| Path planning is the process of generating an optimal sequence of waypoints from a start configuration to a desired goal configuration under constraints (e.g., avoiding obstacles, respecting time/energy budgets). In this thesis, we study the problem of risk-aware planning. Specifically, we design, develop, and experimentally validate optimal paths for Autonomous Underwater Vehicles (AUVs) in the open ocean in the presence of navigational hazards such as ships and other obstacles. A novel aspect of this work is the introduction of ocean current predictions to optimize planning in such settings. This is challenging because current predictions are typ- ically available at non-uniform spatial resolution, noisy, and time-delayed. We designed three risk-aware planners that reason probabilistically about the uncer- tainty in ocean currents predictions. The minimum expected risk planner ensures that the AUV always reaches the goal, while minimizing risk along the way, Risk- aware Markov Decision Process-based planning uses stationary models over a short horizon, and trades off between goal-directed behavior and reducing risk. This is susceptible to finding sub-optimal policies due to stationarity. The non-stationary, risk-aware MDP makes use of variability in the currents where possible to overcome high-risk sections of paths on the way to the goal. In addition to these planners, we develop a taxonomy for risk-aware planning in dynamic settings. Another key contribution is learning the uncertainty in currents to improve planning. Results from extensive simulations clearly show that learning uncertainty helps significantly improve performance of risk-aware planners in uncertain currents, allowing AUVs to be operated in more challenging scenarios than was previously possible. Finally, the planners described in this dissertation have been field tested at unprecedented levels to validate their practical utility (~2000 hours of testing at sea). |
