Yuichi Ohsita

The increasing deployment of Automated Mobile Vehicles (AMVs) in warehouses requires efficient navigation in dynamic environments shared with human workers. Traditional safety strategies, such as stopping or slowing down, often reduce efficiency and do not account for human-robot interaction. To address this, interaction-aware methods have been developed. However, traditional interaction models assume fixed parameters, rely on a single forecast, or use intent labels that are not updated during the interaction, and thus do not explicitly represent intention uncertainty, that is, variability in how people react to robots. When combined with position uncertainty introduced by sensing noise, this unmodeled intention uncertainty can make motion prediction and navigation decisions unreliable, leading to conservative behavior, inefficiency, or collisions in human–robot interactions. To bridge this gap, we propose Belief-based Risk-Aware Anticipatory Navigation (BRAN), a framework that jointly models intention and position uncertainty. BRAN maintains probabilistic beliefs over an agent's position and intention to explicitly represent uncertainty in both quantities. Position uncertainty is represented via sampled observations, and the intent belief is updated online through Bayesian inference so that predictions evolve with the interaction. These beliefs are propagated through the Social Force Model (SFM) to generate predictions of possible agent motion under uncertainty. We then compute collision risk from the predicted motion distribution and incorporate it into model predictive control as a constraint. This constrains predicted collision risk while optimizing motion toward the goal. Experimental results validate the proposed design in controlled settings, demonstrate its effectiveness in scenarios with multiple pedestrians, and confirm its feasibility on a real robot with onboard sensing and actuation.

A Transmission Control Protocol (TCP) throughput map represents communication quality over road networks and supports communication-aware applications in intelligent transportation systems. Maintaining such a map online is challenging because vehicular measurements are sparse and unevenly distributed, network conditions vary rapidly, and sensing-budget constraints limit the number of vehicles from which onboard communication measurements can be uploaded at each time step. This work addresses online TCP throughput map maintenance under sparse vehicular observations and sensing-budget constraints. To support budget-constrained sensing, we combine discoverability-guided vehicle selection and probabilistic map updating within a digital twin (DT)-assisted vehicular sensing architecture. The resulting sensing-and-mapping method, referred to as Discoverability-aware and Statistical Mapping (DISMAP), maintains a spatio-temporal discoverability map to characterize historical sensing coverage and select vehicles that improve the coverage of under-represented regions. It then uses Gaussian Process Regression (GPR) as a probabilistic mapping engine to estimate the mean TCP throughput and predictive standard deviation, where the standard deviation is adjusted using local vehicle density. Simulation results show that DISMAP reduces the mean absolute error (MAE) and mean standard deviation (MSTD) by up to 23.7% and 37.5%, respectively, and achieves a prediction-interval miss rate (PIMR) of 0.048, which is close to the nominal value of 0.05. These results indicate a favorable balance among prediction accuracy, interval sharpness, calibration, and spatial representativeness across different traffic-density conditions.

Finding a correlation between physiological measures and subjective well-being (SWB) or comfort has been an active research area in recent years. We focus on short-term SWB measures and their correlation to electroencephalography (EEG) signals in an office environment. We recorded EEG from 30 participants and asked them to report their SWB every 30 s. We analyzed the correlation between the relative power of different frequency bands at various sensor locations and SWB via k-nearest neighbor (k-NN) classification and linear regression. We also analyzed the correlation of the time series themselves at different sensor locations and how they can be classified into different SWB values via k-NN. Then, we tried to cluster participants into subgroups that had a similar correlation between their EEG recordings and their reported SWB. We found that a correlation between relative power and SWB also holds for short terms. However, the results of every single participant of all analyses vary substantially, and we could not find any consistent clustering into subgroups. That implies a huge individuality when it comes to EEG measures and reported short-term SWB.

Deep learning (DL) models have rapidly evolved, and their scales have become larger. Pipeline parallelism is used to execute a large-scale DL model. In pipeline parallelism, DL models are partitioned and allocated graphics processing units (GPUs) to execute each partition. However, to execute numerous DL services in clusters with heterogeneous GPUs, a DL model partitioning that considers a specific type and number of GPUs available and the GPUs allocated for the service is required. We propose resource aware model partitioning and allocation (RAMPA) to execute more DL services while satisfying the performance requirements. RAMPA minimizes the allocation of important resources for future DL service execution to avoid inhibiting future DL service execution. We define the resource allocation cost based on resource importance. Furthermore, we formulate the impact of model partitioning and allocated resources on service performance. We define an optimization problem to minimize resource allocation costs while satisfying service performance requirements. We evaluated the effectiveness of RAMPA by simulating the execution of DL services in clusters with heterogeneous GPUs. The results demonstrate that more services can be executed while satisfying performance requirements compared to the conventional method. RAMPA enabled efficient GPU utilization to deliver many DL services.

The digital twin is a high-precision, real-time representation of the real world. It has evolved from a digital copy of physical objects, such as jet engines, robots, or any machine, to a broader concept to represent a space, room, building, and city, where many objects and humans interact to engage in productive activities. The digital twin has been attracting attention as an important component of cyber-physical systems in many real-world use cases where safety and trustability are critical. In this article, we argue that risk management is dramatically improved through probabilistic representation of the real world whose true status is only inferred with certain uncertainties. We introduce an evolution of the digital twin and then present a probabilistic digital twin of the real world in which the identity and status of any object are represented in a probabilistic graphical structure. Then, we introduce use cases of the probabilistic digital twin, such as probabilistic object recognition, prediction of the existence of obstacles, and risk-sensitive robot control for robotized automation in various environments such as warehouse operations.

A disaggregated data center (DDC) can efficiently use resources such as CPU and memory. In a DDC, because each resource is independent and connected by a network, communication between resources is required for task execution. Communication delays can be an overhead for task execution, causing performance degradation. Because communication delays depend on the correspondence between resources on the network and the paths over which they communicate, an efficient resource allocation method is required to determine this relationship. Herein, we propose a resource allocation method called RA-CNP to execute many tasks simultaneously while satisfying performance requirements. This method models the impact of the network on the performance of tasks for the provided service. Furthermore, this method defines a resource allocation problem to avoid the allocation of resources that will be requested in the future. We evaluated the effectiveness of our method by simulating various DDC networks, assuming a DDC at the edge. The results demonstrated that RA-CNP could execute more tasks than conventional methods could, without violating performance requirements, based only on current network information in both networks configured by circuit and packet switches. RA-CNP could allocate resources in less than 10 s, even in a relatively large network configured with 64 switches; this capability demonstrates its practicality.

Machine learning (ML) systems that use sensors obtain observations from sensors and use them to recognize and interpret the current situation. These systems are susceptible to sensor-based adversarial example attacks (AEs); if some sensors are vulnerable and can be compromised by an attacker, the attacker can change the output of the system by changing the values of the sensors. The detection of compromised sensors is important to defend the system against sensor-based AEs, because we can check the sensors and replace them by detecting the sensors used by the attacker. In this paper, we propose a method to detect the sensors used in sensor-based AEs by utilizing the features of the attack that cannot be avoided. In this method, we introduced a model called the feature-removable model (FRM), which allows us to select the features used as inputs into the model. Our method detects the sensors used in sensor-based AEs by finding the inconsistencies between the outputs of the FRM obtained by changing the selected features. We evaluated our method using a human activity recognition model with sensors attached to the user’s chest, wrist, and ankle. We demonstrate that our method can accurately detect sensors used by the attacker and achieves an average Recall of Detection of 0.92, and the average Precision of Detection is 0.72.

Subjective well-being (SWB) describes how well people experience and evaluate their current condition. Previous studies with electroencephalography (EEG) have shown that SWB can be related to frontal alpha asymmetry (FAA). While those studies only considered a single SWB score for each experimental session, our goal is to investigate such a correlation for individuals with a possibly different SWB every 60 or 30 s. Therefore, we conducted two experiments with 30 participants each. We used different temperature and humidity settings and asked the participants to periodically rate their SWB. We computed the FAA from EEG over different time intervals and associated the given SWB, leading to pairs of (FAA, SWB) values. After correcting the imbalance in the data with the Synthetic Minority Over-sampling Technique (SMOTE), we performed a linear regression and found a positive linear correlation between FAA and SWB. We also studied the best time interval sizes for determining FAA around each SWB score. We found that using an interval of 10 s before recording the SWB score yields the best results.

Recently, artificial intelligence (AI) based on IoT sensors has been widely used, which has increased the risk of attacks targeting AI. Adversarial examples are among the most serious types of attacks in which the attacker designs inputs that can cause the machine learning system to generate incorrect outputs. Considering the architecture using multiple sensor devices, hacking even a few sensors can create a significant risk; an attacker can attack the machine learning model through the hacked sensors. Some studies demonstrated the possibility of adversarial examples on the deep neural network (DNN) model based on IoT sensors, but it was assumed that an attacker must access all features. The impact of hacking only a few sensors has not been discussed thus far. Therefore, in this study, we discuss the possibility of attacks on DNN models by hacking only a small number of sensors. In this scenario, the attacker first hacks few sensors in the system, obtains the values of the hacked sensors, and changes them to manipulate the system, but the attacker cannot obtain and change the values of the other sensors. We perform experiments using the human activity recognition model with three sensor devices attached to the chest, wrist, and ankle of a user, and demonstrate that attacks are possible by hacking a small number of sensors.