← All projects
Case studyJuly–September 2026Updated 6 October

Underwater ROV: Mechanical Design & Embedded Control

Pool dive test · Synchronized footage, dashboard and attitude visualization

September 2026 · Built prototype · Bench and shallow-water trials

I am developing a modular remotely operated underwater vehicle (ROV) that brings together mechanical design, embedded firmware, live video, and variable ballast. The ROV has progressed from CAD to an assembled prototype with twin thrusters and a browser-based control console connected through an Ethernet tether.

My work spans the complete integration: designing and iterating the printed assemblies, packaging the electronics and plumbing, integrating control and sensing, and using physical tests to guide the next revision. The vehicle is remotely piloted, with sealing, ballast performance, and endurance evaluated through bench and shallow-water tests.

  • Mechanical: a 700 mm hull with an extractable electronics sled, transparent camera dome, and serviceable twin-thruster tail.
  • Embedded and software: a two-ESP32 architecture, Ethernet video and telemetry, differential steering, reversible ballast, and fault-aware controls.
  • Test evidence: working camera transport and control integration on the bench, plus assembled-vehicle propulsion trials in shallow water.
Current ROV electronics sled, battery, flexible ballast bag, pump, and printed twin-thruster tail laid out on the workbench
The current build with the sled withdrawn: electronics, battery, ballast bag, pump, and removable rear assembly.

Designing for access, not just fit

The hull is a 700 mm long PVC tube, 100 mm outside diameter and 94 mm inside diameter. A transparent forward dome gives the camera a view out of the vehicle. Inside, a removable sled carries the electronics, battery, ballast bag, and pump between printed supports and longitudinal rods.

The narrow diameter makes packaging a systems problem: hose bend radii, connectors, cable routing, and maintenance access compete for the same space. I revised the pump installation and rear geometry around those constraints rather than treating the exterior and internals as separate designs.

The current tail combines a straight conical fairing, one dorsal fin, blended motor supports, a pressure-probe cradle, and cable passages. It screws onto a separate mounting interface at the rear plug, allowing the fairing to be replaced without rebuilding the sealed penetrations. The flooded nose and tail fairings shape the exterior; the PVC tube, plugs, dome, O-rings, and penetration seals form the pressure boundary.

Interactive CAD · External design · 6 October 2026

Explore the current build

Rotate the complete external design, including the ballast shells and compact light housing, or inspect the printed end assemblies. Explore the computed flow ↓

Full vehicle

700 mm hull · twin thrusters · ballast shells · light housing

CAD preview: Full vehicle

CAD preview shown. Interactive view loads as you scroll here.

Straight-cone tail

V8 print geometry · single fin · integrated motor supports

CAD preview: Straight-cone tail

CAD preview shown. Interactive view loads as you scroll here.

Dome and nose assembly

Revised fairing · front plug · transparent dome proxy

CAD preview: Dome and nose assembly

CAD preview shown. Interactive view loads as you scroll here.

Full external assembly: V8 straight-cone tail, revised dome fairing, V17 ballast shells, V28 closures and V36 light housing. Purchased components are dimensional proxies. Internal electronics, ballast strips, hoses, wiring, seals, rubber sleeve/clamps and the deployed tether are omitted. This is the design geometry, not an as-built scan or a new print release; the computed-envelope study below is separate. End-assembly views are independently scaled and omit the accessories.

Appearance: dark-grey tube, grey ballast and flashlight holders, white nose and tail assemblies. Download the full external 3D assembly · GLB

Viscous-flow study / 6 October 2026

A wake calculation and a test of the drag claim

Does the light housing actually reduce drag? A controlled, three-mesh SST RANS comparison tests the current cover against the same torch exposed on two simple supports. Everything else stays the same. The detailed full 3D assembly above remains available to rotate.

Computed three-dimensional mean-flow paths around the full vehicle, with recirculation behind both unpowered thruster housings
Computed 3D mean trajectories at 0.30 m/s.

Decision: no drag-reduction claim for this cover

The finest trial returned 0.405 N with the cover and 0.372 N without it. Every trial points toward added drag, not a benefit. However, wall-treatment and convergence checks are incomplete: these are raw screening values, not validated vehicle drag. No percentage improvement or penalty is certified.

Actual RANS mean velocity and modeled turbulent kinetic energy in the horizontal thruster centre plane, including the downstream recirculation region
Top: mean speed and projected in-plane flow directions. Bottom: modeled turbulent kinetic energy, k; the white contour marks zero axial velocity. The wake is produced by the viscous calculation, not added graphically. Its quantitative accuracy is not yet validated.
Comparison, numerical checks and reproducible data

SU2 8.5.0 · incompressible steady SST-2003m RANS · water density 998 kg/m³ · viscosity 0.000998 Pa·s · speed 0.30 m/s · hull Reynolds number 210,000 · zero incidence. Initial/far-field turbulence intensity is assumed to be 1%, with viscosity ratio 1. The 1.5 mm CAD-derived geometry keeps the thruster ducts open but has no propeller blades or powered jets. The tether is excluded.

TrialCovered torchExposed torchChange
Coarse · 10.67 mm0.433 N0.423 N+0.010 N
Medium · 8 mm0.406 N0.372 N+0.033 N
Fine · 6 mm0.405 N0.372 N+0.033 N
Larger domain · 8 mm0.410 N0.366 N+0.045 N

All entries are unvalidated numerical trials, not measurements. Nominal near-body mesh size is listed, not a uniform cell size. Reference area: 0.02040375 m². The larger domain extends each box boundary by a factor of 1.5; its remeshing also contributes to the difference.

These are bounded trials, not a claim that every solution converged. Medium, fine and enlarged-domain runs were stopped with coherent saved fields after the mesh/wall-validation shortcomings became clear; force drift and all residuals remain in the downloadable audits. All eight tabulated cases start from uniform flow, not from a restart.

Raw covered and exposed drag results across three meshes and a larger domain, plus fine-mesh force histories
  • Solver/setup checks: integrated laminar flat-plate friction differs from Blasius by 0.55% on the refined grid (2.20% on the coarser grid). A wall-resolved turbulent flat-plate sample differs by 0.22% from the TMR incompressible reference. That uses a different SST variant and mesh: it is a useful comparison, not experimental validation.
  • Mesh and force checks: all eight vehicle meshes have positive cell quality. The fine covered mesh has 177,805 vertices and 1,048,647 tetrahedra. Independent surface integration agrees with the fine-case drag coefficients within 0.07%.
  • Global flow balance: independent integration of the exported fine-case velocity over the outer boundary leaves a net volume-flux imbalance below 0.007% of nominal throughflow in both configurations. This is a post-processing balance check, not the solver's exact dual-volume residual.
  • Why the drag claim is not accepted: the coarse turbulence residuals did not settle; the mesh sequence does not establish asymptotic convergence. About 63% of the fine covered wall area has 5 < y+ < 30, with no prism layer. Wall-function adequacy, geometry sensitivity, transition and turbulence-model uncertainty remain unverified. No GCI error bar is asserted.
  • Boundary check: enlarging the domain changes the covered and exposed trial forces by +1.20% and -1.71% respectively. That combines boundary and remeshing sensitivity, not a clean isolated uncertainty bound.
  • Visualization check: 130 of 131 candidate trajectories stayed inside the fluid mesh when their display segments were checked at ≤0.5 mm intervals. Halving the integration step from 2 to 1 mm changed matched trajectories by at most 0.42 mm. No random eddies or extra transverse velocities were added.

The local solver build includes a disclosed OpenMP call-site correction to avoid a wall-function threading deadlock; the patch and build hashes are included. Original CAD and all earlier calculations are preserved. The compact download contains source, input geometry, benchmark meshes, all case configurations/histories/force audits, surface solutions and sampled flow fields. Original vehicle volume meshes and the full fine covered volume solution remain in the complete local record; regeneration settings are supplied. Failed boundary-layer meshing attempts are logged locally and were not used as successful meshes.

SU2 release · Turbulence Modeling Resource · NASA grid-convergence guidance

Download the calculation record and reproducible inputs · ZIP · Machine-readable assessment · Trial force table

What would establish drag, stability and pressure accuracy?

Drag benefit: complete a near-wall-resolved, model-sensitive CFD comparison or perform paired tow tests. Tow the covered and exposed configurations at identical measured speed, trim and immersion, with motors off. Alternate the order, repeat each configuration at least five times, exclude acceleration periods, and measure/allow for fixture and cable drag. Compare the paired mean force difference and its uncertainty. An apparent improvement smaller than the combined uncertainty is inconclusive. These CFD trials suggest a difference of only about 0.03 N: an ordinary luggage scale is unlikely to resolve that reliably.

Handling: measure total mass, CG, buoyancy and CB in the final ballast configuration, then record small-angle pitch/yaw release tests and controlled steering tests with the actual tether. Decay versus growth of the motion establishes behavior for that tested condition. One zero-incidence, motors-off force calculation cannot demonstrate dynamic stability or closed-loop handling.

Pressure probe: hold the vehicle at a fixed, independently measured depth and compare repeated readings with motors off and at the intended thrust settings, including asymmetric steering. Repeat at several depths with the actual port, plumbing and mounting. Convert the motor-correlated pressure change to equivalent depth error using Δh = Δp/(ρg), and compare its bias and scatter with a predeclared accuracy target. A modeled external pressure or freestream dynamic-pressure estimate is not a bound on powered-thruster sensor error.

No physical tow, stability or motor-on pressure-validation test was performed in this study. These tests require the real vehicle; the files contain blank measurement templates rather than invented readings. Use a secured, leak-tested vehicle in a controlled test area, keeping people and loose lines clear of thrusters.

The earlier potential-flow study below is retained separately. Its surface acceleration map is not used to calculate viscous drag or turbulent wakes.

Hydrodynamics / computed flow

Where the water accelerates

A 3D potential-flow solution around the CAD-derived vehicle envelope. Streamlines follow the calculated velocity field; surface colours reveal acceleration and pressure changes around the dome, ballast pods, light housing and tail.

0.30 m/sForward-flow condition
10,000Triangular source panels
0.78%Fine-grid velocity change · RMS
Computed envelope · speedPotential flow · U∞ = 0.30 m/s
Computed water-speed contours and streamlines around the CAD-derived submarine envelope at 0.30 metres per second.
Water speed · m/s0.000.30≥ 0.60

Computed static results shown. Interactive results load here.

What the calculation tells us

The light housing introduces a second high-pressure region below the camera dome. Flow accelerates beneath the housing while the upper mid-hull stays close to the imposed speed. These are useful locations to inspect when refining the pod transitions, not a measured drag reduction.

Fixed fluid probeSpeedCpPressure change*
Ahead of camera dome0.143 m/s+0.774+34.76 Pa
Ahead of light housing0.148 m/s+0.756+33.95 Pa
Below light housing0.322 m/s-0.154-6.92 Pa
Above mid-hull0.303 m/s-0.019-0.84 Pa

*Relative to undisturbed water at the same height. Probe points are 9–14 mm outside the envelope, not wall values. Their exact coordinates and values are in the results package. At this condition, q∞ = ½ρU∞² = 44.91 Pa; Cp = 1 − (|u|/U∞)².

Follow-up: the viscous-flow comparison above tests the light-cover drag claim. It does not demonstrate a benefit, and its near-wall/convergence checks do not yet support a validated drag estimate. The potential-flow map remains an acceleration study; handling and powered-probe accuracy require the separate tests described above.

Numerical checks, assumptions & reproducibility

Steady, incompressible, inviscid flow; zero incidence, motors off, unbounded water. The calculation uses a 2 mm CAD-derived envelope with small fittings removed and cavities simplified. It resolves external displacement flow, not viscous separation, turbulent wakes, propeller jets or tether motion. Colour limits are clipped for readability; raw values are included in the download.

The solver represents the closed surface with constant-strength triangular source panels. Exact edge-logarithm and solid-angle influence integrals are used to solve the no-penetration condition at panel centroids: ∇²φ = 0, u = U∞ + ∇φ, n · u = 0. Surface pressure follows Bernoulli’s equation.

CheckResultInterpretation
Sphere analytical benchmark1.16% RMS velocity error / U∞1,280 panels; 915 probes at radii 1.05R, 1.25R and 2R.
AUV: 3,000 → 5,000 panels1.01% RMS field change / U∞Same 1,000 fluid probes for both solutions.
AUV: 5,000 → 10,000 panels0.78% RMS; 4.31% maximumProbes offset 10 mm from the reference mesh; at least 8 mm clearance.
Linear-system relative residual6.33 × 10⁻¹⁵Algebraic solve accuracy, not physical validation.
Net source / (U∞ × surface area)−1.08 × 10⁻⁴A global source-balance diagnostic; ideally zero.
Streamline step-halving0.30 mm maximum path change10 trajectories, including the four nearest the surface; 3.16 → 1.58 mm integration steps.

Mesh refinement tests the chosen envelope, not its fidelity to the manufactured vehicle. Sharp-corner surface peaks remain resolution-sensitive, so no peak-pressure or force claim is made. The surface map uses panel-centroid values. Of 64 upstream seeds, 54 throughflow paths retain more than 2 mm surface clearance and positive streamwise velocity; only these are displayed. They are integrated through the same solved field; moving particles use the computed travel time.

Water density: 998 kg/m³. The hull-length Reynolds number is about 210,000 using ν = 10⁻⁶ m²/s; viscosity and boundary layers are nevertheless omitted by this first-pass model. No empirical wake or drag has been added to the flow field.

Method references: MIT · potential-flow boundary conditions and Bernoulli; NASA TP-2995 · panel methods.

Download the solver, meshes, results & checks · ZIP

Optional calculations, assumptions & test plan

Engineering study · 6 October 2026

Designing for steady footage, not just low drag

The mission is slow, controlled underwater video. I evaluated the complete external design above (nose, straight-cone tail, twin thrusters, wraparound ballast shells and compact light housing) alongside the rear-exiting Ethernet tether.

Separate drag-sensitivity estimates. Geometry is calculated from the assembly. Drag coefficients are deliberately varied assumptions; the resulting loads are scenarios, not measured performance or validated bounds. Pool observations inform the design but do not validate these calculations.

External design envelope
914 × 258 × 162 mm

Length × width × height; no tether

Projected frontal area
0.0204 m²

Union of the CAD silhouettes

Nominal ballast authority
4.90 N

Weight change for 500 mL of water

The cable is part of the vehicle

The tether is 3.8 mm in diameter and 150 ft (45.72 m) long. For a straight 10 m submerged section perpendicular to a 0.30 m/s relative flow, the selected cable coefficient of 1.2 gives 2.05 N of distributed drag. Exposing the entire length to that same crossflow gives 9.36 N. By comparison, the axial vehicle scenarios at 0.30 m/s give 0.27–1.10 N.

This comparison identifies a design priority, not a connector-load prediction: deployed length, current and cable angle can outweigh refinements to the fairing. A curved tether shares loads between its endpoints, and its force need not act in the same direction as vehicle drag. The two values must not simply be added.

Explore the load sensitivity

Assumed drag coefficients · parameter study
Vehicle axial drag · CD 0.3–1.20.27–1.10 N
Cable normal drag · Cn 0.8–1.61.37–2.73 N

Hull Reynolds number: 210,000. Cable crossflow Reynolds number: 1,140.

The same uniform speed is used for both examples. Body flow is axial; cable angle is independent. At 0°, only this model’s normal cable term vanishes; axial cable drag is not zero in reality. Coefficient ranges are selected for sensitivity, not calibrated to this vehicle or cable.

Equations, reference areas and numerical checks

D = ½ ρ U² CD A for the vehicle; Fn = ½ ρ Cn d L (U sin α)² for a uniform straight cable. Freshwater is represented by ρ = 998 kg/m³ and ν = 1.0 × 10⁻⁶ m²/s. The drag-equation reference explains the coefficient/area pairing; Orcina’s Morison-element formulation separates normal and axial flow.

Frontal area is the union of projected mesh triangles, not a sum of overlapping component areas. Refining the raster from 0.25 to 0.125 mm changes it by 0.52%. That checks numerical resolution only; purchased-component proxies and omitted hardware still limit physical accuracy. Side and plan projections are 0.1056 and 0.1129 m². These are not wetted surface areas, and the selected axial coefficient cannot be reused for sideways motion.

The 700 mm hull Reynolds number spans 70,000–560,000 for 0.1–0.8 m/s. Transition, print roughness, steps, through-flow and thruster operation are unresolved. Cable Reynolds numbers are much lower; a coefficient of 1.2 is a nominal scenario, not a universally applicable cylinder value. Added mass, transient damping, free-surface waves, wake interaction and tether dynamics are outside this reduced model.

Trim comes before more fins

The 3.42 kg of nominal fixed weights are distributed in low wraparound holders. That placement is intended to lower the center of gravity, but the complete vehicle’s center of gravity and center of buoyancy still need measurement. A level float test alone does not establish a stability margin.

For a fully submerged vehicle with vertically aligned centers, a center of gravity below the center of buoyancy produces a pitch/roll restoring moment: approximately M = −B h sin θ, taking moments about the center of gravity. Here B is buoyancy force, h is the vertical separation at level trim, and θ is the tilt angle. Longitudinal ballast placement sets trim. This is a hydrostatic relationship, consistent with Fossen’s marine-craft formulation; it does not prove dynamic stability or heading hold.

The aft dorsal fin can oppose sideslip during forward motion, but it is not a pitch-control surface. Fin forces fall approximately with speed squared and provide little authority near hover. Extra front canards are therefore not the first improvement for steady slow footage. At this stage, measuring trim and reducing tether disturbances are more useful than adding decorative surfaces.

The CAD thrust axes are symmetric, 172 mm apart and at the hull centerline height. Equal thrust cancels their ideal yaw moments; a thrust difference ΔT produces a yaw moment magnitude of 0.086 × |ΔT| N·m when ΔT is in newtons. Vertical alignment with the actual center of gravity is unverified, so equal thrust may still pitch the vehicle. Duct/body interference and unequal motor output need a water test.

What the ballast can and cannot change

The 100 mm × 700 mm cylindrical hull envelope is 5.50 L; its gross internal bore is 4.86 L before the sled, battery and hardware. Neither number is the displacement of the whole submarine. The flooded fairings and ballast shells do not displace their entire enclosed envelope: water must be able to enter and trapped air must escape.

Taking 500 mL of water into an internal bag adds approximately 0.499 kg without enlarging the rigid outer hull, reducing net upward force by 4.90 N. That is the nominal maximum, not a measured fill. Trim the empty-bag configuration slightly positively buoyant, then measure how much water the installed pump really transfers.

Why the sealed hull can affect filling

At the surface, slow isothermal compression of initially atmospheric free air gives Δp = p₀ ΔV / (V₀ − ΔV). Filling 0.5 L raises hull pressure by about 34, 20 or 14 kPa for initial free-air volumes of 2, 3 or 4 L respectively. These are examples, not measured free-air volumes. The gross bore volume is not the available air volume.

The pump sees the difference between bag/hull pressure and outside water pressure, plus hose and bag losses. Depth helps the filling pressure balance but makes emptying against the surrounding water harder. Kinks, suction-side hose collapse and pump performance must be checked separately. A nominal bag capacity cannot establish usable buoyancy authority.

Geometry decisions and the next tests

Keep the packaging low

The wraparound holders organize the weights without large downward boxes. Their steps, straps and closure bridges still create disturbance. Keep straps tight, tails trimmed, seams aligned and all weights positively retained.

Protect the inlet flow

The two thrusters set the maximum width. Preserve their intake clearance and avoid routing cable or straps across the openings. The smooth cone is a design intent; attached flow and reduced drag have not been demonstrated.

Measure static pressure

The rear probe is close to powered flow. At 0.30 m/s, one dynamic-pressure head is about 4.6 mm of water; at 0.50 m/s it is 12.7 mm. Local jets may cause a larger bias. Compare motor-off/on readings at fixed depth before claiming depth accuracy.

Manage the tether first

Deploy only the length needed, provide strain relief to the structure rather than the RJ45 connector, and keep the line out of the thrusters. A rear exit does not eliminate pitch or yaw moments from a sideways-pulling tether.

Validation plan and limits
  1. Finish leak and retention checks without vulnerable electronics, then confirm that the free-flooding shells do not trap air. This study assigns no safe depth.
  2. Measure assembled mass, trim, center-of-gravity position and repeatable bag fill with the hull closed. Check slightly positive buoyancy with the bag empty and record the full-bag response.
  3. Measure steady unpowered tow force at 0.2, 0.3 and 0.5 m/s, using a calibrated load cell and controlled relative water speed. Repeat each case at least three times; account for the tow line, current and surface proximity.
  4. At fixed depth, log pressure and attitude with motors off and then at controlled thrust settings. Check both propulsion directions and tether angles, using a safe restrained setup.
  5. Fit drag and moment coefficients to those data before predicting maximum speed, power, endurance, control margins or adding more fins.

Pool testing has qualitatively established a workable weight placement. It has not yet supplied the force, flow and mass-property measurements needed for a validated six-degree-of-freedom model. The renders show a design assembly, not a computational fluid solution.

From CAD to structural analysis

The structural study examines how external pressure loads the hull, dome and printed closures, and how the twin-thruster supports carry an offset thrust load. It uses the current CAD geometry with Gmsh quadratic solid meshes, CalculiX elastic analysis and Python post-processing.

Calculated stress fields mapped onto the assembled ROV, including the hull, front and rear closures, and twin motor supports; logarithmic von Mises scale in MPa
Pressure case
147.1 kPa

15 m freshwater differential

Closure load
1.155 kN

Equivalent axial end load

Mounting case
20 N / motor

Axial load with bending

Follow the load path

The 700 mm PVC hull has a 100 mm outside diameter and a 3 mm wall. The pressure case combines exterior pressure with the equivalent closed-end load. At the hull midspan, the inner-wall von Mises stress is 2.19 MPa and inward radial movement is 0.028 mm.

At the front, pressure on the 3 mm acrylic dome transfers through its flange into the printed socket. At the rear, the plug transfers the equivalent end load through its seating shoulder. Idealized attachments isolate each component's response; O-rings, cable potting, adhesive and the rubber sleeve/clamps are not represented by a contact or leakage model.

The tail is flooded, so it does not carry the sealed hull's pressure differential. Its separate load case includes the 41.68 mm offset between the thrust axis and mounting seat, represented by a force and bending couple. This is a mounting study, not a measured thruster limit; impact, torque and fastener contact remain separate questions.

Model verification

The hull's midspan stress agrees with the analytical cylinder solution within 0.054%. Force/reaction balance closes within 0.001%, and the sampled quadratic-element Jacobians are positive. Refinement changes maximum displacement by 0.03% for the dome, 11.8% for the front closure, 1.8% for the rear plug and 3.9% for the tail. The front closure remains mesh-sensitive; local stress peaks are not treated as converged fracture predictions.

Material and modeling assumptions

Representative short-term elastic properties are used: PVC E = 3.2 GPa and acrylic E = 2.8 GPa. The dome's material and 3 mm thickness are confirmed; its 80 mm diameter and flange shape follow the existing CAD.

Printed parts use an approximate 1.25 mm skin with E = 2 GPa around a weak core with E = 0.2 GPa. The core stiffness is an assumed continuum value, not a property inferred from the rear plug's 10% infill. Layer adhesion, anisotropy, toolpaths, creep and water aging are not explicitly resolved. Skin/core assignment and ideal restraints influence local stress patterns.

15 m is an analysis case, not a qualified operating depth. These elastic results do not establish a breaking depth. External-pressure buckling, printed-layer failure, joint retention and sealing still require dedicated analysis and physical validation. A preliminary generic hull buckling screen is sensitive to ovality and load duration, so low elastic displacement alone does not demonstrate safety.

Components and design decisions

I prioritized affordable, available modules that could fit the hull and support incremental testing. The choices below describe the current prototype, not the earlier proposed component list.

Hardware choices, their purpose, and the constraints they introduce
ComponentWhy it fits this design
ESP32-WROOM DevKit + ESP32-CAMA compact, low-cost split between camera acquisition and the main controller. The ordinary ESP32 provides accessible GPIO for control and sensing; a dedicated SPI link carries camera frames to it. This avoids adding a Linux single-board computer, at the cost of limited video throughput.
W5500 Ethernet interfaceOne wired tether carries video, telemetry, and commands without relying on an underwater Wi-Fi connection. The trade-off is cable drag and a communications path that needs explicit disconnect recovery.
2 × APISQUEEN U2 Mini thrusters + independent ESCsCompact ducted propulsion with differential steering: varying left and right thrust avoids adding a moving rudder linkage. Their size and mounting loads drove several tail and motor-support revisions.
ZC-A210 reversible gear pump + bidirectional motor driverThe smaller pump fits the constrained internal bay and allows one pump to fill and empty the ballast bag. Installed flow and pressure capability still need characterization with the hull closed.
Nominal 500 mL flexible ballast bagA low-cost reservoir that conforms to the available space. Adding water changes vehicle mass without changing external volume, but the expanding bag compresses the sealed hull air and needs kink-free plumbing.
4–20 mA pressure probe + current-to-voltage converterA packaged submersible probe keeps the pressure-sensing element outside the hull. The converter makes its current output readable by the ADC; the assembly is bulky and requires calibration against known depths.
ADS1115 ADC + battery divider + leak probeOne external ADC collects the conditioned pressure signal, divided battery voltage, and a simple two-electrode water-detection input. Shared diagnostics and saved calibration make faults more visible; the leak probe is a backup, not a seal or power isolator.
MPU6050 inertial sensorA compact, inexpensive source of acceleration and angular-rate data for attitude monitoring in the piloting console.
3S battery + regulated 5 V UBECBattery power serves the actuators while the UBEC supplies the low-voltage electronics. This separates voltage requirements, not electrical grounds; power integrity remains part of integrated testing.
PVC hull + PLA printed assembliesA readily available cylindrical enclosure and fast, inexpensive geometry iteration on a FlashForge Dreamer. Removable interfaces reduce rework, while sealing surfaces and printed parts still require physical validation.
Electrical architecture: battery and UBEC power; laptop through Ethernet and W5500 to the main ESP32; a dedicated SPI camera link; ESC and ballast driver outputs; ADS1115 battery, pressure and leak inputs; and a separate MPU6050 I2C bus.
Open full-size schematic

From camera frames to a piloting console

The main ESP32 handles Ethernet, control, and sensor acquisition; the ESP32-CAM sends captured frames over a dedicated internal SPI link. A local laptop service presents the live feed, battery and depth readings, sensor state, and commands in a browser.

Measured bench result: two receiver tests totaling 500 frames delivered 800 × 600 JPEG video at 7.42–7.56 fps, with a maximum arrival gap below 153 ms. This characterizes the camera transport under the recorded bench conditions, not end-to-end display latency or in-water endurance.

The operating console supports W/A/D steering, on-screen controls, and hold-to-fill/empty ballast commands. Releasing a control, leaving the active window, or losing fresh communication clears motion commands. Reconnection requires new input rather than resuming a previous movement.

Battery voltage is calibrated against a multimeter and retained in non-volatile storage. The pressure input has a saved atmospheric zero; its depth scale remains provisional until checked against known immersion depths.

Engineering the failure cases

I added a conductive leak probe and a visible water-detected warning. A wet reading latches a software stop for the thrusters and pump; it does not disconnect the battery. Network recovery also preserves the leak latch and clears stale motion commands.

Intermittent connection failures led to a more explicit recovery sequence: invalidate commands, stop outputs, rebuild the Ethernet services, and initialize the ESC signal channels at idle before accepting fresh input. A controlled console-outage test restored video and the idle controller state without rebooting either ESP32. Further battery-powered testing is needed to resolve the original intermittent fault and validate recovery with powered actuators.

The ballast system exposed a different integration problem. Early filling tests moved less water than expected; later tests improved. I am checking excess service-loop tubing for kinks and priming restrictions, alongside the backpressure caused by compressing air in the sealed hull. There is no fitted flow or bag-volume sensor, so run time is not reported as a measured fill percentage.

Built, tested, and still improving

The assembled grey PVC underwater vehicle in a shallow inflatable pool, with water flowing behind its two rear thrusters
Assembled-vehicle propulsion testing in shallow water.

The pool trials put the packaging, propulsion, and seals into the same physical test. They demonstrate an assembled vehicle with operating twin thrusters, not just a CAD concept or isolated electronics.

Immersion checks also revealed small leaks, with cable penetrations appearing to be one source. Reworking and isolating those paths is now a validation task. The vehicle does not yet have a qualified operating depth.

Next milestones: repeatable dry-hull immersion tests, measured fill/empty performance with the hull closed, longer battery-powered runs, and controlled trim and depth-calibration checks.

The main outcome so far is a serviceable electromechanical prototype and a testable control stack. The project has sharpened how I work across subsystem boundaries: turning packaging constraints into CAD revisions, turning observed faults into firmware behavior, and distinguishing a demonstrated result from a design assumption.

Continue exploring

More selected work

↗