Search

This paper describes the design and implementation of a rail-guided mobile robot-based monitoring system that can be used for autonomous monitoring of the facilities and rapid response to the on-site event. It is necessary to improve the safety and efficiency of the monitoring system when applying a robot system, and the implementation of intelligent functions can increase the level of automation to facilitate operation. In order to increase driving safety and lower the level of control, a two-track rail structure of circular pipes, U-shaped wheels, and a differential gear are used to develop a driving system that enables rotational path driving even with forward and backward movement. For the proper monitoring of facility conditions, high-temperature occurrence, and environmental quality, cameras (RGB and thermal imaging) and air quality sensors are chosen by considering the range of the monitoring site and target subjects. Data communication between the robot and the remote control room(a remote server) is built up to patrol in a planned schedule and deliver acquired video streams and environmental sensor data. Emergency alarms about high temperatures and the abnormality of air conditions can be generated by processing the transferred data. Additionally, deep network-based detection services on video data are implemented on the server for the detection of emergency situations; fall downs. Integrating the developed driving system, image/sensor data communication, and alarm function through data analysis in an operating framework reduces management work through automatic patrol and enables efficient on-site response when an alarm occurs. The validity of the proposed system is demonstrated by building and operating the system in the existing multi-use facility. The automatic patrol function and the manual operation function are performed through a test operation similar to the facility monitoring task of a human operator. The operator can control the robot to move to, 2023 IEEE/SICE International Symposium on System Integration (SII), 17-20 January 2023, Atlanta, GA, USA, identifier:https://dspace.jaist.ac.jp/dspace/handle/10119/18197

An object’s inertial parameters, such as the mass, the Center of Mass (CoM), and the moment of inertia, affect the response to the external forces exerted on it. It is important to estimate these parameters as accurately as possible in order to facilitate robot-led automation including grasping and manipulation. Traditionally, the estimation is conducted by employing special equipment in a controlled environment, which may not be always available for a small batch production system dealing with unknown objects. We propose an efficient exploratory type framework for estimating an object’s CoM via force sensor-less high-speed robotic pushing, which only requires the use of a vision system for detecting the change in the object’s pose. Accurately estimating an object’s CoM aids robotic grasping and expands manipulation scenarios. We conducted intensive simulation and real robot experiments to show the accuracy of the estimation, robustness to friction variation, and generalization capability to novel objects, that the framework ensures only with a limited number of pushes., 2023 IEEE 19th International Conference on Automation Science and Engineering (CASE), Auckland, New Zealand, August 26-30, 2023

It is one of the most fundamental and challenging problems for current dialogue generation systems to maintain the consistency of dialogue logic during the conversation. Also, the lack of an open-source annotated personabased dialogue dataset may lead to insufficient training volume for the model. Besides, the computational time and computational memory required by the attention mechanism have become drastically large. In order to overcome these problems, in this work, we propose a vertical-structure model based on the BERT model with the sentence embedding method. The model generates a raw response based on the sentence embeddings of context and persona and finally revises the raw response according to the persona. Moreover, an understanding task is designed for the BERT decoder to have a better revision ability. Considering the difference between the generation and the understanding models, three kinds of input methods are designed for each part of the model. Comparative and experimental results are presented using publicly available datasets., 2023 23rd International Conference on Control, Automation and Systems (ICCAS 2023), Yeosu, Korea, October 17-20, 2023

In recent years, medical image segmentation using deep learning methods has become more and more popular and developed with the aim of both reducing human-related errors and the time required for manual segmentation. One of the pioneers in deep learning-based biological image segmentation networks, U-Net was proposed back in 2015. Since then, several models have been proposed to extend U-Net. However, the trade-off between computational complexity and accuracy remains a major challenge. To address this trade-off, we use a new Involution kernel for spatial information and propose a model lightweight medical image segmentation network, Residual Involution U-Net (RDIU-Net). Involution, Residual, and Dense structures are incorporated into the U-Net model to extract both channel and spatial features. Evaluations have been carried out on three different datasets of ultrasound, X-ray, and dermoscopic images. The proposed model RDIU-Net showed superior results in accuracy, processing speed, training stability, and convergence compared to U-Net., 2023 23rd International Conference on Control, Automation and Systems (ICCAS 2023), Yeosu, Korea, October 17-20, 2023

The segmentation of ultrasound (US) images plays a crucial role in the development of end-to-end smart diagnosis systems. In the diagnostic stage, specific diagnosis programs can be applied to well-cropped sub-regions within a single US image, catering to different medical interests. In this study, we propose and test a neural network model designed to perform end-toend segmentation on abdominal US images, with a focus on five different anatomical structures: liver, kidney, vessels, gallbladder, and spleen. The main contribution of our work lies in the exploration of multi-organ/tissue segmentation. Unlike previous research, our approach takes into account two inherent features of US images: (1) significant variations in spatial sizes among different organs and tissues, and (2) the relatively consistent spatial relationships among anatomical structures within the human body. To address these considerations, we introduce a novel image segmentation model that combines the feature pyramid network (FPN) and the spatial recurrent neural network (SRNN). In our paper, we describe the utilization of FPN for extracting anatomical structures of varying scales, as well as the implementation of SRNN to capture spatial context features within abdominal US images. Our model incorporates both top-down and bottom-up pathways, enhancing both semantic features and spatial context features. We refer to this as the ”two-path augmented” approach. Furthermore, we incorporate a directional attention mechanism, which selectively leverages spatial context information from four principal directions. This is the essence of our ”directional context aware” component. The performance of our proposed model is evaluated through both quantitative and qualitative measures. The evaluation results demonstrate the competitiveness of our approach, and the inclusion of spatial contextual information has resulted in improved performance compared to using the pure feature pyramid network alone., 2023 IEEE International Conference on Mechatronics and Automation (ICMA), August 6 - 9, Harbin, Heilongjiang, China