| 米野 尚斗 | D, 中間発表 | ロボットラーニング | 松原 崇充, 和田 隆広, 佐々木 光 |
|
title: Utilization of Flexible Structures for Extended Tactile Sensing and Application to Manipulation Control
abstract: Human labor has already been replaced by robots in many workplaces. However, there are still many tasks that can only be performed only by humans. One of the reasons for this is the difference in physicality between robots and humans. Humans are covered with skin all over their bodies, especially the flexible structure of their fingers, allowing them to manipulate complex physical phenomena through deformation and slip. In this study, we propose a new robot system with such flexible structures. The first is the extension of tactile sensing capability using the flexible structure as a medium. By utilizing the fact that the flexible structure deforms in accordance with the object, texture recognition without sliding motion and simple incipient slip detection can be realized. The other application is to manipulation control. Dynamic utilization of nonlinear deformation of flexible structures enables dexterous manipulation. As a preliminary study, we present analysis and control using the Koopman operator with a flexible inverted pendulum. In future work, we aim to realize a tactile servo by combining the proposed tactile sensing method and manipulation control. language of the presentation: Japanese | |||
| OH HANBIT | D, 中間発表 | ロボットラーニング | 松原 崇充, 和田 隆広, 佐々木 光 |
|
title: Exploiting Human Behavior Characteristics in Interactive Imitation Learning for Robot Manipulation
abstract: Imitation learning, as an alternative to explicit programming of the robot, provides an intuitive means of enabling robots to learn skills via observing a human expert demonstrator. A central problem of this approach is that even slight deviations from the human-demonstrated coverage can compound errors during policy execution; this phenomenon is commonly referred to as covariate shift. Human-robot interaction technique reduce covariate shift by iteratively a robot requests augmented demonstrations from a human and captures human behaviors from accumulated demonstrations. However, its applicability is still limited in real-world manipulation tasks, which often involve complex human behaviors since it is human behavioral characteristics-agnostic. To address this, we propose human behavioral characteristics-aware approaches, designing human-robot interaction with a focus on utilizing human behavioral characteristics; thereby embodying principles for capturing and exploiting actual demonstrator behavioral characteristics. Therefore, our idea is embodied in two main interactive IL frameworks as follow: 1) Bayesian Disturbance Injection (BDI), that typifies human behavioral characteristics by incorporating model flexibility, robustification, and risk sensitivity. Bayesian inference is used to learn flexible non-parametric multi-action policies, while simultaneously robustifying policies by injecting risk-sensitive disturbances to induce human recovery action and ensuring demonstration feasibility; 2) Exploiting human action speeds for risk-sensitive interactive IL that performs a robot policy and cedes operation authority to a human in risky states for collecting corrective actions; thus, robustifying a policy while ensuring exploration safety. Both methods' effectiveness is evaluated on a real-robotic reaching task and a risk-sensitive wall-avoidance simulation, respectively. language of the presentation: English | |||
| 谷口 太一 | M, 2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 |
|
title: Waste crane operation planning based on waste surface shape estimation
abstract: Automation of waste cranes in waste incineration plants is an important problem in the real world. However, the waste pits that store the waste are not installed with sensors to observe the shape and other conditions of the inside of the pits. Additionally, the collected waste has a non-uniform density, and the weight does not match the grabbed volume. Therefore, it is difficult to automate the crane operation efficiently because the bucket falls on undulating surfaces and the crane has to re-grab the waste due to lack of weight at low-density locations. This study proposes a framework for crane operation planning based on waste surface shape estimation. The waste pit's internal shape is estimated as a shape representation function based on observation data. This model predicts the shape change due to the grabbing and dropping of waste and plans the efficient waste crane operation. In addition, we developed a simulated bucket equipped with a proximity distance sensor to obtain the suitable volume of the grabbed waste. And, the waste volume is estimated by modeling the waste shape based on the observed waste sparse surface. language of the presentation: Japanese 発表題目: ごみ形状推定に基づいたごみクレーンの動作計画 発表概要: ごみ焼却施設におけるごみクレーンの自動化は実社会における重要な問題である． しかし，ごみを集積するごみピットには，内部のごみ形状など状態を観測するセンサが取り付けられていない．さらに集積されたごみは密度などが不均一であり，掴み体積に対して重量が一致しない． そのため掴み動作の際に，凹凸のある位置でのバケットの転倒や，密度が低い位置で重量不足による掴み直しが発生し，効率的な動作の自動化が困難であった． そこで本研究では，ごみの形状モデルを活用したクレーンの動作計画の枠組みを提案する．これはごみピット内部の形状を，観測データから形状表現関数としてモデル化を行う．そしてこのモデルから，ごみの掴みや落としによる形状変化予測を行い，効率的なごみクレーンの動作計画を行う． 加えて掴んだごみの量を適切に把握するため，近接距離センサを取り付けた模擬バケットを作製した．そして，観測したごみの位置情報から形状をモデル化し，掴んだごみの体積を推定する． | |||