Performing autonomous robotic tasks in the field, such as ocean monitoring and aerial surveillance, requires planning and executing paths in dynamic environments. In these uncertain and changing environments, it is not uncommon to see a large difference between the path planned by the robotic vehicle and the path that the robotic vehicle realizes while executing that path. This difference can decrease the performance of the robotic system by introducing additional risk or forcing the vehicle to miss important survey areas. Existing systems do not consider large-scale disturbances, do not consider the differences between the planning and control models, and do not incorporate new information about disturbances found online during planning and execution. To address these shortcomings, this thesis provides algorithms that incorporate disturbances directly into planning, reason about the robot's low-level controller, and utilize information gathered during execution about both disturbances and the robot's dynamics. The impact of these improvements is a reduction in risk and improvement of the quality of robotic information collection.
This thesis provides three contributions to help reduce this difference between planned and executed trajectories. First, we introduce a stochastic optimization framework which utilizes an action-space path representation to remove the need for expensive reachability calculations. This action-space formulation allows for a more natural representation of the effects of disturbances on vehicles with low actuation, and the stochastic optimization technique allows the mapping of a state-space based reward function to the action-space while being efficient enough to be used in a sequential allocation framework for planning for multiple vehicles. We demonstrate the computational efficiency of this algorithm against other state-of-the-art planners in a simulated ocean environment of the Gulf of Mexico.
Second, we present a novel algorithm, Energy-Efficient Stochastic Trajectory Optimization (EESTO), which allows vehicles with moderate levels of actuation to plan energy-efficient trajectories thorough strong and uncertain disturbances. In addition to this algorithm, we introduce a framework which can utilize the efficiency of EESTO to account for information gathered online about the disturbances that the vehicle is moving through. We demonstrate the capabilities of the algorithm and framework in both a simulated ocean environment off the coast of California near the Channel Island as well as on hardware on a lake near Eugene, Oregon.
Lastly, we present a framework for increasing the realizability of planned paths for high-actuation vehicles, which allows the robotic system to reason about the capabilities of the on-board low-level controller. By incorporating the capabilities of the low-level controller into execution and planning, this framework is able to increase the realizability of the planned information gathering path. We demonstrate the capabilities of this framework through extensive simulation trials and on hardware on a lake near Corvallis, Oregon.