2020 Fiscal Year Research-status Report
Deep Learning for Planetary Rover Localization
| Project/Area Number |
20K14706
|
|---|
| Research Institution | National Astronomical Observatory of Japan |
|---|
Principal Investigator |
Wu Benjamin 国立天文台, アルマプロジェクト, 特別客員研究員 (50868718)
|
|---|
| Project Period (FY) |
2020-04-01 – 2024-03-31
|
| Keywords | machine learning / computer vision / deep learning / space exploration |
|---|
| Outline of Annual Research Achievements |
We have configured and installed a new GPU server at NAOJ which can be used for accelerated machine learning.
We are developing a new neural network that can take as input monocular panoramas from the lunar surface and output location coordinates on a corresponding satellite map. Thus far, we have built data generation tools and have spent many CPU and GPU hours to generate large datasets (1.6 million images) using simulated satellite maps and elevation models. I built prototype Siamese neural networks that are able to learn if two surface-perspective images are within a certain distance from one another. I presented results from the deep learning rover localization project at the i-SAIRAS conference in Oct 2020. Currently, 1 paper is in preparation.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The GPU server setup proceeded smoothly and several other NAOJ researchers are already taking advantage of the new compute capabilities.
Our new data generation tools are working well. The rate of data generation took slightly longer than expected because the software we used (Blender) for one of the datasets does not render in real time like the software used for the paired dataset (Unreal Engine). However, I was still able to build prototype models using partial data. I am currently developing the neural networks to be able to take advantage of the full dataset. I was happy to be able to present our previous and preliminary results at a conference (i-SAIRAS) so quickly.
|
| Strategy for Future Research Activity |
I plan to expand the current Siamese convolutional neural network in order to utilize the full resolution of the images and also take advantage of the larger dataset. This should improve performance over the prototype models and allow for more accurate localization. Once the model achieves good performance, we will tune the hyperparameters and perform ablation studies in order to understand the effects of various design choices in the network. I plan to finish and submit our paper to a robotics and machine learning publication.
This method uses 2D images from monocular cameras, but we can extend the work further to 3D point clouds from stereo cameras or lidars. This may be explored in future projects.
|
