The Georgia Tech Miniature Underwater Robot (GT-MUR) is a small, low-cost robot designed to support research for underwater robotics. The GT-MUR has a versatile and modular design. It is capable of communicating through several modes, including Wi-Fi, radio frequency (RF), acoustic, and infrared radiation (IR). The platform has an inertial measurement unit (IMU), depth sensor, two vertical and two horizontal motors and propellers, and a forward-facing monocular or stereo camera. Our lab has a pending patent for the GT-MUR:
Q. Tao, J. Cha, X. Chen, S. Maxon, C. Qin, L. Seguin, H. Xie, J. Y. Zheng, F. Zhang, “Miniature Underwater Robot for Research and Education,” Provisional patent (62/669,571) filed on May 10, 2018.
The GT-MUR was one of the first robots I have ever worked on. During my first semester in the lab, I created a prototype for the motor control module, which involved programming mbed microcontrollers to communicate with motor electronic speed controls (ESC), along with laser cutting the GT-MUR’s acrylic housing and rubber pads.
An early prototype of the GT-MUR’s motor control capability.
Another project that I worked on with the GT-MUR was implementing a deep learning approach to detecting holes in fishnets underwater. This has applications to fish farm monitoring, as a hole in a fish enclosure can result in large monetary losses for the fish farming company, and could also lead to disastrous environmental consequences. Additionally, employing a diver to constantly check the health of these large fish farms can be expensive. The GT-MUR would be a low-cost solution to autonomously monitoring the status of the fish farm.
This was the first deep learning project that I was able to lead for creating a proof of concept to tackle a real world problem. To implement a deep neural network on the GT-MUR, which has low compute power, we used the Movidius neural compute stick, which is small enough to fit onboard the GT-MUR. I also had to build a test rig for capturing data of a fishnet inside a pool, and organize labeling efforts among students in the lab. We found that a deep convolutional neural network (CNN) is very capable of detecting holes in fishnets, likely because nets have very consistent patterns, and any significant deviation from this pattern could potentially be a hole.
Example of a CNN detecting a hole in a fishnet underwater.