2023 Fiscal Year Final Research Report
Multi-Fingered In-Hand Manipulation Based on Multi-Type Tactile Information Using Deep Learning
| Project/Area Number |
22K17979
|
|---|
| Research Category |
Grant-in-Aid for Early-Career Scientists
|
| Allocation Type | Multi-year Fund |
|---|
| Review Section |
Basic Section 61050:Intelligent robotics-related
|
|---|
| Research Institution | Waseda University |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2022-04-01 – 2024-03-31
|
| Keywords | 触覚センサ / 多指ロボットハンド / 深層学習 |
|---|
| Outline of Final Research Achievements |
In the first year, we proposed a model that combines Transformer and LSTM, applying it to the tactile sensors and joint angle information of the multi-fingered hand. As a result, the model learned to focus on important modalities such as tactile and joint information, successfully adapting its motions to unknown grasping positions and object characteristics. In the second year, to achieve more dexterous manipulation with the multi-fingered hand, we developed and improved high-resolution optical tactile sensors. Human fingertips are characterized by a near-spherical shape at the tip, while the shape of the finger pads becomes flatter when closer to the first joint. We confirmed that the fingertip shape is easy to manipulate while still capturing fine-grained grasping states.
In the next six months, we will advance the research on a motion generation model that integrates the developed deep learning model and optical tactile sensors.
|
| Free Research Field |
知能機械学
|
| Academic Significance and Societal Importance of the Research Achievements |
本研究では，各触覚センサの特徴や限界を逆手に取り，人の手上の固有受容器の分布を模倣するため，光学式触覚センサのような高い解像度を持つがスペースを取るセンサを指先に，センシング点は少ないが薄く容易に貼り付けられる磁気式の触覚センサを指腹や掌に搭載した．さらに重要なこととして，今まで深層学習は画像や関節角度などの情報をマルチモーダルに処理してきた実績があるが，触覚単体でも，様々なモダリティが混在する難易度の高い情報であり，しかしながら人間のように巧みな操りを行う上では欠かせない観点と考えた．これにより人間の作業を行うことを目指すことで学術的意義のみならず社会的意義も見出そうとした．
|
