s construction sector. Its contribution is an integration-based decision-support artifact rather than a new PPE detection algorithm. The study addresses a persistent practical gap: many vision-based PPE systems can detect violations, yet they rarely convert image-level detections into auditable, location-aware, and managerially defensible interventions. Using a design science research approach, the paper re-specifies the original detection-centered concept as a socio-technical artifact composed of six tightly coupled layers: multimodal site capture, RF-DETR-based PPE perception, explanation generation, BIM spatial anchoring, AHP-TOPSIS-driven prioritization, and governance-oriented analytics. The assessment remains analytical, and field validation is left for future work. The framework formalizes an event schema that links each alert to confidence, explanation evidence, anchor confidence, zone semantics, response ownership, and closure status. It also introduces a resource-aware deployment path suited to fragile and connectivity-constrained projects by combining smartphone inspections, CCTV streams, offline buffering, staged BIM anchoring, and selective explanation triggering. The main contribution is therefore not merely improved PPE recognition; rather, it is the conversion of real-time vision outputs into trustworthy safety intelligence that supports prioritization, hotspot discovery, accountability, and progressive digital-twin readiness.
This work evaluates several state-of-the-art VLMs across two datasets and multiple prompting strategies to test whether an explicit distinction between hazard and anomaly changes model behavior, and shows that explicitly separating anomaly from hazard provides a more informative evaluation of VLM safety reasoning and exposes failure modes that binary safety judgments can obscure.
Smart-city airspace is transforming Uncrewed Aerial Vehicles (UAVs) from passive sensing platforms into cyber-physical decision makers that must follow operational rules under degraded observations and ambiguous language. Existing UAV and multimodal benchmarks evaluate perception, navigation, collaboration, and reasoning, but few assess whether physical evidence, protocol constraints, and action risk remain coupled during critical decisions. We introduce MulRobBench, an offline, protocol-conditioned benchmark for Vision-Language-Action (VLA) UAV agents in smart-city environments. MulRobBench integrates real UAV multimodal observations, protocol-level security policies, and action-level cyber-physical safety into a unified evaluation framework. The benchmark contains 3,024 samples spanning 17 task taxonomy nodes and 12 scoring dimensions across four stages: operational context understanding, multimodal evidence arbitration, degradation-aware reasoning, and risk-aware action planning. Evaluation combines semantic scoring with structural diagnostics, including policy compliance, format compliance, unsafe actions, parsing failures, and dimension-level validity. Across 17 multimodal models, the best semantic protocol-decision score reaches only 0.5141, while the best strict mean scoring-dimension accuracy is 0.1599. A controlled 20-anchor modality-ablation study changes 4-15 action selections per model, confirming that both visual and textual inputs influence decisions. Analysis identifies modality-trust selection, constraint extraction, glare, missing data, and operator shorthand as the primary causes of decision instability. MulRobBench provides a reproducible benchmark for trustworthy multimodal UAV decision making under realistic operational constraints.
Logistics Cyber-Physical Systems (LCPS) generate large volumes of regulatory and operational texts that encode early signals of safety risks. Converting short, noisy, and domain-specific records into actionable intelligence is difficult due to industrial semantic drift and the limited auditability of black-box predictors. This paper proposes Neuro-Symbolic Logistics Risk Awareness (NS-LRA), a dual-channel framework that integrates lightweight semantic perception with constraint-aware topological reasoning. NS-LRA first maps raw texts to a standardized schema of $K=20$ risk nodes using a dual-weighted embedding mechanism that combines TF-IDF and Word2Vec to mitigate short-text sparsity. It then constructs a directed risk graph by fusing co-occurrence evidence with a domain constraint mask, and derives hierarchical propagation via ISM level partitioning with deep-driver identification via MICMAC analysis. We evaluate NS-LRA on $N=8,435$ records, validated against an annotated subset $(\mathcal{D}_{\text{ann}}=1,500)$) with Fleiss' $\kappa=0.82$ and an expert-defined gold graph. NS-LRA achieves Micro-$\mathrm{F} \mathrm{1}=\text{0. 8 7 9}$ for risk mapping and approximately $22 \times$ lower perrecord CPU latency than fine-tuned BERT on the same test split under the same environment. For topological inference, NS-LRA reports $\text{E P}=\text{0. 9 2 4}$ and $\text{T C}=\text{0. 9 5}$ against the gold graph. These results indicate that NS-LRA can provide an efficient and traceable pipeline for proactive risk governance in LCPS.
Driver drowsiness is a persistent road -safety problem whose episodic and under -reported nature complicates both prevention and measurement. This paper presents AI Safety Guard, a low -cost edge- AI prototype that combines non -contact facial-landmark analysis with bounded auditory and optional olfactory alerts. The proposed artefact uses local camera processing to estimate sustained eye closure, mouth opening and yawn patterns, and head -pose deviation; temporal decision fusion then triggers an active warning through a speaker or buzzer and, when enabled, a short, atomised scent pulse. Unlike cloud-dependent monitoring, the prototype is designed to retain no video and to record only minimal local event information. The study adopts a design -science and safety -by-design methodology: it reconstructs system requirements, specifies the hardware and inference architecture, formalises the tri- channel decision logic, and evaluates the credibility and limits of preliminary prototype evidence. Project documentation reports operation on Raspberry Pi-class hardware at approximately 10-15 frames per second, local event logging, hard -coded ac tuator duration, cooldown lockout, manual acknowledgement, and a scent opt-out. A website event trace reports 116 ms from a detection event to alert activation, whereas a separate pitch document claims 0.001 s actuation latency; this discrepancy is treated as an unresolved measurement issue rather than evidence of validated performance. The paper therefore distinguishes artefact feasibility from safety efficacy. It proposes a five -phase validation programme covering bench metrology, public -dataset evaluation, simulator experiments, closed -track trials, and regulatory readiness against functional-safety, safety-of-the-intended-functionality, privacy, and human -machine-interface requirements. The principal con tribution is an evidence -bounded blueprint for translating a student -developed prototype into a testable driver -monitoring system while preserving privacy and explicitly managing intervention risk. The system is not positioned as a substitute for sleep, rest, or safe pull-over behaviour, but as a supplementary warning device requiring independent validation before road deployment.
Using SAFERELBENCH to evaluate seven open- and closed-source VLM-driven embodied agents, this work finds a substantial gap between task success and process-level safety compliance, and shows that safe embodied intelligence requires not only stronger perception and planning, but also reliable reasoning about how object relations shape risk during interaction.
