Informative Path Planning (IPP) is an important planning paradigm for various real-world robotic applications such as wildfire monitoring and predicting infection spread in crops. IPP involves planning a path that can learn an accurate belief of the quantity of interest, while adhering to planning constraints. Traditional IPP methods are effective only in static, time-invariant environments and they typically require high computation time during execution. This has given rise to reinforcement learning (RL) based IPP methods. Despite these advances, existing RL-based methods do not address spatiotemporal environments, which present unique challenges due to variations in environment dynamics.

This thesis introduces a robust RL-based IPP framework specifically designed to enable robots to operate effectively across dynamic spatiotemporal environments. Our approach combines domain randomization with our proposed dynamics prediction model (DPM). The DPM constitutes a key contribution of our framework, explicitly modeling how environments evolve over time and extracting latent representations of their specific characteristics. Through extensive evaluations in a wildfire spread prediction task, we demonstrate that our DPM successfully infers environment dynamics online, enabling the RL policy to maintain consistent performance across environments with significantly different dynamics.

Committee:
Prof. Katia Sycara (advisor)
Prof. Sebastian Scherer
Ananya Rao