Policy Learning for Autonomous Feature Tracking

We consider the problem of tracing the structure of oceanological features using autonomous underwater vehicles (AUVs). Solving this problem requires the construction of a control strategy that will determine the actions for the AUV based on the current state, as measured by on-board sensors and the historic trajectory (including sensed data) of the AUV. We approach this task by applying plan-based policy-learning, in which a large set of sampled problems are solved using planning and then, from the resulting plans a decision-tree is learned, using an established machine-learning algorithm, which forms the resulting policy. We evaluate our approach in simulation and report on sea trials of a prototype of a learned policy. We indicate some of the lessons learned from this deployed system and further evaluate an extended policy in simulation.