Trajectory Planning for Autonomous Vehicles Performing Information Gathering Tasks

Thumbnail Image


Publication or External Link





This dissertation investigates mission scenarios for autonomous vehicles in which the objective is to gather information. This includes minimizing uncertainty of a target's estimated location, generating coverage plans to cover an area, or persistent monitoring tasks such as generating informative patrol routes. Information gathering tasks cannot be solved with shortest path planning algorithms since the rewards are path-dependent. Further, in order to deploy such algorithms effectively in the real world, generated plans must safely avoid obstacles, account for the motion uncertainty (e.g. due to swift currents), and constraints on the vehicle's dynamics such as maximum speed/acceleration. This work extends state-of-the-art information gathering algorithms by generating dynamically feasible trajectories for autonomous vehicles that are able to exploit the environment to find higher quality solutions, reducing mission costs. We also reduce mission risk without sacrificing the amount of information gathered. The focus of this dissertation will be to solve three related information gathering tasks that require generating dynamically feasible trajectories for reliable plan execution.

When searching for targets, minimizing target location uncertainty with autonomous vehicles improves the effectiveness of ground relief crews. We investigate the use of mutual information for efficiently generating long duration multi-pass trajectories to minimize target location uncertainty in natural environments. We develop epsilon-admissible heuristics to create the epsilon-admissible Branch and Bound algorithm to gather the most information. Next, we investigate coordination techniques for underwater vehicle teams conducting large-scale geospatial tasks such as adaptive sampling or coverage planning. It is advantageous to exploit the currents of the ocean to increase endurance, which requires accounting for forecast uncertainty. We adapt Monte Carlo Tree Search and Cross Entropy Method to maximize path-dependent reward, and introduce an iterative greedy solver that outperforms state-of-the-art planners. Finally, we investigate persistent monitoring tasks such as sentry patrol routes and monitoring of harmful algae blooms in a littoral environment. Such an automated planner needs to generate collision-free coverage paths by moving waypoints to locations that both minimize path traversal costs and maximize the amount of information gathered along the path. We extend previous Lloyd's based algorithms by factoring in the ocean currents and introduce greedy methods that minimize mission risk while maximizing information gathered.