Most soft-body organisms found in nature exist in underwater environments. It is helpful to study the motion and control of soft robots underwater as well. However, a readily available underwater soft robotic system is not available for researchers to use because they are difficult to design, fabricate, and waterproof. Furthermore, submersible robots usually do not have configurable components because of the need for sealed electronics packages. This work presents the development of a submersible soft robotic arm driven by hydraulic actuators which consists of mostly 3D printable parts which can be assembled in a short amount of time. Also, its modular design enables multiple shape configurations and easy swapping of soft actuators. As a first step to exploring machine learning control algorithms on this system, two deep neural network models were developed, trained, and evaluated to estimate the robot's forward and inverse kinematics. The techniques developed for controlling this underwater soft robotic arm can help advance understanding on how to control soft robotic systems in general.