| 荒木 駿佑 | M, 2回目発表 | ロボットラーニング | 松原 崇充 | 和田 隆広 | 柴田 一騎 | 鶴峯 義久 | 佐々木 光 | |
|
title: Imitation of Human Force Tactile Skills Using Bilate Control for Grinding Operation
abstract:Automation of grinding operations using robots requires consideration not only of workpiece geometry and pressing posture, but also of the density within the workpiece and the condition of the grinding belt. These conditions affect the removal volume and removal resistance, which are difficult to capture with external sensors. In this study, we developed a framework that mimics human grinding skills and uses force-tactile information to grind workpieces of different densities with the appropriate amount of force. For verification, we conducted tests on an actual workpiece with constant workpiece geometry and grinding depth, and confirmed that the appropriate force level can be inferred from the force-tactile information in a time series, and that grinding can be performed in a shorter time than with conservative grinding. language of the presentation: Japanese | ||||||||
| 稲垣 海 | M, 2回目発表 | ロボットラーニング | 松原 崇充 | 和田 隆広 | 柴田 一騎 | 鶴峯 義久 | 佐々木 光 | 権 裕煥 |
|
title: Capability Estimation in Multi-Robot Cooperation Using Multi-Agent Reinforcement Learning
abstract: Multi-agent reinforcement learning (MARL) methods have achieved remarkable success in recent years in the field of robotics as a means of problem solving in which multiple agents cooperate to perform tasks. In conventional methods, in order to account for individual differences among robots, it is necessary to define individual differences as capabilities in advance. This is because simultaneous learning of ability representation and cooperative measures is difficult due to the expansion of the search space caused by the combination of abilities. In contrast, this study proposes a stepwise learning approach. First, trajectory data for each agent's capabilities are collected using ID-based MARL, and then latent vectors of capabilities are learned using an encoder-decoder model. Finally, MARL is performed again based on the acquired capability vectors to obtain efficient cooperation strategies. With this method, capability estimation is possible even for unknown agents, and it is expected to learn general-purpose cooperative strategies. language of the presentation: Japanese | ||||||||
| 梶原 隆太郎 | M, 2回目発表 | ロボットラーニング | 松原 崇充 | 和田 隆広 | 柴田 一騎 | 鶴峯 義久 | 佐々木 光 | |
|
title: Hierarchical Imitation Learning with Entanglement Representation for Cable Bundling Tasks
abstract: In server and factory control systems, many devices and control panels are connected by cables, and it is known that bundling these cables improves maintainability. This task is performed manually, and automating this task using robots is expected to improve work efficiency. Imitation learning is a method for learning control policies for automation from human work data. It is an effective method for automating cable manipulation, which is a deformable object whose actions are difficult to evaluate. This study proposes a framework for automating cable bundling tasks by extracting features from cable images using topological representations and hierarchical imitation learning. language of the presentation: Japanese | ||||||||