CoralSight is an autonomous underwater vehicle being built to inspect, photograph and listen to coral reefs — and to flag bleaching before it is visible to the eye. An 8″ pressure hull, four ducted thrusters with no thrust ever aimed at the reef, and 67 TOPS of edge-AI compute running the classifiers on board.
Marine heatwaves, coral disease and local stress are reshaping reefs season by season. Traditional survey methods — divers, tethered ROVs, boats with trailing sensors — are slow, infrequent, and can be physically damaging to the fragile polyp colonies they try to measure.
CoralSight is being built to change that equation: a compact, neutrally trimmed AUV that images reefs with fluorescence and calibrated colour, listens to them with a hydrophone, and reports what it finds — with every safety limit designed in rather than bolted on.
Every subsystem is chosen to survive seawater, respect coral, and produce data a reef scientist can trust.
Four fully ducted thrusters: two side-mounted for surge and yaw, two on the dorsal surface for heave and pitch. No downward-pointing thrust exists anywhere in the design, and speed is capped at 0.15 m/s inside coral protection cells.
Blue excitation and a longpass filter reveal fluorescence shifts; an onboard neural-network classifier on calibrated colour imagery grades severity from 0 to 4. An alert only fires when both methods agree, which keeps sediment and fouling from raising false alarms.
A hydrophone payload records the soundscape for passive acoustic monitoring. Spectral analysis (FFT/STFT) is a core mechanism, with the vehicle’s own thruster and modem noise filtered out before any acoustic index is trusted.
A factor-graph estimator fuses IMU, depth, stereo-visual odometry, altimeter and acoustic positioning. Every mission starts by relocalising against the stored reef map, and terrain-aided matching catches drift against prior bathymetry.
A rear-facing camera on the lead vehicle tracks an accompanying diver, and a thin-glove wrist controller offers Hold, Resume, Home and Abort. The diver is treated as a protected peer with its own clearance envelope.
Dual leak sensing, emergency surfacing with an acoustic and RF beacon, and a burn-wire ballast release that is electrically isolated from the main bus, so it can fire even if the whole electronics stack is dead.
The same hull carries imaging, acoustic and environmental payloads, each wired and logged into one versioned observation record.
Pre-visual bleaching detection with a severity grade you can audit.
The reef’s soundscape as a health signal.
The water chemistry and clarity that explain what the cameras see.
Layout reference CS-DWG-SYSTEM-LAYOUT-002, revision 1.4 — the figures that matter when you are building a payload, a cradle or a mission plan.
Overall vehicle length and the dome profile for the 8″ hull are being finalised and are intentionally not stated here.
The hull is arranged front to back into dedicated zones. Internal wiring follows six signal classes routed through four zones, separating power, USB 3.0, camera ribbons and acoustic signals to keep noise out of the science data.
Forward dome port. Stereo depth camera and the high-resolution science camera behind its 500 nm longpass filter, with 450 nm blue excitation LEDs.
A 160×100×4 mm aluminium tray carrying the edge-AI compute module (8 GB, 67 TOPS), the autopilot, microcontroller I/O and a USB hub.
Two 14.8 V packs held side by side in a separate cradle — a layout that saves 146 mm of hull length against mounting them end to end.
Flat acrylic window with the rear diver-tracking camera looking out from inside the hull, so no extra penetrator is needed, plus power distribution.
Both vertical thrusters sit on the dorsal surface, so the underside stays clear for the ballast-release assembly and no jet ever points down at coral. Losing any single thruster degrades the mission gracefully rather than ending it.
Common thrust sets depth; differential with the aft duct trims pitch for inspecting overhangs.
Mirrors the fore duct. Together they hold depth without attitude drift and without downward jets.
Mid-hull duct angled outward. Common thrust drives forward; differential against starboard turns the vehicle.
Yaw authority at any speed, including hover, keeps cameras pointed at the transect without rolling the vehicle.
A factor-graph estimator provides the position; three nested layers of collision avoidance use it, each trading range for reaction time.
Reads the stored traversability map, assigns survey cells, and raises altitude floors over coral before the vehicle ever enters the water.
Runs onboard at 5 Hz, projecting the vehicle’s own path and its peers’ ahead, and yielding early when paths converge.
Independent of comms and planning: visual looming, an anomalous sonar return or lost localisation triggers an immediate response.
These are design requirements. Each one has an explicit test, and recoverability is verified by deliberate fault injection on the bench before any in-water autonomy.
No two vehicles ever come within 1.5 m; a 5 m envelope is held for 99 % or more of operating time.
A 1.5 m minimum altitude over flat seabed, with extra margin over rough terrain.
Reduced speed, planned contours around dense coral, and no downward thrust over sensitive habitat.
Any unrecoverable fault triggers an emergency ascent and beacon; no fault sequence may leave the vehicle stranded on the seabed.
Divers are protected peers: a 5 m soft envelope and a 2 m hard floor, with the swarm holding if the diver is lost.
Development is gated: every phase has an exit test that must pass before the next begins. We start with one vehicle and grow into a swarm.
8″ hull, electronics tray and battery cradle, and the internal cabling architecture are finalised for the current development hardware.
The fluorescence and colour-calibration pipeline is built and validated on synthetic data. Navigation and SLAM development is sequenced in nine phases, the first four testable without any vehicle hardware. Offline training of the bleaching classifier comes next.
Live sensor bring-up on the development compute module: IMU and depth fusion, stereo visual odometry, and the sensor-to-flight-controller link. A production compute carrier follows once the development stack is validated.
One vehicle in a test tank: localise, run a short waypoint mission, and prove every recovery behaviour with deliberately injected faults.
A supervised single-vehicle MVP at a Caribbean national marine park (site partner to be confirmed), validating species and bleaching classification against the park’s own survey baseline.
A lead and two workers with acoustic mesh networking, a single fused multi-vehicle map, and degraded-fleet operation proven by deliberate vehicle-loss drills.
A ten-vehicle production fleet, then a scaled mixed fleet, then multi-site persistent operation with moored charging so reefs can be monitored without a boat on station.
We specify what each part must do and the figures it must meet, not a brand. Where a marine-grade standard part meets the requirement we use it; where none did, we designed our own. The full bill of materials is kept in a versioned register.
Every component that touches the water should survive a full field season, so we choose anodised aluminium, marine-rated seals and proven suppliers over novelty.
Every dimension we publish traces to a manufacturer datasheet or a physical measurement. When we find a figure that was inferred rather than verified, we correct it in a dated notice and fix it everywhere it appeared — including this page.
The goal is a vehicle a research team can assemble, deploy, service and maintain in a single field season, without a machine shop on the boat.
We are building toward our first reef deployment and looking for marine research groups and conservation teams to help shape it as pilot partners. If that sounds like you, we’d like to hear from you.
Get in touch →