3D Tactile-Driven Two-Fingered In-Hand Manipulation with Finger-Specific CNN

The use of reasonable robot hands could be practical for industries. An Allegro Hand which is inexpensive and easy to maintain but inferior in performance to some highly sophisticated hands (e.g., TWENDY-ONE Hand) was used in our conventional research. To compensate for inferiority performance, we focused on getting abundant tactile information and curved fingertips which is important for in-hand manipulation. Moreover, such tactile information needs to be processed to provide appropriate feedback for the robotic hand control. To achieve this, we proposed using uSkin, 3-axis distributed tactile sensor and using convolutional neural networks (CNNs). We prepared the model that was trained with measurements from 3-axis tactile sensors, 6-axis F/T sensors and joint angles. In the experiment, the model executed in-hand manipulation with the ball and the cylinder 10 times. As a result, the successful manipulation was achieved 8 times out of 10 for the ball and 7 times for the cylinder.