Presented at IEEE International Conference on Robotics and Automation
The study on "Transfer Learning through Data Augmentations Using Omnidirectional Camera Images" by Kosuke Tahara and Noriaki Hirose was accepted for the presentation at The 2022 IEEE International Conference on Robotics and Automation.
With the progress in deep learning in recent years, robots can be controlled using only camera images. However, dataset collection to cover various environments is one of the potential issues in robotics application of deep learning because it is time-consuming and costly. Our research team focused on the fact that the dataset collection is hardware-dependent as well as on the difficulties in sharing data between different types of robots. Therefore, this study proposes a method where training data could be shared among robots with different degrees of freedom using data augmentations, in which omnidirectional camera images are virtually rotated. We verified that, using data collected from only a compact wheeled robot with 3 degrees of freedom, it is possible to train the control models for an arm robot with 6 degrees of freedom, which is large in size and difficult to carry, and control the arm robot using only the camera images. It is inferred that this method could also be applied to other types of robots, thereby reducing the data collection cost for robot control deep learning.
Title: Ex-DoF: Expansion of Action Degree-of-Freedom with Virtual Camera Rotation for Omnidirectional Image
Authors: Tahara, K., Hirose, N.
Appears in: IEEE International Conference on Robotics and Automation
Presented: May 24, 2022
https://doi.org/10.1109/ICRA46639.2022.9812301