Abstract:
Gathering information from the physical world plays a crucial role in many applications—whether it be scientific research, environmental monitoring, search and rescue, defense, or disaster response. The utilization of robots for information gathering allows for the leveraging of intelligent algorithms to efficiently collect data, providing critical insights and facilitating informed decision-making. These autonomous robots provide unparalleled advantages in situations where human access is constrained, dangerous, or logistically challenging. Moreover, the scalability and speed of autonomous information gathering robots significantly enhances the rate of information gathering by not capping the number of robots based on the number of human operators. Informative path planning (IPP) is key to creating intelligent paths for robots to execute actions that maximize their inherent potential and advantages in these data gathering situations.

Much research has been done in the space of IPP and has shown the inherent challenges of creating algorithms that are adaptable, efficient, and scalable. These challenges include incorporating real-world disturbances into the path-planning framework, solving IPP problems in high-dimensional spaces with non-trivial sensor constraints, creating accurate and useful world belief models, and evaluating the IPP algorithms in ways that accurately reflect their performance in the real world.

In this thesis, I seek to address these challenges and improve the performance of IPP in autonomous real-world applications. First, I present our work on incorporating real-world disturbances into planning, focusing mainly on wind-aware planning for Uncrewed Aerial Vehicles (UAVs) which includes how to estimate urban wind fields online, quickly solving time-optimal paths in wind, and how to predict the energy usage and risk of a UAV flight in wind. I then present our ongoing work building upon our IPP algorithm, TIGRIS, with modifications to make it iterative, faster, and incorporate nuanced reward functions. For my proposed work I will explore different aspects of improving world belief models through new belief representations and objective functions as well as using predictive models from prior data. Next, I will build upon our planning work by finding how to create a long-horizon and high-dimensional IPP planner that adapts and performs well in situations with or without strong information map priors. Over the course of these proposed works I would like to create a clear and public evaluation pipeline that effectively benchmarks the impact of our methods and allows for others to do the same.

Thesis Committee Members:
Sebastian Scherer, Chair
David Wettergreen
Maxim Likhachev
Geoff Hollinger, Oregon State University

More Information