| 和田 遥香 | M, 2回目発表 | ユビキタスコンピューティングシステム | 安本 慶一, 小笠原 司, 荒川 豊, 藤本 まなと |
|
title: Efficient Care Recording Application with Location-based Automatic View Transition and Information Complement
abstract: In nursing homes or hospitals, It is necessary to take detailed records (e.g. excretion in the toilet, the meal at the dining, etc.) including the location information. In this research, we propose a mobile memo system that switches the view according to the location automatically and assists inputting the necessary information. In the system, staffs use the small size device for making reports. This device can switch the view and make report by receiving signals send from the BLE beacons attached to the residents and each location of a nursing home. To test the effectiveness of the proposed system, I carried out evaluation experiments in the real nursing home. As a result, I found that the proposed system can identify the nursing area or the care recipient and staff can input records easily and correctly. language of the presentation: Japanese | |||
| 松岡 潤樹 | M, 2回目発表 | 知能システム制御 | 杉本 謙二☆, 松原 崇充, 小笠原 司 |
|
title: Learning Food Arrangement Policies with Imitation Learning
abstract: The food arrangement is one of the challenging task in kitchen work due to its ambiguous evaluation criteria. As a first step for automating it by a robot, in this presentation, we focus on the food arrangement planning problem and propose a framework for learning food arrangement policies using an imitation learning method so-called P-GAIL. The effectiveness of our framework is evaluated in simulations. language of the presentation: Japanese | |||
| 宮本 知弥 | M, 2回目発表 | 知能システム制御 | 杉本 謙二☆, 松原 崇充, 小笠原 司 |
|
title: Efficient shape estimation of rigid body covered with flexible materials
abstract: In a real environment, there are many objects covered with flexible materials. For such objects, it is hard to estimate its internal shape by visual information. Tactile information is useful but maybe inefficient since it only provides limited information at once. In this study, we introduce a shape estimation method combining the uncertain visual with more certain tactile information using Gaussian process implicit surface. language of the presentation: Japanese | |||
| 鶴峯 義久 | D, 中間発表 | 知能システム制御 | 杉本 謙二☆, 松原 崇充, 小笠原 司 |
|
Title: Deep Reinforcement Learning for Robotic Cloth Manipulation
abstract: Deep Reinforcement Learning (DRL) has drawn much attention in robot control since it enables agents to learn control policies from very high dimensional states such as raw images. However, it is difficult to apply DRL to robot control due to the following problems: P1) DRL learns from a large number of samples, but a real robot has a limited number of samples. P2) DRL requires a reward function to evaluates the policy, and design of a reward function suitable for each task is required. In this research, we approach these problems by following two methods: M1) Propose sample efficient DRL, M2) Learning reward functions from expert demonstrations. Each of the proposed methods was applied to a clothing manipulation task using a dual-arm robot, and the proposed method was demonstrated to be effective even in a real environment. language of the presentation: Japanese 発表題目: 実ロボット衣類操作タスクに適した深層強化学習 発表概要: 深層強化学習は画像等の高次元センサを入力とする方策を学習することができ，ロボット制御への適用に注目が集まっている．しかし，以下の問題でロボット制御への適用は困難である．P1)深層強化学習は膨大なサンプルから学習するが，実際の動作を通して実行するロボットは収集できるデータが少ない，P2）学習には方策を評価する報酬関数が必要であり，タスクごとに適した報酬関数の設計が求められる．これらの問題に対して本研究では以下の2つの方法でアプローチした．M1）サンプル効率の良い深層強化学習の提案，M2）エキスパートデータから報酬関数の学習．本研究ではそれぞれのアプローチを双腕ロボットによる衣類操作タスクに適用し，実環境でも提案手法が有効であることを確認した． | |||
| 森谷 友香 | M, 2回目発表 | 生体医用画像 | 佐藤 嘉伸, 加藤 博一, 大竹 義人, スーフィー マーゼン |
|
title: Musculoskeletal Segmentation and Shape Analysis of Metal Artifact-Reduced Maxillofacial CT Images
abstract: The segmentation of the musculoskeletal structures, e.g. masseter muscle, in maxillofacial computed tomography (CT) images is essential for the diagnosis and treatment planning of maxillofacial disorders. However, the presence of a metal prosthesis, such as a dental filling, leads to metal artifacts in the CT images that affect the segmentation accuracy. In our previous study, we developed a musculoskeletal segmentation method applied to metal artifact-reduced CT images and evaluated it based on the manual trace produced from metal artifact-reduced CT images. In this research, instead, we simulated metal artifacts in the CT images and evaluated the segmentation accuracy using labels produced from CT images without metal artifact. In addition, we proposed an improvement of the conventional Normalized Metal Artifact Reduction (NMAR) method, and compared the impact of three metal artifact reduction methods on the segmentation accuracy. Future work will focus on applying the proposed method to larger datasets, and the shape analysis of the musculoskeletal structures. language of the presentation: Japanese | |||
| 大久保 達矢 | M, 2回目発表 | サイバネティクス・リアリティ工学 | 清川 清, 加藤 博一, 酒田 信親 |
|
title: Development of a Shopping Support System for Visually Impaired People
abstract: It is generally difficult for a visually impaired person to find an item at a shop that he or she wants to buy by himself or herself, and asking somebody for assistance is not always feasible. To address the problem, we have designed a shopping support system for visually impaired. Our system features three functions; (1) creation of an environmental map, (2) path-finding and navigation of a user. In this presentation, we report on the detail of the system design and current progress. language of the presentation:Japanese | |||