Estuaries, mangroves, and tidal flats are awkward to work in. Traditional survey vessels draw too much water, wading disturbs the habitat, and aerial drones can’t collect samples or run sonar. Macquarie’s Estuarine Research Vessel (MERV) is a practical attempt to fill that gap.
The Build
MERV is built on a race-boat hull (we have two: an MHZ Mystic and a Pro Boat Miss Geico — the same parameter set works on both). The autopilot is ArduRover, paired with a Holybro CM4 companion computer in a custom 3D-printed housing alongside the power distribution board.
There are no servos or rudders. Turning is entirely skid-steered via two ESCs and two motors. One thing worth flagging early: the motors in these race hulls are seriously powerful. You do not want to open them up at full throttle, and full throttle in reverse is something you want to discover in a controlled setting, not in the field. The precise motor control available through ArduPilot is genuinely useful here as it lets you creep through sensitive habitat at low throttle without the kind of wash that disturbs wildlife or kicks up sediment.
The ground station is a Lenovo Legion Go, connecting over WiFi or 4G depending on location. You can fail-over to a traditional RC transmitter and receiver, but doing so loses the live camera and sonar feeds, both of which turn out to matter quite a lot in practice.
Weight is under 2 kg, draft is around 5 cm, and the flat hull plus powerful motors means you can usually free yourself if you do run aground on sand. One person can carry and deploy it without assistance.
What We Have Used it For
The persistent, high-bandwidth connection back to the ground station is what makes this platform genuinely flexible. So far we’ve run:
- Initial camera + sonar integration, streaming live to the ground station
- Testing 4G on-water connectivity around the east coast of Australia
- Automated eDNA sample collection
- Shallow-water sonar development (this work has since produced some solid results — see the NSSN writeup)
- Camera-based AI object detection
- Camera-based AI follow-me
- Custom firmware with alternative motor slew control code
That last one came directly out of the throttle behaviour we observed in manual mode. The stock slew handling in ArduRover is reasonable for most platforms, but race-boat ESCs are a different environment. The torque response is abrupt enough that it causes real problems in practice . We implemented and tested alternative motor slew control logic in a custom firmware build to get smoother, more predictable behaviour, particularly at the low-throttle end where most of the useful work happens in sensitive environments. However, in doing so we found motor slew is a give-here and take-there problem and we are very keen to hear from others interested in that area.
The 2024 on-water AI work was particularly interesting because the image environment on water is quite different from aerial drone footage, and there’s not much prior work to draw on. Research students used MERV to collect training data and benchmark how well standard approaches held up. For object detection (for both avoidance and following), YOLO could be supplemented to work well, but that supplementation was custom, which is never ideal.
Getting Involved
MERV was developed with support from the NSW Smart Sensing Network (NSSN) and CEE HydroSystems. The platform is available to universities and research organisations interested in water sampling, autonomous algorithms, radio protocols, or AI development.
If any of that is relevant to work you’re doing, get in touch with Matt Roberts at Macquarie University: matthew.roberts@mq.edu.au
