Posts

Bench build closed out. Highlights: RM3100 calibrated, Priority 1, Pixhawk happy (took three units, turns out I'd been soldering at 800F when you want ~600, cooked the first one). Battery is 4S16P, 192 cells, ~14kg. Needed a plate to hold sixteen 21mm cells in a 103mm tube, so I pulled the optimal circle-packing coords from Packomania, two irregular rings, not a tidy circle. All sixteen dropped straight in. Top plate has crush ribs, dialled to 103.75mm (printed a test gauge with four diameters on one part... then still printed a few more to hone it). Deadman STM32 now watches two separate heartbeat wires instead of one shared line, so a broken connector no longer looks like both brains dying. Only fires when both go quiet. And LTE is up on the Jetson, once I fixed the modem's USB mode (NVIDIA doesn't ship the driver in the kernel). Tailscale for the shared-NAT problem, SSH in over cellular from anywhere. #RM3100 #Pixhawk #STM32 #Jetson #4S16P #3dprinting #Tailscale #circlepacking #embedded #AUV

Modem defaulted to a USB mode NVIDIA doesn't ship a driver for in the Jetson kernel, so nothing bound and it insisted the modem didn't exist while sitting right there in the device list. One setting changed, worked instantly. LTE up at 160ms. https://divewithdanu.com/log#entry-25

Danu runs three independent computers by design — not for speed, but for fault isolation. No single failure should take down the whole sub. The Supervisor is an ST Nucleo-F446RE running bare-metal (one hand-written program, no OS to crash). Its only job: failsafes — abort the dive, drop weights, surface, and fire the strobe beacon. The Pixhawk 6C is the flight controller running ArduSub, holding depth and heading via the thrusters. The Jetson is the AI brain — mission planning, cameras, and shipwreck vision — and it's where the experimental code (and the bugs) live. The key rule: the least trustworthy computer has the least authority. The Jetson can't override the Supervisor. All three talk over a tiny gigabit switch that fits inside the 4-inch tube. Cooperation, with a check and balance. 🛠️ #AUV #Robotics #EmbeddedSystems #ArduSub #Pixhawk #JetsonAI #FaultTolerance #MarineRobotics #BareMetal

Milestone on the bench: depth/leak/temp sensors live, first thruster spin proving battery→ESC→motor, and the flight controller ↔ AI computer link up with mission control steering *through* the AI stack. Plus a defective depth sensor that cost us days. Write-up: https://divewithdanu.com/log#entry-24

Danu's 3-brain split isn't for speed — it's fault isolation. Supervisor (Nucleo-F446RE, bare-metal) owns failsafes + strobe. Pixhawk 6C runs ArduSub for depth/heading. Jetson does AI/vision. Least-trusted brain, least authority. https://divewithdanu.com/learn/system-three-brains-that-don-t-trust-each-other

This week Danu's nervous system came alive on the bench. Three sensors reading live: a pressure sensor for depth (the one value she can always trust underwater), a leak sensor whose whole job is to trigger abort-and-surface, and water temp resolved to 0.1°C. First thruster spin proved the full chain end to end — battery bus → ESC → motor, driven by a slider on a laptop. Bigger milestone: the flight controller (reflexes) and the AI computer (decisions) established their first data link, with mission control now seeing and steering *through* the AI computer — exactly the topology for autonomous ops. Building in the open means sharing the floods too: a defective-out-of-box depth sensor that masqueraded as power/wiring/bus faults for days, and a leak sensor that needed one buried config setting its own manual said we'd never touch. Every gremlin was a bench win — and would've been a disaster sealed in a hull. Dry-test everything. 🛠️ #AUV #Robotics #FlightController #ROV #Autonomy #Sensors #SystemsIntegration #EmbeddedSystems #BuildInPublic #MarineRobotics

Requirement #02 is live: depth. Danu's operating target is 300 m — chosen because Lake Michigan bottoms out around 281 m, and most wrecks sit far below the ~40 m scuba limit. Depth here isn't a flex; it's what puts the targets in reach. The rating logic worth stealing: a vehicle's depth isn't set by its toughest part, it's set by its weakest wet one. Danu's pressure tubes are good to ~900 m and lights to 500 m, but the depth sensors, buoyancy foam, and nav sonar top out at 300 m — so 300 m is the number. Shallowest link wins. And a spec sheet is a promise, not proof: every hull, dome, and seal is pressure-tested to >1.5× operating depth (450 m for a 300 m sub) before any real dive. Fittingly, this was the week the first depth sensor came alive on the bench — press on it, watch the pressure reading jump. Requirement, meet hardware. Eight cards to go. #AUV #UnderwaterRobotics #PressureTesting #DepthRating #MarineEngineering #Sonar #Buoyancy #Robotics #SystemsEngineering #DiveWithDanu

Req #02: depth. Target 300 m (Lake Michigan bottoms ~281 m). Rated by the weakest wet part — tubes ~900 m, lights 500 m, but sensors/foam/sonar cap at 300. Every hull tested >1.5× (450 m) before diving. First depth sensor just came alive on the bench. https://divewithdanu.com/log#entry-23

Milestone: full depth-sensing chain is live. A 30 BAR pressure sensor → hair-thin wiring → a small splitter board → the flight computer, with a real depth value landing on screen. Sanity check that actually feels good: a gentle breath into the sensor reads as ~1 m, because the sensor just sees pressure and can't distinguish lungs from a column of water. The compass was the humbling half. We tried a bargain magnetometer — unlabeled pins, fussy voltage, a rat's nest of hand-soldered wires just to enumerate it. We got it working, but "working" isn't the bar for something sealed in a hull bound for cold, dark water. Ordered the proper purpose-built board instead. Sometimes the cheapest part is the expensive one. Depth proven, heading pending. That's the honest shape of building real hardware — one validated sense at a time. 🛠️ #AUV #Robotics #Magnetometer #PressureSensor #Embedded #SensorFusion #FlightComputer #HardwareEngineering #BuildInPublic #Mechatronics

Depth chain validated end-to-end: 30 BAR pressure sensor → splitter board → flight computer, real depth on screen. A breath reads ~1m (sensor can't tell lungs from water). Bargain magnetometer got scrapped for a proper board — heading next. https://divewithdanu.com/log#entry-22

After a month of firmware, USB gremlins, and flight-controller heartbeats, we stepped back and wrote the requirements doc — the whole dream, stated precisely. Every part and every line of code now answers one question: which requirement does it serve? The targets: 300 m operational depth, 3–5 day missions on ~3.6 kWh, 30–50 km round trip at a patient cruise (drag grows with the cube of speed), all in a hull under ~2.2 m and ~40 kg so one person can launch from shore. That last constraint shapes everything — writing multi-day endurance against one-person launch is exactly the tension that resized our battery tube last week, while it was still a purchase order, not a rebuild. Safety is deterministic: the AI advises, but geofence, energy-reserve, depth, and deadline validators decide — the safety chip doesn't even know the AI exists. Fail-to-surface is stacked: drop-weights, independent timer release, strobe, acoustic beacon. These are design targets; reservoir tests grade the homework. Requirements are public, released as numbered cards. #01 drops today. #AUV #Robotics #JetsonNano #SystemsEngineering #MarineTech #Sonar #EmbeddedSystems #RequirementsEngineering #BuiltInTheOpen

Danu's requirements are published: 300 m depth, 3–5 day missions on ~3.6 kWh, 30–50 km range, <40 kg single-person launch. AI advises; deterministic validators decide. Fail-to-surface, stacked. Card #01 drops today. https://divewithdanu.com/log#entry-21

Danu's flight controller is alive. We flashed a Pixhawk 6C with ArduSub — the open-source underwater autopilot — to give it its reflexes: hold depth, stay level, command thrusters, fail safe. The careful part: powering it. Instead of a battery, we used a bench supply with a current limit (0.5 A), then deliberately shorted the leads to confirm the safety net held before trusting real hardware. Power runs through a power module that reports voltage and amperage back to the controller. Snags: the module cable didn't fit the port (we sourced the right one rather than splice and risk the board), and the board went silent over USB — turned out the LAPTOP needed the reboot. Payoff: booted on its own power at ~0.1 A, reporting 14.95 V in QGroundControl, matching the supply to the decimal. #Pixhawk6C #ArduSub #QGroundControl #AUV #flightcontroller #benchtesting #powermodule #embeddedsystems #robotics #buildinpublic

Danu's Pixhawk 6C is up: flashed ArduSub, powered via a power module behind a 0.5A current-limited bench supply (short-tested first). Booted on its own power at ~0.1A, reported 14.95V in QGroundControl — matched the supply exactly. https://divewithdanu.com/log#entry-20

Danu's compute is alive: an NVIDIA Jetson Orin Nano — the board with the big fan in the photos. It'll eventually run the vision models that classify what the cameras see and drive her navigation decisions. Getting there was two jobs: flash Ubuntu onto a fast NVMe SSD (the little stick on the underside), then add the AI stack — CUDA and friends — for quick model inference. The snag: a 256GB USB-C drive wouldn't even show up as a boot option. An old 32GB USB-A thumb drive booted instantly, no network needed. Lesson logged: if your Jetson won't boot the USB, try a smaller plain USB-A stick. After that, AI libraries installed and we reached the board headless over SSH — exactly how we'll talk to it once it's sealed in the hull. Up next: the flight controller, depth & motion sensors, and first thruster spin. 🔧 #Jetson #JetsonOrinNano #NVIDIA #EdgeAI #CUDA #NVMe #Ubuntu #SSH #Robotics #AUV

Danu — engineering log #0. A father–daughter build, fully in the open (engineer by training, dentist by day). The platform: Jetson Orin + Pixhawk/ArduSub + an independent STM32 supervisor, 300 m depth target, LTE + Iridium comms, DVL bottom-lock nav, DIY variable-buoyancy engine. Status: just getting onto the bench — first power-on now. → divewithdanu.com/build

Danu's brain is online: Jetson Orin Nano, Ubuntu on NVMe, CUDA stack for vision. Gotcha: a 256GB USB-C drive wouldn't boot — an old 32GB USB-A stick worked instantly. Now reachable headless over SSH. Next: flight controller + sensors. https://divewithdanu.com/log#entry-18