Indoor localization of Unmanned Vehicles is critical for tasks requiring low latency and low power consumption. Channel State Information (CSI) systems provide a promising approach by capturing fine-grained signal features, but often suffer from limited accuracy in complex environments or high computational cost. To address this, we propose MambaLoc, an indoor localization framework that balances efficiency and precision by leveraging Mamba architectures with cross-modal knowledge distillation. In offline training, a Transformer based on Ultra-Wideband (UWB) data acts as a teacher to guide the CSI-based Mamba student, improving robustness to multipath effects without relying on UWB during inference. The distillation process is further enhanced using a Gaussian Mixture Model (GMM) in both teacher and student, enabling effective transfer of UWB-derived spatial representations. Evaluations on real-world environments show that MambaLoc reduces MAE by 22.53% over a baseline Mamba, while cutting FLOPs by 65.1% and inference time by 58% compared to a Transformer model, making it a practical and deployable solution for UV localization.

Despite significant advances in autonomous driving, cross-city deployment remains challenging due to variations in traffic patterns, urban driving environments, and local regulations that require expensive retraining for each new location. Existing end-to-end driving systems trained on specific cities exhibit poor generalization when deployed elsewhere, limiting scalability. In this paper, we propose GeoDrive, a meta-learning framework that enables rapid adaptation to new cities with minimal local data. We combine Model-Agnostic Meta-Learning (MAML) with spatial context from OpenStreetMap to learn generalized driving policies that rapidly adapt to city-specific traffic dynamics. GeoDrive integrates a Large Language Model (LLM) module for interpretable decision-making that incorporates region-specific regulations and behavioral patterns. Experimental results show 18% and 32% improvements in trajectory prediction for intra-city and cross-city settings, respectively.

Modern mobility platforms have stored vast streams of GPS trajectories, temporal metadata, free-form textual notes, and other unstructured data. Privacy statutes such as the GDPR require that any individual's contribution be unlearned on demand, yet retraining deep models from scratch for every request is untenable. We introduce MobText-SISA, a scalable machine-unlearning framework that extends Sharded, Isolated, Sliced, and Aggregated (SISA) training to heterogeneous spatio-temporal data. MobText-SISA first embeds each trip's numerical and linguistic features into a shared latent space, then employs similarity-aware clustering to distribute samples across shards so that future deletions touch only a single constituent model while preserving inter-shard diversity. Each shard is trained incrementally; at inference time, constituent predictions are aggregated to yield the output. Deletion requests trigger retraining solely of the affected shard from its last valid checkpoint, guaranteeing exact unlearning. Experiments on a ten-month real-world mobility log demonstrate that MobText-SISA (i) sustains baseline predictive accuracy, and (ii) consistently outperforms random sharding in both error and convergence speed. These results establish MobText-SISA as a practical foundation for privacy-compliant analytics on multimodal mobility data at urban scale.

Schistosomiasis remains a significant public health challenge in Africa, where it is considered endemic, particularly in areas with inadequate access to clean water. This study focuses on mitigating the spread of schistosomiasis by monitoring the physicochemical parameters of water sources to detect the proliferation of infected snails, which serve as the intermediate host for the parasite. We propose a system that leverages sensing technology to monitor water physicochemical parameters, e.g., temperature (Temp), pH, and electrical conductivity (EC) and uses AI to predict their future values, alongside the dynamics of the infected snail population, which are crucial for schistosomiasis transmission. By forecasting these trends, the system facilitates proactive interventions, such as water treatment and sanitation improvements. To accomplish this, we leverage FNet, a model that effectively captures temporal features through discrete Fourier transforms (DFT), delivering Transformer-level accuracy while being more resource-efficient. While FNet is originally tailored for language processing tasks, we adapt it for time series analysis by employing a quantization technique. This approach is particularly advantageous in contexts where computational and operational efficiency is critical, making it well-suited for environments with limited infrastructure or financial flexibility. Our evaluation demonstrates that the FNet model outperforms transformers in terms of resource efficiency while maintaining similar performance, making it a promising tool for real-time monitoring and early detection of schistosomiasis risk. The results highlight the potential of this approach to enhance disease control efforts and contribute to improved public health outcomes in endemic regions.

This doctoral research proposes a framework for developing sociallyaware robot navigation systems by integrating the cognitive capabilities of Large Language Models (LLMs) with the demands of real-world Human-Robot Interaction (HRI).Our work follows a four-stage plan that systematically addresses the challenges of applying LLMs to time-sensitive, safety-critical tasks.This paper details the completion of the first two stages, wherein we developed and evaluated a foundational navigation model.Our system features a meticulously designed multimodal fusion pipeline that integrates LiDAR and camera data, processed by a YOLO model and a Hungarian algorithm for semantic association, providing rich, contextual input to the LLM.Through knowledge distillation and fine-tuning on data from a custom simulator, our model demonstrates robust spatial reasoning and superior performance in low-frequency decision-making scenarios compared to traditional reinforcement learning methods.We successfully validated this foundational model and identified its inference latency as a key challenge.These results establish a solid basis for our future work.This includes developing a "brain-cerebellum" hybrid architecture for real-time performance and exploring multi-robot social compliance.This research contributes to HRI by creating more predictable and trustworthy robots, and to the LLM field by investigating the symbol grounding problem through embodied intelligence.

Due to global climate change, developing robust and accurate disaster prediction systems has become essential worldwide, as severe disasters are increasingly occurring even in previously safe regions.Two key technical challenges in disaster prediction are data scarcity and regional heterogeneity.Disasters are rare and extreme, resulting in limited data, which hinders data-driven approaches like machine learning (ML).Moreover, hydro-meteorological disasters depend on region-specific factors such as rainfall patterns, soil, vegetation, and infrastructure.These, combined with dynamic environmental changes from earthquakes, climate, and urbanization, complicate the generalization of models across time and space.We propose a self-adaptive physics-hybrid disaster prediction system that autonomously adapts to diverse and evolving environments to address this.Our system consists of: (1) a physics-hybrid disaster prediction model integrating physics-based hydro-meteorological simulation and multi-modal ML models, and (2) an environmental adaptation utilizing test-time adaptation (TTA) to update parameters without labeled disaster data(i.e., disaster event data).Our physics-hybrid model is anchored in the physics model to balance stability and adaptability, leveraging strengths from both physics and ML models.Furthermore, the strong inductive bias of the physics model mitigates overfitting and catastrophic forgetting during TTA, enabling robust adaptation to unseen regions.

Electric scooters (e-scooters) are reshaping urban mobility, providing eco-friendly and cost-effective transport.However, the surge in their usage in mixed-traffic environments poses significant safety risks, especially for vulnerable road users (VRUs).This paper presents a lightweight vision-based safety assistance system optimized for real-time inference on edge devices.The core modules include semantic segmentation enhanced by semi-supervised learning, accurate bird's-eye view (BEV) transformation, real-time motion prediction, and adaptive path planning.Extensive evaluations demonstrate high segmentation accuracy and low-latency execution suitable for resource-constrained hardware.Future research directions include knowledge distillation for model adaptability, cooperative multi-scooter perception, intersection-based AI infrastructure, and the application of large language models (LLMs) to optimize urban traffic flows.

Accurate monitoring of indoor environments via wireless sensing is essential for applications such as adaptive crowd flow control, energy management in smart buildings, and rapid response in healthcare and emergency scenarios.Traditional vision-based methods suffer from lighting issues, occlusion, and privacy concerns, while RSSI-based approaches are vulnerable to multipath fading and device inconsistencies, and CSI-based systems lack interoperability.We present a device-free and device-independent framework that continuously records round-trip signal delays as a rich time series, supporting both occupancy estimation and spatial localization without specialized hardware.A deep feature extraction strategy suppresses environmental noise and non-line-of-sight distortions, delivering accurate real-time performance on commodity WiFi equipment even in cluttered spaces.This approach establishes a foundation for ubiquitous privacy-preserving indoor environment monitoring with potential extensions to multimodal sensor fusion and integration with next-generation wireless standards.

Machine learning and deep learning are becoming more popular in the area of indoor localization leveraging the Wi-Fi Received signal strength demonstrating high performance and scalability. However, such models usually perform poorly in noisy environments, leading to reduced trust in their reliability. In this work, we propose a novel framework that enhances trust in AI-based localization by using explainable AI techniques for automatic real time auditing of model behavior. Our approach detects and corrects any unusual behavior on-the-fly. Experiments in a testbed and two challenging real-world settings demonstrate that our framework effectively stabilizes performance and maintains trust in the localization models.

WiFi-based indoor localization is essential but often hampered by its dependence on specific environments and network configurations. We propose a novel solution using Language Models, like BERT, to overcome these limitations. Our approach transforms WiFi RSSI data into a language-like format, enabling the system to generalize across different buildings and AP setups. Key to our method is the Virtual Space Mapper, which standardizes user and Access Point (AP) locations into a common coordinate system, and then the system adapts BERT to predict locations within this virtual space. This design allows for robust, accurate localization without the need for extensive recalibration. Our results show significant improvements in adaptability and accuracy, making our approach a scalable, flexible alternative to traditional WiFi localization methods.

Indoor localization is critical for enabling a wide range of location-based services such as navigation, security, and contextual computing in complex indoor environments. Despite significant advances, the deployment of indoor localization systems in real-world settings remains limited due to challenges posed by non-independent and identically distributed (non-IID) data and device heterogeneity. In this paper, we propose SimDeep, a novel Federated Learning (FL) framework designed to tackle these challenges. SimDeep introduces a Similarity Aggregation Strategy to aggregate model updates based on client similarity, thereby effectively addressing the non-IID issue. Experimental results demonstrate that SimDeep achieves 92.89% accuracy, outperforming traditional federated and centralized techniques, making it a promising solution for practical deployment.

Accurate taxi-demand prediction is essential for optimizing taxi operations and enhancing urban transportation services. However, using customers' data in these systems raises significant privacy and security concerns. Traditional federated learning addresses some privacy issues by enabling model training without direct data exchange but often struggles with accuracy due to varying data distributions across different regions or service providers. In this paper, we propose CC-Net: a novel approach using collaborative learning enhanced with contrastive learning for taxi-demand prediction. Our method ensures high performance by enabling multiple parties to collaboratively train a demand-prediction model through hierarchical federated learning. In this approach, similar parties are clustered together, and federated learning is applied within each cluster. The similarity is defined without data exchange, ensuring privacy and security. We evaluated our approach using real-world data from five taxi service providers in Japan over fourteen months. The results demonstrate that CC-Net maintains the privacy of customers' data while improving prediction accuracy by at least 2.2% compared to existing techniques.

Human Activity Recognition (HAR) systems hold great potential in aiding disabled and elderly individuals to live independently. Various approaches have been suggested for identifying human activities, including sensors, cameras, wearables, and contactless microwave sensing. This latter method is gaining significant attention due to its ability to address privacy concerns arising from cameras and alleviate discomfort caused by wearables. However, current microwave sensing techniques have a key limitation, requiring controlled and ideal conditions to achieve accurate activity detection. In this paper, we propose RISense, a deep learning-aided system for HAR using Re-configurable Intelligent Surface (RIS). RISense introduces novel modules designed to generate a human activity representational space that ensures separability between activity classes, even in the presence of noisy and distorted Channel State Information (CSI) measurements. These representations are then fed into a Recurrent Neural Network (RNN), which learns the sequential changes in features to accurately estimate the user activity. The evaluation of RISense across two realistic settings, encompassing both non-line-of-sight and multi-floor scenarios, showcases its efficacy. Specifically, RISense achieves an activity recognition accuracy of 98.7%. This performance overcomes the accuracy of state-of-the-art systems by over 15%. These findings highlight the superior efficacy of the proposed methodologies, including integrating RIS and advanced learning techniques.

Indoor localization has become a critical area of research with increasing relevance in applications. While numerous technologies have been explored, WiFi-based fingerprinting solutions using Received Signal Strength Indicators from multiple access points have garnered substantial attention due to the ubiquity of WiFi networks. However, these methods often encounter challenges, including fluctuation of access points and their noise signal measurements. Furthermore, they struggle to adapt effectively to cluttered or dynamically changing environments. In this paper, we introduce GraphLy: a Graph Neural Network-based model explicitly designed to tackle these challenges. GraphLy captures complex spatial relationships between different locations and adapts to environmental complexities and clutter, offering a robust solution for indoor localization. Our experiments demonstrate that GraphLy outperforms state-of-the-art WiFi-based localization techniques in two cluttered and propagation complex testbeds. In particular, we achieved a performance improvement of at least 37% and 65% in the two environments, respectively. These findings underscore the potential of GraphLy to enhance indoor localization accuracy and reliability for various real-world applications.

WiFi fingerprinting systems for indoor localization have improved drastically with the advent of deep learning. However, these systems are often designed to be used on the same testbeds they are trained on, making generalizations to unknown testbeds hard. Additionally, any changes to Access Point configurations can drastically impact system performance. These systems often require collecting new data, retraining, and/or fine-tuning, which are time-consuming and costly. To address these issues, we propose a novel localization framework that can adapt to varying environments without recalibration. This is achieved by utilizing a "virtual space" via Graph Neural Networks, enhancing adaptability and system performance.

The growing demand for ride-hailing services has led to an increasing need for accurate taxi demand prediction. Existing systems are limited to specific regions, lacking generalizability to unseen areas. This paper presents a novel taxi demand forecasting system that leverages a graph neural network to capture spatial dependencies and patterns in urban environments. Additionally, the proposed system employs a region-neutral approach, enabling it to train a model that can be applied to any region, including unseen regions. To achieve this, the framework incorporates the power of Variational Autoencoder to disentangle the input features into region-specific and region-neutral components. The region-neutral features facilitate cross-region taxi demand predictions, allowing the model to generalize well across different urban areas. Experimental results demonstrate the effectiveness of the proposed system in accurately forecasting taxi demand, even in previously unobserved regions, thus showcasing its potential for optimizing taxi services and improving transportation efficiency on a broader scale.

Indoor localization systems have become increasingly important in a wide range of applications, including industry, security, logistics, and emergency services. However, the growing demand for accurate localization has heightened concerns over privacy, as many localization systems rely on active signals that can be misused by an adversary to track users’ movements or manipulate their measurements. This paper presents PassiFi, a novel passive Wi-Fi time-based indoor localization system that effectively balances accuracy and privacy. PassiFi uses a passive WiFi Time Difference of Arrival (TDoA) approach that ensures users’ privacy and safeguards the integrity of their measurement data while still achieving high accuracy. The system adopts a fingerprinting approach to address multi-path and non-line-of-sight problems and utilizes deep neural networks to learn the complex relationship between TDoA and location. Evaluation in real-world testbed demonstrates PassiFi’s exceptional performance, surpassing traditional multilateration by 128%, achieving sub-meter accuracy on par with state-of-the-art active measurement systems, all while preserving privacy.

The growing demand for ride-hailing services has led to an increasing need for accurate taxi demand prediction. However, the use of real passenger data to train predictive models raises serious privacy concerns. To address this challenge, we present a privacy-preserving taxi demand prediction system that employs a generative model to synthesize synthetic trajectory data, preserving privacy while retaining the statistical properties of the original data. The system also overcomes the challenge of location dependence of latitude-longitude values by encoding the representation into region-independent space, making it more general and applicable to different geographical areas. The system was evaluated on real-world data collected from a major taxi service provider in Japan over a period of six months. The results showed that the system can effectively defend against 98% of all attempted attacks on passenger data and against 60% of state-of-the-art attacks on the learning-based prediction models. Additionally, the proposed system ensures the prediction performance, with a barely noticeable decrease of 2.9% compared to using the original data.

In this demonstration paper, we present STM: a new system for securing and management of vehicle trajectory data using a generative model that balances privacy and utility. For instance, traditional methods for taxi-demand prediction pose the risk of privacy breaches from both the data and the model. To address this challenge, we deploy Spatiotemporal-GAN to generate synthetic trajectories that meet privacy regulations such as GDPR. We assess the quality of the generated data by constructing several taxi-demand prediction models. Moreover, we evaluate the privacy risk by implementing trajectory user linking attacks against the generated data and membership inference attacks against the prediction model. Our system is designed with rich interactivity and visualization, enabling the audience to use these modules. Overall, our approach demonstrates the potential of generative models in preserving privacy while maintaining data utility in the context of taxi-demand prediction.

Taxi demand prediction is an essential process for enabling efficient taxi operations and customer satisfaction. However, most existing solutions are vulnerable to leakage of passengers' privacy or membership inference attack. In this study, we propose a privacy-preserving taxi-demand system built without real customer data. Specifically, we employ LSTM-GANs to generate synthetic trajectories reflecting the typical Spatio-temporal semantics of the original data without privacy leakage. The system evaluation was held on a real taxi service provider in Japan for six months. The results show the ability of the system to keep privacy with 86% of the cases with barely a negligible decrease in the prediction performance compared to using the original data.

We propose SoLoc, a lightweight probabilistic fingerprinting-based technique for energy-free device-free indoor localization. The system harnesses photovoltaic currents harvested by the photovoltaic cells in smart environments for simultaneously powering digital devices and user positioning. The basic principle is that the location of the human interferes with the lighting received by the photovoltaic cells, thus producing a location fingerprint on the generated photocurrents. To ensure resilience to noisy measurements, SoLoc constructs probability distributions as a photovoltaic fingerprint at each location. Then, we employ a probabilistic graphical model for estimating the user location in the continuous space. Results show that SoLoc can localize the user at sub-meter accuracy in a real indoor environment.

Recently, commercially available off-the-shelf devices support the IEEE 802.11-2016 standard that implements a fine-time measurement (FTM) protocol for measuring the WiFi signal round trip time (RTT). This opens the door for a practical time-based localization where the clock synchronization problem between the transmitter and the receiver no longer exists. However, time-based techniques are highly affected by multipath propagation and non-line-of-sight transmissions. To combat these challenges, we present MagttLoc, a hybrid deep-learning-based indoor localization system that fuses magnetic measurements with WiFi RTT. MagttLoc yields a stable localization performance of decimeter-level accuracy improving upon the state-of-the-art systems in all experiments.

Indoor localization is a key component of pervasive and mobile computing. Due to the widespread use of WiFi technology, WiFi fingerprinting is one of the most widely utilized approaches for indoor localization. Despite advancements in WiFi-based positioning approaches, existing solutions necessitate a dense deployment of access points, time-consuming manual fingerprinting, and/or special hardware.

Location services are one of the most used applications today on mobile devices. The vast majority of localization systems propose solutions for locating the user in a 2D single floor environment. However, accurate estimation of the user's floor level, in tall multistory buildings, is a crucial basis for many applications, especially for emergency services.

Location-based services have undergone massive improvements over the last decade. Despite intense efforts in industry and academia, a pervasive infrastructure-free localization is still elusive. Towards making this possible, cellular-based systems have recently been proposed due to the wide-spread availability of the cellular networks and their support by commodity cellphones. However, these systems only consider locating the user in a 2D single floor environment, which reduces their value when used in multi-story buildings.

Robust and accurate indoor localization has been the goal of several research efforts over the past decade. Towards achieving this goal, WiFi fingerprinting-based indoor localization systems have been proposed. However, fingerprinting involves significant effort; especially when done at high density; and needs to be repeated with any change in the deployment area. While a number of recent systems have been introduced to reduce the calibration effort, these still trade overhead with accuracy.

The demand for a ubiquitous and accurate indoor localization service is continuously growing. Despite the pervasive nature of cellular-based solutions, their localization quality depends on the number of cell towers provided by the phone. According to the standard, any cell phone can receive signal strength information from up to seven cell towers. However, the majority of cell phones usually return only the associated cell tower information, significantly limiting the amount of information available to the location determination algorithm. In this paper, we present SoloCell: a novel deep learning-based indoor localization system that utilizes the signal strength history from only the associated cell tower to achieve a fine-grained localization. SoloCell incorporates different modules that lessen the data collection effort and improve the deep model's robustness against noise. Evaluation using different Android phones shows that SoloCell can track the user with a median localization error of 0.95m This accuracy demonstrates the superiority of SoloCell compared to the state-of-the-art systems by at least 210%.

The demand for a ubiquitous and accurate indoor localization service is continuously growing. Despite the pervasive nature of cellular-based solutions, their localization quality depends on the number of cell towers provided by the phone, which is typically limited. Specifically, according to the standard, any cell phone can receive signal strength information from up to seven cell towers. However, the majority of cell phones usually return only the associated cell tower information, significantly limiting the amount of information available to the location determination algorithm, degrading its performance.