Autonomous science robots encounter coupled planning and modeling challenges when placed in an unknown environment. When little prior information is provided about a survey domain, many exploration algorithms are prone to exhibit behavior resembling random exploration or local gradient-following. Furthermore, a data-collecting agent’s belief about its environment is often modeled as a continuous, uncertain Gaussian process, which makes the definition of policies for the agent a complex task. To address these challenges, we propose an autonomous agent that learns about the conditional relationships between multiple variables or features over a survey field in the form of a graphical model. We demonstrate that such an agent can overcome the problem of weak priors by making inferences, and we define an active learning policy for such an agent to intelligently gather information about variable dependencies. Our method has applications in learning with side information, transfer learning, multi-agent coordination, and surveillance of nonstationary domains.

I am participating in SuperUROP because I want to gain experience doing scientific research in robotics. I hope to gain a better understanding of predictive and model-based algorithms and techniques to learn how to develop, evaluate, and express my ideas, and to contribute towards the research goals of my group.