PhD · NTNU · adaptive sampling
Teaching an underwater robot where to look
Bayesian models and a planner that ran onboard an autonomous underwater vehicle in Trondheimsfjord, steering it along a river-plume front instead of surveying evenly.
- 2.5 hautonomous mission in the fjord
- 5peer-reviewed papers from the PhD
- Onboardmodels updated live from the robot's own data
The question
An autonomous underwater vehicle has a few hours of battery and a fjord several kilometres across. It cannot measure everywhere, so the real question is: given what it believes now, where should it go next?
The method
- A belief about the ocean. Gaussian processes and sparse Gaussian Markov random fields model salinity in space, fast enough to update onboard from every new measurement.
- A reason to move. Expected-improvement and value-of-information criteria score where a measurement would most reduce uncertainty about the front between river and sea water.
- A loop that closes at sea. Sense, update, plan, and send waypoints to the vehicle's controller, repeated for the whole mission.
Did it work
Every planner was compared in Monte Carlo simulation on a GPU cluster before going to sea, then run onboard LAUV vehicles in field campaigns with SINTEF Ocean and LSTS (Porto), including a 2.5-hour autonomous mission and side-by-side adaptive and pre-planned runs. The adaptive vehicle spends its time where the front is, and ignores the water it already understands.
The simulator now runs in your browser, rewritten from scratch in JavaScript.