Coordinating distributed mobile agents under dynamic spatial perturbations is challenging. In satellite formations, satellites must jointly adjust transmission parameters to maintain beamforming despite position and attitude disturbances. Existing methods rely on centralized optimization with repeated objective evaluations, resulting in high cost and runtime unsuitable for real-time use. We propose a surrogate-guided graph policy learning framework for satellite beamforming. A graph neural surrogate approximates the beamforming objective and captures intersatellite interactions, while a distributed graph policy is trained with surrogate-guided optimization and imitation regularization. The learned policy performs decentralized inference through localized message passing. Experiments on 4-100 satellites show that it matches or outperforms Simulated Annealing with 0.54-1.25 ms inference latency, while improving gain by up to 36.84 dB and directivity by 1.88 dB in large constellations.

Mobility systems are evolving into large-scale cyberphysical infrastructures that continuously sense, learn from, and act upon human movement. From indoor positioning and activity recognition to ride sharing and shared urban services, these systems increasingly shape how people interact with buildings, campuses, and cities. As intelligence becomes pervasive, trust in sensing, learning, and data use emerges as the central challenge for sustainable deployment. This talk presents a vision for trustworthy cyber-physical mobility systems, arguing that trust must be designed holistically across the entire pipeline rather than addressed at isolated stages. We begin at the sensing layer, highlighting privacy conscious modalities such as LiDAR and wireless sensing, which enable rich spatial and behavioral understanding without directly capturing identifiable visual information. This reflects a broader shift toward responsible sensing as a foundational design principle. We then examine learning and collaboration in distributed mobility ecosystems, where federated learning and GenAI enable collective intelligence while keeping raw data at the source. While promising, such approaches are not inherently safe, as threats such as membership inference attacks expose subtle privacy leakages even in decentralized settings. These challenges reveal a fundamental utility–privacy trade-off that future mobility systems must explicitly manage. Finally, the talk highlights transparency and visualization as key enablers for making cyber-physical intelligence observable and interpretable, outlining a path toward mobility systems that are intelligent, resilient, and worthy of long-term trust.

Global climate change has increased the frequency and intensity of extreme rainfall events, causing landslides even in regions with little or no historical disaster records. Accurately predicting such events in urban environments requires integrating heterogeneous environmental data within a system-level co-design framework. However, two key challenges hinder reliable prediction: (i) limited observability due to scarce disaster records and the difficulty of measuring essential physical variables such as subsurface hydrology, and (ii) strong regional heterogeneity, where disaster data exhibit inherently non-IID characteristics across regions. Traditional physics-based models often rely on simplified assumptions because observable variables are limited, whereas deep learning models tend to overfit under scarce and imbalanced data. To address these issues, this study proposes a physics-integrated deep learning system that combines a three-layer tank model as a physics-based model with a data-driven deep learning model. The physics-based outputs are incorporated both as input features and through a skip-connection that adaptively controls their contribution under abnormal rainfall conditions. In addition, Stochastic Feature Augmentation is employed to improve robustness against regional data distribution shifts. Experiments conducted across 19 regions in Japan demonstrate that the proposed system achieves improved generalization and superior PR-AUC in even unseen regions compared with state-of-the-art methods.

In this paper, we formulate a new vehicle dispatch optimization problem, called Nursing Care Taxi Dispatch, as a variant of the Vehicle Routing Problem, considering constraints related to wheelchair use, user compatibility, pick-up and drop-off times, and vehicle limitations. Previous neural-based methods for Vehicle Routing Problems have typically addressed a few simple constraints, while our new problem involves multiple complex constraints, resulting in having fewer destinations to select. This complexity makes it more difficult to obtain solutions that allow all nodes to be visited with a limited number of vehicles. To balance low violation rate, computational efficiency, and solution quality, we propose a supervised machine learning approach based on the Transformer architecture. We first obtain a set of high-quality solutions using an integer linear programming solver for given inputs and then train our learning model through supervised learning. Additionally, we introduce the post-processing of the paths generated by the learning model, ensuring that all constraints are satisfied. We compared each instance’s objective function value (operating time), execution time, and constraint violation rate across different methods: our proposed method and some existing methods including integer linear programming and machine learning-based methods, using real-world facility data. Our method successfully produced balanced solutions regarding operating time, execution time, and constraint violation rate. Notably, we observed a decrease in the operating time for all problem sizes and regions, while keeping constraint violations to a minimum compared to existing methods. Especially, the decrease reached up to 8% for problem sizes with fewer than 30 users.

Unmanned Aerial Vehicles (UAV)-mounted Reconfigurable Intelligent Surfaces (RIS) have emerged as a promising architecture for enhancing wireless coverage, spectral efficiency, and energy performance in 6G networks. By combining programmable electromagnetic wave manipulation with aerial mobility, UAV-RIS systems enable dynamic blockage mitigation, adaptive beamforming, and flexible deployment across terrestrial, maritime, and satellite-integrated environments. However, joint optimization of UAV trajectory, RIS phase configuration, and resource allocation incurs high computational complexity, which is incompatible with the strict energy and onboard processing constraints of UAV platforms. Lightweight AI techniques offer practical solutions to this challenge. Hence, this paper provides a comprehensive overview of lightweight AI techniques for UAV-mounted RIS systems, including Reinforcement Learning (RL), meta-learning, Federated Learning (FL), Multi-Armed Bandits (MAB), and energy-aware optimization. We present a detailed taxonomy and comparative analysis of existing work, highlight computational-energy trade-offs, and identify open research challenges for scalable, energy-efficient airborne intelligent surfaces. Furthermore, we present a case study demonstrating the effect of MAB schemes on throughput and energy efficiency in UAV-mounted RIS.

This artifact accompanies the paper:"A Teaching Assistant for Teacher–Student Learning: Knowledge Transfer from Skeleton to Inertial Sensing for Activity Recognition in Industrial Domains [1]." The artifact provides the implementation, configuration files, and preprocessed data required to reproduce the experimental results reported in the paper. Specifically, it supports training and evaluation of the proposed Teacher–Assistant–Student (TAS) framework for transferring knowledge from skeleton-based models to IMU-based activity recognition models in industrial environments. The code, data and supplementary information can be publicly downloaded from the following repository: https://github.com/qiaohongyin/TAL_research.

The proliferation of low-Earth orbit (LEO) satellite constellations presents unprecedented opportunities for distributed machine learning (ML) applications. However, the inherent challenges of sparse connectivity, heterogeneous communication windows, and non-independent and identically distributed (non-IID) data across satellites hinder the effectiveness of conventional federated learning (FL) frameworks. To address these challenges, we propose Model Contrastive Federated Learning (MCFL), a novel framework tailored for LEO satellite constellations. MCFL introduces a two-stage approach: (1) similarity-based satellite clustering to mitigate intra-cluster data imbalance by grouping satellites with aligned data distributions, and (2) collaborative staleness-aware learning that employs semi-asynchronous model aggregation within clusters to balance convergence speed and model accuracy. The key contributions include a contrastive loss function for robust representation learning under class imbalance, gradient sparsification to minimize communication overhead, and an inter-cluster knowledge-sharing mechanism to prevent cluster-specific model bias. Extensive simulations on the EuroSAT dataset demonstrate that MCFL achieves an improvement of 15% in test accuracy and reduces training time 3× compared to state-of-the-art FL baselines while reducing communication costs by 40%. This work bridges the gap between distributed learning theory and practical satellite constraints, offering a scalable solution for real-time ML applications in dynamic space-terrestrial networks.

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 prediction system, harnessing the strengths of multiview graph neural networks to capture spatial-temporal dependencies and patterns in urban environments. Additionally, the proposed system CROSS-Net employs a spatially transferable approach, enabling it to train a model that can be deployed to previously unseen regions. To achieve this, the framework incorporates the power of a Variational Autoencoder to disentangle the input features into region-specific and region-agnostic components. The region-agnostic features facilitate cross-region taxi demand predictions, allowing the model to generalize well across different urban areas. Experimental results demonstrate the effectiveness of CROSS-Net 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.

Autonomous Vehicles (AVs) must make reliable decisions in dense urban environments where pedestrian behavior is variable, sometimes abnormal, and often unseen during training. Reinforcement learning (RL)-based AV control systems perform well in structured traffic but struggle to generalize to unpredictable pedestrian interactions and out-of-distribution scenarios. Their reliance on handcrafted rewards and opaque decisions further limits their suitability for safety-critical, pedestrian-rich environments. To address these limitations, we introduce a Large Language Model (LLM)-based decision-making framework for pedestrian-aware behavioral planning. The system converts structured scene observations into natural-language reasoning prompts, enabling the LLM to infer pedestrian intent, anticipate risk, and generate cautious tactical driving decisions. These decisions are executed by a motion planner that ensures smooth, kinematically feasible control. We evaluate the framework in SUMO across multiple pedestrian-interaction scenarios, including unexpected jaywalking, turn-back crossing, hesitation, and bidirectional crossing. In zero-shot evaluation, the LLM-based agent achieves a 68

Rainfall-induced landslides pose a growing risk worldwide as climate change intensifies extreme rainfall events. To provide sufficient evacuation time, landslide early warning systems (LEWS) for real-time disaster monitoring must estimate near-future landslide risk by integrating observed rainfall with short-term rainfall forecasts from spatio-temporal environmental data streams. Although recent landslide prediction methods have improved predictive performance using statistical and deep learning approaches, most assume accurate rainfall inputs. In operational settings, however, landslide prediction relies on rainfall forecasts, which often contain spatial displacement of rainfall fields due to forecasting uncertainties. Such displacement can alter local accumulated rainfall and degrade prediction accuracy. To address this challenge, we propose a novel LEWS robust to rainfall field displacement. The key idea is to learn latent representations from rainfall and terrain data that remain stable under displacement in rainfall field motion, enabling reliable geospatial data integration for landslide risk estimation. The landslide prediction model is trained using Rainfall-Motion-Aware Contrastive Learning (RMCL), which introduces temporally correlated rainfall field perturbations to emulate forecast-induced displacement in rainfall-driven spatio-temporal environmental data streams. Experiments were conducted using two years of rainfall and terrain data across Japan, covering 19 regions with landslide events. The proposed system achieved up to 37

Accurate and reliable landslide prediction is essential for proactive disaster management, where risk assessments must support timely decision-making. However, landslide prediction remains challenging due to data scarcity, severe class imbalance, and regional distribution shifts, which degrade both discriminative performance and probability calibration. In this study, we propose a physics-integrated deep learning system that produces decision-ready, calibrated probability estimates across regions. The proposed framework combines three key components: (1) integration of physically grounded features to enhance learning under limited data, (2) Region-Conditional Contrastive Learning (RCCL) to learn region-invariant representations, and (3) calibrated training with MDCA-based objectives to improve confidence reliability. By jointly addressing generalization and calibration, the proposed system mitigates overconfidence and maintains predictive consistency under regional shifts. Experimental results on multi-region datasets demonstrate that our approach outperforms state-of-the-art methods in both predictive performance and calibration, particularly in unseen regions, enabling more reliable and actionable risk assessment.

Human activity recognition (HAR) in industrial domains is important for workflow optimization, throughput estimation, and bottleneck detection. Skeleton-based models achieve high HAR accuracy by exploiting rich spatial and temporal cues, but they are difficult to deploy in industrial sites due to occlusions, camera placement, and privacy concerns. IMU sensors, especially smartwatches, are practical for deployment but lack spatial awareness, resulting in weaker performance. This work aims to enable robust HAR using only a wrist-worn IMU by distilling knowledge from richer modalities. Knowledge distillation allows transferring information from a skeleton teacher to a single-IMU student, but the large modality gap has limited the success of prior teacher–student approaches. To address this issue, we propose a teacher-assistant-student (TAS) learning framework, in which a multi-IMU assistant model bridges the skeleton-based teacher and the single-IMU student. To support TAS, we develop the following techniques: (i) Dense temporal Contrastive Learning, aligning structural representations of skeleton and IMU segments; (ii) Spatial Relationship Learning, guiding models to capture spatial priors from skeleton data; and (iii) Temporal Attention Transfer, distilling attention patterns for key atomic actions. We further boost the robustness to behavioral variation with motion-guided IMU data diversification using physics-based simulation. Experiments on industrial HAR sensor data show that our framework consistently improves single-IMU recognition across diverse operational scenarios, highlighting its potential for practical deployment.

Accurate indoor localization on resourceconstrained edge devices remains challenging due to the tension between modeling fidelity and computational efficiency. Wi-Fi Channel State Information (CSI) enables low-cost deployment using existing Wi-Fi infrastructure but is heavily distorted by multipath propagation. Transformer-based solutions mitigate these effects but their quadratic complexity limits real-time edge deployment. Ultra-Wideband (UWB) provides geometrically precise ranging yet requires dedicated infrastructure. We present MambaLoc, a cross-modal distillation framework that improves the accuracy-efficiency balance in CSI-based localization. MambaLoc uses a linear-time Selective State Space Model (Mamba) as a lightweight CSI student, guided during training by a UWB-Transformer teacher. A Feature Alignment Module bridges the representational gap between heterogeneous RF modalities, allowing the CSI model to inherit UWB-derived geometric structure without UWB hardware at inference. A probabilistic Gaussian Mixture regression head models multi-modal spatial ambiguity from multipath effects, avoiding the physically implausible averaging of deterministic regression. Evaluations on a real-world dataset show MambaLoc achieves a median localization error of 0.230 m, reducing error by 22.2 % over a standalone Mamba baseline, while requiring 65.1% fewer FLOPs and 58% lower inference latency than Transformer-based approaches.