Marlin V1’s hull is a 9.5-inch diameter acrylic tube, 25.875 inches long, and sealed by two anodized 6061-T6 aluminum endcaps.

The front endcap is sealed to the end of the tube with 3M DP420 epoxy, and the front camera dome is sealed using a face o-ring. An aluminum collar was added to the back of the tube, so that the back endcap seals to aluminum instead of casted acrylic, ensuring seal integrity.

The back end cap uses o-rings 3% smaller than normal specification and was given a 10-degree chamfer to ease insertion and extraction of the electronics rack. Four spring-draw latches on the rear of the hull help keep the back endcap seal watertight by preventing the endcap from sliding out of the tube. Thirteen waterproof SubConn connectors interface between Marlin V1’s internal and external electronics, and can be quickly detached and reattached without disrupting the watertight seal. Two aluminum mounts attach the hull to the horizontal plane of the submarine.

Marlin V1 maneuvers using eight VideoRay M5 thrusters, chosen for their high thrust and efficiency. Four are arranged horizontally at 45-degree angles and control strafe, yaw, forward, and reverse movements, while the other four are oriented vertically and control depth, pitch, and roll. The thrusters use up to 750W and generate up to 23 lbs of thrust.

The frame is built from 6061-T6 aluminum components for high strength and relatively low weight. Sheets and L-pieces are machined using CNC machines and standard industrial tooling machines, and then anodized with a type II coating. A horizontal plane provides secure mounting points for the hull mounts, surge thrusters, dropper, torpedoes, and many other components, while four vertical side panels support the weight of the plane and provide mounting points for the vertical thrusters, grabbers, and hydrophones. Additional panels were added in the front and back of Marlin V1 to avoid warping the horizontal plane under load.

To manage the complex software system with multiple collaborators, the team uses Git version control. Git allows the software team to work on the same files independently while keeping backups of previous versions of the code. Marlin V1’s software stack runs on the Debian OS (GNU/Linux). Dubbed Aquastorm, it consists of multiple processes that communicate over a shared mutex and named pipes. Each process works independently of the others, while the mutex ensures thread safety. Most software is written in C/C++ and Python.

Interface is the master program that starts each process on the software stack. Each process is run in a separate thread toOur low-level controller runs on the ATMega2560 microcontroller. It connects to Aquastorm, the AHRS, and the thruster controllers via serial UART pins. It uses six PID filters, one for each bearing of the submarine. The resulting PID vector is multiplied by the thruster matrix, mapping the effect of each thruster on each element of the state, to produce the desired thruster power. The state consists of the (X, Y) position, computed by integrating the thrust vectors, the depth, determined from the analog pressure sensor, and the roll, pitch, and yaw from AHRS. Control sends the current state to the control process in Aquastorm, and receives the desired setpoint. It also activates the relays on command. Because of the simple serial protocol, Marlin V1 can be controlled manually through a terminal for testing and debugging use. Furthermore, all configuration variables (e.g., PID gains and the thruster matrix) can be adjusted over serial communication to empirically tune the optimal values on the fly.

In order to robustly navigate obstacles and account for imprecise positioning and observations, Modeling localizes the sub to its environment. Our modeling algorithm uses Monte Carlo Localization, which initially constructs a uniform particle filter. Each particle represents a possible state of the sub. As the particle filter is first set to be mostly uniform, the sub has little knowledge about its surroundings. However, as the submarine moves around its environment, it is able to resample its particles with motor and sensor updates. Eventually, the particles converge upon the actual location of the submarine, leading to successful localization.

Each competition task is associated with a vision function within Aquastorm, that returns an observation based on what objects it detects. Most vision processes initially run preprocessing to generate more contrast in the image. Afterwards, they use a variation of machine learning, color mapping, edge, contour, and blob detection with the aid of the OpenCV library. Our machine learning algorithms utilize the Tensorflow Object Detection API to detect objects within an image.

Mission generates a list of goals to complete throughout the competition run. Each goal contains a set of instructions, location, and point value. Based on the location of both the sub and the goal, we can compute an estimated time to completion for each goal, and visit the goals in an optimal fashion, maximizing our point value. Once we select a goal to complete, mission tells the corresponding vision function to start returning observations.

Otolith, written using LabVIEW, runs on the FPGA and reads from the four hydrophones located outside the main hull. The analog-to-digital converter (ADC) receives amplified signals from the hydrophones at 500 thousand samples per second. Using each analog signal, Otolith applies a Butterworth low pass filter to remove extraneous noise, then looks for a ping of the designated frequency. The time at which the ping is detected is recorded for each hydrophone, then the time difference of arrival (TDOA) is computed and sent to the main computer for localization. On the main computer, a small C program receives the TDOAs and computes the location of the pinger using trigonometric ratios between the TDOAs; this method, while introducing a slight built-in error (which decreases as the sub points towards the pinger), is less susceptible to slight errors in physical measurements and ping detection when compared to multilateration (solving the system of hyperboloids), making it much more robust. It sends to Interface the corresponding Gaussian distribution.