Path planning for autonomous information collecting vehicles

In many environments where autonomous air or ground vehicles are used to collect information, there will be a known prioritization of areas of the environment where most valuable information will be found. Over time, priorities may change with areas losing value or suddenly becoming important. In this paper, we present an approach to planning paths for vehicles collecting information in such environments, such that they maximize the overall system information gain over time. A key feature of this path planning problem is that there is not a single or small set of goal points to which the vehicles should try to reach, instead information is collected over the entire path without a particular goal in mind. We present a planning approach, which rapidly expands a search tree, inspired by an RRT planner by choosing promising nodes to expand and expanding them randomly. Genetic algorithms are used to learn sets of configuration parameters for the planner, i.e., how to expand which nodes. Results show that the learned planner gets more substantially information than pre-defined paths in a variety of domains.