日時: 12月2日(金)3限(13:30-15:00)

会場: L2

司会: 磯山 直也
照岡 肇 M, 1回目発表 ロボットラーニング 松原 崇充, 池田 和司, 鶴峯 義久, 花田 研太
title: ***Multi-task Reinforcement Learning Based on Time-Phase Logic Focused on Task Progression ***
abstract: *** In the reinforcement learning problem of generating actions according to instructions, methods using linear temporal logic that can design instructions for a variety of multitasks have been actively studied. However, in existing research, there are cases in which the progress status is misrecognized due to changes in the positional relationships of objects in the environment, etc., and the task is not correctly accomplished. In this study, we aim to propose a method for such situations that corrects the misrecognition of the progress status. ***
language of the presentation: *** English or Japanese (choose one) ***
発表題目: *** タスクの進行に着目した時相論理に基づくマルチタスク強化学習***
発表概要: *** 指示に従って動作を生成する強化学習問題において、多様なマルチタスクの指示を設計できる線形時相論理を用いた手法が盛んに研究されている。しかし、既存研究では環境内のオブジェクトの位置関係の変化等によって進行状況を誤認識し、正しくタスクを達成できない場合がある。。本研究では、そのような状況に対して、進行状況の誤認識を是正するような手法を提案することを目指す。 ***
市 尚都 M, 1回目発表 ロボットラーニング 松原 崇充, 池田 和司, 花田 研太, 佐々木 光

title:Optimal Control Method for Explosive Motion with Soft Under-actuatede Manupilator  

abstract:Unlike typical industrial robots, soft under-actuated manipulators can accomplish explosive motion(e.g..,push-up motion utilizing swing-up motion). However, optimal control of this motion requires stable long-term predictive model and identification of nonlinear dynamics. In this study, we focused on data-driven modeling based on Koopman operator theory and its optimal control called Deep Koopman Network, and verified the prediction errors of LQR and multi-step. It was found that future works should focus on long-term prediction and handling of high-dimensional state inputs and states. 

language of the presentation:Japanese 

長田 瑛綺 M, 1回目発表 ロボットラーニング 松原 崇充, 和田 隆広, 佐々木 光, 清川 拓哉
title: Self-supervised grasp learning with prior knowledge for automatic sorting of mixed industrial waste on a conveyor.
abstract: Mixed industrial waste treatment plants are still being manually sorted, thus, automating manual sorting process with robots is expected. The mixed industrial waste items come in a variety of shapes and materials, and are densely gathered and moving on a conveyor. Previous self-supervised grasp learning makes it difficult to learn efficiently and to learn grasping policy of moving objects. To tackle the moving object grasping, we proposed a framework including modules specialized for predicting the grasping position and object movement error. To achieve efficient learning, the self-supervised learning modules can be started from by using prior knowledges of graspable areas.
language of the presentation: Japanese
発表題目: コンベア搬送される混合産業廃棄物の自動仕分けのための事前知識を用いた自己教師あり把持学習
発表概要: 混合産業廃棄物の処理場での人手による仕分け作業について,汎用性の高いロボットを用いた自動化が求められている.コンベア上を移動する密集した多様な形状や材料の廃棄物を効率的にロボットで仕分ける必要がある.従来の試行錯誤的な把持学習では,効率的な学習や移動物体の把持が困難である.本研究では,物体把持位置と移動誤差を予測することに特化させたモジュールを自己の試行錯誤による成否から学習させ,事前知識を活用することで,移動物体の把持を効率的に学習可能な枠組みを構築する.
森田 俊平 M, 1回目発表 ロボットラーニング 松原 崇充, 和田 隆広, 鶴峯 義久, 佐々木 光
title: Motion planning for grasping and placing food ingredients for food arrangement by a robot
abstract: In recent years, labor shortages have created a need for automation in the restaurant industry. One particularly difficult task is food arrangement. Although there is research on the position and posture of arrangements, the lack of feasibility using robots is a problem. Since it is not always possible for a robot to make the arrangement target made by a human, it is necessary to consider the circumstances of grasping and arranging the food. In this study, we develop a grasping and arrangement of ingredients motion plan for food arrangement by a robot.
language of the presentation: Japanese
発表題目: ロボットによる料理盛り付けのための食材把持・配置動作計画
発表概要: 近年、労働力不足によりレストラン産業の自動化が求められている。とくに難しい作業として食材の盛り付けがある。盛り付けの位置や姿勢に関する研究はあるが、ロボットによる実現性がないことが問題である。人間が作った盛り付け目標をロボットが作ることができるとは限らないため、把持・配置の事情を考慮する必要がある。本研究では、ロボットによる料理盛り付けのための食材把持・配置動作計画の開発を行う。
CHEN TIEN HAO M, 1回目発表 数理情報学 池田 和司, 金谷 重彦, 久保 孝富, 福嶋 誠, 日永田 智絵

title: Omics-data analysis for subtying delirium

abstract: Delirium is a neuropsychiatric syndrome which is characterized in higher cortical functions impairments. Moreover, the elderly population have been on the rise in recent years with following phenomena such as mortality, longer hospitalization, dementia caused by delirium results. To clarify the etiology of delirium, the investigation of subtyping the complex symptoms is essential. In present study, we would like to discover establishment of complex categorization of subtype delirium using linear models of classification and biomedical statistical methods to clearly understand subgroups of delirium patients.

language of the presentation: English

山田 浩太 M, 1回目発表 大規模システム管理 笠原 正治, 松本 健一, 笹部 昌弘, 原 崇徳

title: *** On Cloud/Fog Computing Resources Allocation for Decentralized Applications *** 

abstract: ***Decentralized applications (DApps), which enable the people to independently design a smart contract without centralized controls, have attracted many researchers, entrepreneurs and/or developers. Since the DApp works on the permissionless blockchain, block mining is required to maintain the transaction consistency, where block mining aims at solving a certain cryptographic puzzle by using a large amount of computational resources. Most DApp users, however, cannot be a block miner due to lack of their computational resources, which may result in the degradation of blockchain integrity. To tackle this problem, the existing work considered to offload the computational tasks of miners to a cloud/fog computing provider (CFP) and formulated a cloud/fog computing resource allocation problem as an auction mechanism between miners and a CFP. However, this work does not consider the resource assignment over time and competitive relationship among different CFPs. In this research, we formulate the cloud/fog computing resource allocation problem as a repeated stochastic game. In addition, we propose a Lyapunov optimization based resource allocation offloading the computational tasks to the multiple CFPs while holding the coarse correlated equilibrium (CCE) constraints, which can be calculated faster than Nash equilibrium, where CCE is an equilibrium that rationally chosen actions by all players match the suggestions by a game manager under the assumption that no-participating players cannot receive the suggestions. *** 

language of the presentation: *** Japanese ***