Engineering

Abstract: In light of strong expansion of renewables, electrolysis may act as an alternative to conventional grid enforcement to overcome grid constraints in a timely and effective manner as it creates additional flexible load and thus enables renewable production peaks to be absorbed. This paper examines whether hydrogen electrolysis can serve as a cost-effective alternative to conventional grid reinforcement in Austrian electricity distribution networks. It goes beyond the current state of art by using three real-world case studies in Styria, using an integrated techno-economic framework combining distribution-grid simulations, market optimization, PEM electrolysis modelling and cost-benefit analysis to compare the grid-supportive electrolysis against conventional grid enforcement. The results demonstrate that grid-supportive electrolysis can become competitive under suitable hydrogen market conditions. When operated in grid-supportive mode only, the capacity factor for the electrolysis lies below 5%. Economic viability emerges only when electrolysers are allowed to combine grid-supportive operation with market-driven hydrogen production. Thereby, the hydrogen price proves to be the key determinant: a hydrogen price above approximately 6 EUR/kg incentivizes market based operation and naturally resolves grid congestion without further intervention by the DSO. In some locations, smaller electrolysers (around 20–60% of the theoretically required size) deliver the best economic performance, recovering most curtailed renewable energy while limiting investment costs compared to conventional grid extension.

Abstract: Digital technology has revolutionized the communication between suppliers and buyers in logistics and transportation. Previous studies focused on individual technolo-gies and their benefits but have not sufficiently explored how these technologies interact together as a communication interface, and what specific data they exchange. The review conducts a systematic literature analysis of 51 peer reviewed papers (2016-2026), screened from 711 records from Scopus and Web of Sciences. The paper provides three contributions, first it identifies digital communication mechanism that enables supplier buyer coordination with documented outcomes from in-cluding reducing freight negotiations from weeks to hours and warehouse waiting time by 80%. Second the review consolidates exchange data parameters into the Digital Logistics Communication Architecture (DLCA), a four-layer taxonomy offering empirically grounded data specification. Moreover, third it finds that digital communication contrib-utes to sustainability primarily through environmental and economic outcomes, with so-cial outcomes remaining underrepresented. The most significant finding is structural; across most 51 peer reviewed papers it is observed that each digital communication system depends on human initiation. Building on this structural gap, the study proposes agentic AI as a Tier 5 extension of the DLCA framework, positioned as the technology most capable of overcoming human-initiation dependence. A future research agenda is outlined.

Abstract: Peripheral neuropathy from diabetic, oncological, or traumatic etiologies degrades gait mechanics, increasing peak plantar pressures (PPP) and tissue ulceration risks. This feasibility study evaluates a customizable fabrication protocol using thermo-mechanically synthesized re-entrant auxetic foam insoles (sizes 6–11) with targeted dome-shaped inserts. Bilateral dynamic gait analysis was conducted across a heterogeneous neuropathic cohort (N=9; 5 females, 4 males) utilizing a P-Walk 600 pressure plate and a MARVUE 2D motion capture system. To control structural shoe variance across testing conditions, all subjects were equipped with standardized footwear featuring pre-installed conventional memory foam. To overcome small-sample parametric limits and evaluate structural predictability, a non-parametric bootstrap resampling analysis (B=1,000) was executed. The native auxetic foam demonstrated mechanical reliability, yielding a tightly bound distribution with a highly constrained bootstrap standard deviation (σ*) of 12.063 kPa and a baseline Consistency Rank of 1—indicating a 100% statistical probability of outperforming traditional controls. While a structural break-in period was required for re-entrant cell adaptation, custom insoles ultimately reduced PPP by up to 62.2% in systemic polyneuropathies, shifting the loading signature beneath the clinical safety target threshold (<200 kPa) in 55% of cases. Conversely, customization triggered an adverse volumetric crowding effect in focal mononeuropathies, identifying un-customized native foam as the superior intervention for localized trauma by preventing premature cellular self-contact. Re-entrant auxetic structures reliably accommodate heterogeneous pathomechanics, offering a mathematically verified pathway for advanced patient-specific orthotic interventions.

Abstract: ANSI/IES TM-30-24 designates CES #15 and CES #18 as skin-representative color-evaluation samples, summarized by their mean individual fidelity (R_f,CES15,18). This analysis tested how well that two-sample mean represents average fidelity and category-level variation across a 100-spectrum skin-reflectance library stratified into five lightness-defined categories, evaluated across 1,432 architectural LED sources from the DTU PhotoLED database. R_f,CES15,18 closely tracked mean fidelity across the full reference library (r = 0.9976; RMSE = 0.84 Rf points). Benchmarked against all 4,950 possible two-spectrum means from the same library, it exceeded 85.7% of pairs in correlation and 82.7% in RMSE, favorable performance that was not unique. Raw category means within individual sources differed by 4.74 R_f points on average and by as much as 10.66 points. This ordering tracked chroma and was largely captured by the lightness-chroma geometry of the sampled spectra; little category-specific structure remained once that geometry was modeled, and the ordering was not fixed across the two reflectance libraries. Compact subsets were tested under source-only holdout and under simultaneous source-and-skin-spectrum holdout. Greedy selection held a modest source-held-out error advantage over an optimized category-matched random search at 9 to 10 samples. Under simultaneous holdout, subsets of that size preserved source-rank recovery, while reconstruction of unseen category means carried median errors near 1 R_f point. CES #15/#18 therefore serves as a favorable aggregate indicator within this library. Characterizing variation across the sampled skin-reflectance locus requires broader reflectance sets.

Abstract: Automation of semantic segmentation of binary detection of focal liver lesions (FLLs) in triphasic computed tomography (CT) scans is critical for knowledge extraction of hepatic malignancy, staging a disease, and planning a treatment. But standard models struggle with class imbalances, boundary differences, and tissue heterogeneity in different phases. This study proposes a modified architecture of a parallel dual-encoder network for FLL boundary refinement optimization. The proposed architecture implements an early feature fusion by stacking the non-contrast (NC), arterial (ART), and portal venous (PVP) phases in the channel dimension. Alongside a per scale, lightweight fusion strategy integrating localized convolutional details to global transformer context. The two architectures were modified and optimized: VGG-19 paired with a Swin Transformer and ConvNeXT-Small paired with a Cross-Shape Window (CSwin) Transformer. The study used a patient-grouped stratified 3-fold validation on 517-case MCT-LTDiag dataset. The networks were extensively benchmarked against standard U-Net, DECTNet and nnU-Net baselines across multiple performance metrics: generalization, segmentation, boundary, and computational. Statistical significance was precisely tested by using a two-tailed paired t-test with Benjamini-Hochberg False Discovery Rate (BH-FDR) adjustments. The ConvNeXT-Small + CSwin Transformer configurations turned out to be the best architecture, with an elite average Dice score of 0.9091 and a leading Intersection over Union (IoU) of 0.8924. This model reduced boundary errors and achieved the lowest 95th percentile Hausdorff Distance (HD95) of 13.86. It resolved the precision-sensitivity trade-off plaguing the baseline models by maintaining a leading sensitivity of 0.9484 and a precision of 0.9352. The champion architecture sustained a tight footprint of ~29M parameters and achieved an optimized average inference speed of 0.111 seconds per slice, even though it has an advanced attention mechanism. This cut the process delay of the VGG-19 variant in half. Qualitative and quantitative results confirm the synergy of modern localized depthwise convolutions and global cross-shaped attention mechanisms removed the bloated false-positive masks and broken under-segmentation. It is delivering a highly stable, precise, and reproducible tool for automated clinical workflows.

Abstract: Modern hotel management increasingly depends on data-driven decision-making based on Key Performance Indicators (KPIs), yet many hotels face significant challenges: operational data arrives in heterogeneous CSV formats with inconsistent structures, manual KPI calculation is time-consuming and error-prone, and staff training in KPI interpretation requires substantial resources. This paper presents an intelligent agent that integrates generative AI with automated data processing to address these challenges in a unified conversational system. The system leverages Google Gemini 2.0/2.5 Flash for natural language understanding and code generation, integrated with a Python-based architecture (FastAPI backend, Chainlit conversational interface). Core capabilities include: (1) automated CSV preprocessing with format detection, missing value imputation using K-Nearest Neighbors, and temporal variable normalization; (2) AI-driven KPI suggestion and automated calculation through secure code generation with multi-tier validation and sandboxed execution; (3) interactive HTML dashboard creation using Plotly with temporal organization and year-over-year comparisons; (4) contextual training delivery with adaptive question generation across multiple formats (true/false, multiple-choice, open-ended, situational); and (5) automated assessment with personalized feedback. Experimental validation at Technology Readiness Level 4 (TRL 4) demonstrates system feasibility in controlled laboratory conditions. Testing with 30 datasets (12 real hotel data, 18 synthetic) achieved 91.5% success rate across 200 functional test cases, including 100% success in CSV preprocessing, 87.5% first-attempt code generation success, 95% dashboard generation accuracy, and strong correlation (r=0.82) between automated and expert scoring for open-ended questions. Performance benchmarking shows acceptable response times (median 11.5s for KPI calculation) and resource utilization suitable for moderate-scale deployment. These results establish technical feasibility for advancing to TRL 5 validation in operational hotel environments, demonstrating that generative AI can effectively automate hospitality analytics while maintaining security and reliability standards.

Abstract: Reliable tactile object‑shape recognition on robotic hands is often achieved using dense sensor arrays or vision‑based tactile skins, which increase fabrication complexity and computational cost. This work demonstrates that high recognition performance can instead be achieved through principled sparse sensing. A minimal multimodal tactile system is developed by fusing soft capacitive stretch sensors at finger PIP joints with a compact six‑element palmar pressure array, integrated on a human‑like hand mechanically constrained to emulate robotic grasping. Using an ANOVA‑based channel selection, low‑informative metacarpophalangeal signals are identified and removed, reducing the sensing configuration from 16 to 11 channels while improving classification accuracy. A lightweight multilayer perceptron operating on this low‑dimensional input achieves 95.4% size‑invariant recognition accuracy across 12 rigid objects representing four geometric primitives: cuboid, sphere, cylinder, and cone, outperforming the denser baseline. Ablation studies confirm the complementary roles of finger‑joint deformation, which encodes curvature cues, and palmar force distribution, which captures contact topology; neither modality alone achieves comparable performance. Beyond accuracy, the proposed design reduces sensor count, wiring, and computational requirements, enabling embedded‑ready deployment. The results show that data‑driven sensor placement, rather than sensor density, can yield robust grasp‑based shape recognition, offering practical guidance for tactile perception in resource‑constrained robotic hands.

Abstract: In this paper, we present an optimization framework for green hydrogen (GH) production integrating photovoltaic generation, reverse-osmosis desalination, proton exchange membrane electrolysis, and battery energy storage for continuous operation under solar intermittency. The study introduces a two-step ordinal optimization (OO) method to explore efficiently the large design space and identify subsystem sizes that minimize the levelized cost of hydrogen ($/kg), including production, storage, and transportation, at an average daily output of 60 tons of GH per day. First, the designs are evaluated using a simple, but computationally efficient model based on a two-week simulation. The evaluated designs are then scaled to a yearly operation and ranked by increasing hydrogen costs. Second, the top-S designs are re-evaluated using an accurate annual simulation model. OO theory predicts the number of top-S designs that need to be evaluated accurately to ensure that the optimum is included with a 95% alignment probability. We applied this framework to case studies for producing GH in Tunis and shipping it to Genoa in Italy and Hamburg in Germany, at costs of $3.91 and $6.40 per kg, respectively. The study leverages the potential of renewable energy (RE) production in Tunis and its proximity to Europe.

Abstract: Classical machine learning (ML) models, including Logistic Regression (LR), Support Vector Machines (SVM), and Random Forests (RF), remain widely used in practical applications due to their efficiency, interpretability, and low computational costs. However, their security properties against different adversarial threats are often evaluated independently rather than within a unified comparative framework. This paper presents a unified empirical evaluation of LR, Linear SVM, and RF models across image (MNIST and CIFAR-10) and text (AG News) classification domains. The models are analyzed under three adversarial scenarios: training-time poisoning, inference-time evasion, and black-box model extraction attacks. The empirical results reveal architecture-dependent security trade-offs: while RF showed higher resistance against random label noise but remained vulnerable to targeted poisoning strategies. In contrast, linear models (LR and Linear SVM) exhibit higher susceptibility to black-box extraction, achieving replication fidelities up to 96.22% under active querying strategies, while showing more predictable performance degradation under the evaluated evasion settings. Furthermore, the study identifies an optimization-related effect in which gradient limitations may contribute to Linear SVM exhibiting higher apparent resistance against iterative attacks compared with single-step perturbations. The evaluation shows that model robustness depends on both the attack type and the underlying architecture. Random Forest achieved lower degradation in some poisoning scenarios, while linear models showed different sensitivity patterns under targeted attacks.

Abstract: Fatigue is widely recognized as one of the primary failure mechanisms affecting metallic structures. Consequently, reliable prediction of fatigue-induced failure is essential for the safe design and lifetime prediction of engineering structures and components. In this context, the present work provides an overview of some of the main approaches employed for fatigue modelling in metals. First, the classical total-life methods, namely the stress–life and strain–life approaches, are reviewed. Subsequently, fracture-based approaches to fatigue are presented, including formulations grounded in fracture mechanics, continuum damage mechanics, and phase-field theory. The finite element method and the extended finite element method are also discussed as representative numerical frameworks for the implementation of these formulations. The reviewed approaches are then critically compared, and their main characteristics are summarized in a comparative table. Finally, the principal features of the presented approaches are highlighted.

Abstract: The Intuition Controlled Smart Glasses are a wearable device that uses embedded AI and non-invasive EEG signals to allow hands-free and voice-free control of smart environments. For on-device classification, the system combines EEG acquisition, preprocessing, feature extraction, and a small CNN. BLE and MQTT are used to send commands to external devices. The system can distinguish at least three mental commands with an accuracy of more than 80 percent, according to evaluation using publicly available EEG datasets (motor imagery, attention, and SSVEP).

Abstract: This paper argues that the combination of systems mapping and opportunity identification offers a coherent pedagogical frame to support sociotechnical practice. Opportunity identification is treated as the cognitive capacity to recognize leverage points for intervention in complex systems, while systems mapping tools provide the scaffolding through which students learn to navigate ambiguity, integrate non-technical knowledge, and locate those leverage points. We report empirical evidence from an undergraduate engineering elective in which students investigate a wicked problem of their choice using multiple systems mapping tools. The course emphasizes problem exploration over solution convergence and is explicitly designed to help students engage with communities, paradigms, and knowledge fields outside engineering. The paper draws on qualitative evidence from semi-structured post-course interviews with 13 students across three course cohorts. The interview protocol included a case-based section to assess students' ability to transfer systems mapping tools and opportunity-identification reasoning to an unfamiliar sociotechnical challenge. Interview data were analysed using a combined deductive--inductive coding approach. The study is anchored in a new framework that integrates three bodies of literature: Richmond's Systems Thinking process model, the KEEN Entrepreneurially Minded Learning framework, and Ardichvili et al.'s Opportunity Identification and Development model. The paper offers an empirically grounded account of how systems mapping pedagogy develops students' capacity to navigate ambiguity, integrate non-technical knowledge, and identify intervention possibilities in complex sociotechnical systems and offers practical design insights for educators seeking to build sociotechnical capacity in undergraduate engineering curricula.

Abstract: Light detection and ranging (LiDAR) sensors are widely known for their applications in 1 robotics, autonomous cars, remote sensing, and object tracking. These compact sensors are favored by 2 automation technologists for their accurate and long-range object detection capabilities. As a result, 3 a multitude of LiDAR sensors with diverse specifications have been introduced in the commercial 4 market. However, existing 3D LiDAR sensors lack the ability to customize their specifications, such 5 as measuring range, Field of View (FoV), angular resolution, number of scan points, and number of 6 scan layers, according to specific applications. This limitation poses several challenges for engineers, 7 including processing excessive data, requiring high computation power, demanding more storage 8 space for post-processing, and leading to false feature detection. To address these issues, this paper 9 presents a novel reconfigurable 3D LiDAR technology called 3D Customizable LiDAR (3D CS LiDAR), 10 which offers customization of sensor specifications based on the application requirements. The paper 11 discusses the mechanical and electrical systems of the developed LiDAR sensor and evaluates its 12 object detection performance by comparing it with a commercially available sensor. The results 13 demonstrate that the proposed reconfigurable 3D LiDAR system outperforms the commercial sensor 14 in all considered scenarios, indicating its potential for various perceptional applications. This research 15 contributes to the advancement of LiDAR technology by introducing a customizable approach, 16 addressing the limitations of existing sensors. The findings showcase the initial progress made 17 towards developing a comprehensive reconfigurable 3D LiDAR system, which holds promise for 18 diverse practical applications.

Abstract: Olive is a major agricultural crop extensively cultivated throughout the Mediterranean region. However, olive trees are vulnerable to several diseases that can negatively affect productivity and yield. One of the most widespread foliar diseases is olive leaf peacock spot, caused by the fungus Cycloconium oleaginum. Early detection of this disease is essential for preventing leaf drop, limiting disease spread, maintaining tree health, and reducing treatment costs before the infection reaches an advanced stage. In this study, a multimodal hybrid deep learning framework is developed to detect peacock spot disease in olive leaves and assess disease severity based on visual and numerical features. The proposed framework integrates olive leaf images with soil conditions, environmental conditions, and vegetation and stress indices to provide a more comprehensive disease analysis than image-only approaches. A ResNet50-based convolutional neural network is used to extract visual features from leaf images, while a multilayer perceptron processes the numerical sensor-based and index-based data. These features are then fused within a unified learning framework to classify disease stages and estimate leaf damage severity, including lesion coverage and yellowing percentage. The performance of the proposed model was evaluated using standard performance metrics suitable for both classification and regression tasks. For classification, the model was evaluated on 494 testing samples and achieved an overall accuracy of 97.77 %, with a macro F1-score of 0.9809 and a weighted F1-score of 0.9776. In addition, the model achieved low regression errors, with mean absolute errors of 1.16 % for lesion coverage and 1.42 % for yellowing estimation. These results demonstrate the effectiveness of the proposed multimodal framework for accurate peacock spot detection and severity assessment, supporting its potential use in smart agricultural monitoring and disease management.

Abstract: Recent progress in robot learning has relied on two investments: larger datasets and more capable models. Vision-language-action (VLA) policies now report success rates above 90% on standard benchmarks, yet perturbation studies show the same policies collapsing to near 0% when object positions, instructions, or scene layouts shift, exposing memorization where competence was claimed. This survey asks whether progress comes mainly from data, from models, or from an interaction that current evaluations often obscure. We review more than 200 papers spanning VLA architectures, world models, reinforcement-learning post-training, robot manipulation datasets, data generation pipelines, scaling studies, and perturbation benchmarks, including a structured analysis of a 100-paper survey set centered on the ICLR 2026 world-model literature. We catalogue every major public manipulation dataset with size, embodiment coverage, collection method, and known weaknesses; we reconstruct the evidence on data scaling laws and data quality; and we trace the evaluation crisis from benchmark inflation through memorization diagnoses to factor-level robustness decompositions. Our synthesis is that the question is ill-posed as a dichotomy: data diversity dominates in-distribution gains, model class and training objective dominate out-of-distribution retention, and current benchmarks confound the two because train and test conditions coincide. We state the conditions under which each answer holds, identify bottlenecks per subfield, and propose falsifiable research directions, including counterfactually structured datasets, world-model-regularized policies, and factor-controlled evaluation protocols.

Abstract: In oil production, the formation of oil emulsions due to reservoir water breakthrough is widely observed. Since the viscosity of these emulsions, which are considered polydisperse systems, can increase sharply depending on the degree of water cut, they create considerable difficulties in well-gathering systems and also increase hydraulic losses. The rheological properties of oil emulsions depend on the phase ratio, flow velocity, degree of dispersion, and numerous other parameters. There is no generalized model for the rheological description and determination of the properties of oil emulsions, which belong to anomalous and rheologically complex systems. Therefore, a diagnostic method for determining the viscosity of stable emulsions, taking into account the effect of increasing water content, is of great importance. In this article, the existing empirical expressions currently used for diagnosing the rheological properties of oil emulsions are examined. It has been determined that their application in oilfield practice is associated with certain difficulties and, in most cases, they are not considered suitable for solving engineering problems. In the article, a mathematical model has been developed, tested, and shown to provide good results for determining and predicting the viscosity of structurally stable oil emulsions depending on the degree of water cut.

Abstract: Professional UAV thermal imaging systems are widely used for inspection, monitoring, and emergency applications, but their cost limits their use in educational, preliminary, and low-resource scenarios. This study evaluates a low-cost indirect UAV thermal sensing workflow based on a DJI Mini 4K consumer drone, a Servo King900 smartphone, and a UTi260M smartphone-connected infrared camera. The smartphone displayed and recorded the thermal stream, while the UAV onboard RGB camera recorded the smartphone-displayed infrared video during flight. The system was tested under no-payload and payload conditions, daylight and nighttime illumination, and several low-altitude operating heights. Motor temperatures were additionally inspected using a UTi260T thermal camera. The complete UAV–payload configuration had a measured mass of approximately 340 g, corresponding to an effective payload of 91 g and a payload-to-UAV mass ratio of 36.5%. Payload operation reduced flight endurance from approximately 25 min to 14 min 40 s and produced increased and asymmetric motor heating. Nighttime operation provided better display readability than daylight operation, with the best usability observed at approximately 5–15 m. The proposed workflow is feasible for short-range preliminary thermal screening, but it is limited by payload mass, suspended-load oscillation, display readability, endurance reduction, motor loading, and the absence of raw radiometric data.

Abstract: Cowpea is a crop of great importance worldwide, which is why many heirloom varieties and improved cultivars are explored. Consuming pods and green beans provides vitamins, minerals, and functional components for people with limited access to vegetables. The pods and green beans of these materials have intrinsic characteristics that distinguish them. Therefore, the objective was to adjust machine learning models to identify cowpea from digital images of pods and green beans using artificial intelligence techniques. Digital images of four heirloom Creole of the cowpea genotypes (Sempre Verde, Rabú de tatu, Corujinha, and Paulistinha) and nine cultivars (BRS No-vaera, BRS Olhonegro, BRS Verdejante, BRS Exuberante, BRS Pajeú, BRS Miranda, IPA 206, BRS Tapaihum, and BRS Pingo de Ouro) were processed using four deep learning architectures for feature extraction (vectorization): InceptionV3, SqueezeNet, VGG16, and VGG19. Six machine learning algorithms were evaluated: K-Nearest Neighbors (KNN), Decision Tree, Random Forest (RF), Gradient Boosting (GB), Support Vector Machines (SVM), and Multi-Layer Perceptron (MLP). The MLP (Artificial Neural Network) and SVM models, particularly when integrated with the InceptionV3 embedder, demonstrated superior performance. For pod classification, these models achieved near-perfect performance, with Area Under the Curve (AUC) and Classification Accuracy (CA) of 1.000. For green beans, the MLP maintained high accuracy (CA = 0.977) and better probabilistic calibration (lower Log-Loss) than the SVM. Digital image-based identification associated with machine learning is an efficient, non-destructive approach for the morphological characterization and discrimination of cowpea genotypes, supporting high-throughput phenotyping (HTP) applications.

Abstract: This study investigates the impact of industry training programs and innovation competitions on microelectronics graduate employability in a private university located in China's Guangdong-Hong Kong-Macao Greater Bay Area. Using two-year cohort data (N=220) with cleaned, mutually exclusive labeling of competition roles, we apply Gradient Boosting Decision Tree (GBDT) with SHAP analysis to identify which factors most strongly predict job-major match. A cross-year validation (training on the larger, higher-participation Class of 2022, N=128; testing on Class of 2021, N=92) provides more reliable estimates. The ranking of the four features remained identical across the two cohorts, with SHAP values showing only slight differences, demonstrating the stability of the analytical approach. Results show that Camp A — an intensive, project‑based, inclusive training program — is the strongest predictor of job-major match (SHAP=0.1754), followed by competition awards (0.1255) and competition participation (0.0888), while Camp B — a short, tool‑oriented program — shows minimal contribution (0.0433). Notably, unlike national competitions that primarily attract top-performing students, Camp A intentionally recruits students with diverse strengths (technical, marketing, project management), making its employability benefits more broadly accessible. This cohort‑level advantage — serving a more diverse student population — explains why a well‑designed training program can have greater overall impact than competition participation. These findings suggest that well‑designed, inclusive industry training can have greater cohort‑level employability value than competition participation in resource‑constrained private universities.

Abstract: Visual impairment affects over 338 million people worldwide, creating a need for context-aware assistive technologies. Existing real-time scene-description systems remain stateless: frames are processed independently, with no memory of prior scenes, no distance awareness, and no reasoning over navigation actions. This paper presents My Eye AI v2, a mobile-cloud assistive vision system that extends prior v1 work with three new components: (1) a Temporal Memory Layer using Retrieval-Augmented Generation backed by ChromaDB; (2) a Depth Awareness Module based on Depth Anything V2 for vision-only metric distance estimation; and (3) an Agentic Decision Layer powered by Claude Haiku that reasons over detection, depth, and memory outputs to produce navigation guidance. We introduce the Temporal Threat Score (TTS), a composite urgency metric combining detection confidence, distance, and temporal persistence. An ablation across four configurations shows the full v2 system achieves 100% action accuracy on 40 labeled scenarios, versus 72.5% (depth-only) and 45% (memory-only). Memory recall reaches 100% at a cosine distance threshold of 0.40. Depth MAE is 0.314 m raw, reduced to 0.285 m after linear calibration. On GPU (RTX 5080), parallel inference averages 236 ms (23× CPU speedup), with full-system p50 of 1,353 ms and p95 of 3,637 ms. A user study with visually impaired participants is planned as future work.