Research Projects
Learning Environmental Models Using Dynamical Systems Theory
Successful interactions of robots and other machine learning agents with the environment first requires understanding the environment. Mapping, or modeling the environment, is essential to operating in complex domains, allowing robots to perform in more realistic scenarios. Creating rich spatial representations that take into account large volumes of data and uncertainty is critical for robot teams to safely navigate in dynamic environments, such as in the ocean or in space.
This works explores the use of dynamical systems and transfer operator theory in machine learning algorithms that are leveraged to create spatial representations of the environment. These spatial representations can identify key features, such as coherent sets, that can be used in planning and decision-making.
Publications and Presentations
Online Estimation of the Koopman Operator Using Fourier Features
Learning and Leveraging Features in Flow-Like Environments to Improve Situational Awareness
Adaptive Sampling of Spatiotemporal Phenomena with Multi-Robot Teams
Multi-robot teams have emerging applications in environmental monitoring including tracking and predicting concentrations of hazardous chemicals, such as explosives or toxins in the atmosphere, or forecasting time series imagery of phenomena, such as weather patterns or species migration in oceans.
We design frameworks for multi-robot systems to study how different actuation and sensing allows for rich distributed and adaptive tracking and inferencing of dynamic environments, including studies on heterogeneous sensing agents and agents with only intermittent communication capabilities.
Publications and Presentations
Adaptive Sampling and Reduced-Order Modeling of Dynamic Processes by Robot Teams
paper | slides | video | poster
Adaptive Sampling and Energy Efficient Navigation in Time-Varying Flows
Heterogeneous Multi-Robot Systems for Modeling and Prediction of Multiscale Spatiotemporal Processes
Asynchronous Adaptive Sampling and Reduced-Order Modeling of Dynamic Processes by Robot Teams via Intermittently Connected Networks
