Amortizing Adaptive Informative Path Planning for Autonomous Underwater Seagrass Mapping

Posidonia oceanica meadows are a protected Mediterranean habitat and a significant carbon sink, but mapping their spatial distribution at scale is difficult: satellite methods lose reliability with depth, diver surveys cover little ground, and AUV lawnmower transects spend most of time gathering redundant data over homogeneous areas. Adaptive informative path planning instead maintains a probabilistic map of habitat coverage and steers the vehicle toward regions of high predictive uncertainty. We reproduce the AVIG framework of Guerrero et al. inside a Stonefish digital twin, giving a complete perception-to-planning loop from stereo segmentation through a sparse Gaussian process map to a Monte Carlo tree search planner. Against a lawnmower baseline, adaptive planning recovers the meadow structure substantially sooner. This advantage is paid for in onboard computation: every replanning step simulates a fresh batch of candidate paths on a power-constrained vehicle with no offboard compute. We propose to move this search offline by training a policy in advance that maps the Gaussian process belief directly to an action, reducing onboard cost to belief maintenance plus a single forward pass. Training environments would be built from segmented photomosaics of real meadows, with entire sites held out for evaluation. We outline two candidate belief representations: the planner's existing graph of sampling locations, and a vehicle-centred image passed to a convolutional encoder.