04 / AUTONOMY
Autonomy is layered: a safety controller keeps the vehicle recoverable, while an edge computer interprets the scene and proposes the next action.
The research shortlist includes Jetson Orin NX (up to 100 TOPS in a 10–25 W class) and, for heavier experiments, AGX Orin Industrial (up to 275 TOPS). Those are candidate envelopes, not selected hardware. A Raspberry Pi 5 with a Hailo or Movidius accelerator is a lower-power alternative.
YOLOv8/YOLOv10-style detection, Sea-thru or WaterGAN-inspired image correction, and multi-sensor fusion are research candidates. A deterministic state machine still owns the vehicle: launch, survey, avoid, sample, return, abort, and recover. A model may suggest a route; it never disables the abort path.
Bench tests, pressure tests, a clear-water tank, seeded gravel, turbidity trials, tether tests, and a supervised river pilot build the evidence ladder. Permits, wildlife rules, navigation requirements, and landowner permissions are part of the product—not paperwork after the fact.
ROS 2 is a practical way to separate camera, sonar, LiDAR, PI, navigation, and sampler nodes while preserving timestamps and diagnostics. The edge computer can publish compressed operator previews and full-rate data to local storage, while a safety MCU owns motor current limits, leak response, emergency stop, and a hardwired recovery output. Time synchronization and dropped-message counters belong in the mission bundle.
A lawnmower survey pattern is only the opening state. When a route cell contains a sonar trap shape, a repeatable PI waveform, or a camera anomaly, the planner can slow, increase light, take a macro pair, repeat the pass, and drop a virtual pin. The system returns to the planned route only when confidence, tether geometry, battery, and current are inside their limits.
