Recent advances in machine learning have unlocked remarkable capabilities for autonomous systems, enabling them to perform increasingly complex tasks in the physical world. As learning-enabled control becomes increasingly prevalent in the physical world, the risks associated with its failures increase correspondingly. My research focuses on the control architectures and statistical foundations of learning-enabled autonomy, with the goal of developing principled approaches for reliable deployment of learning-enabled systems in the physical world.
The increasing scale and complexity of modern autonomous systems introduce significant challenges in control design. Such systems increasingly combine different layers of functionality, from perception and language to planning and control, with components operating at different timescales and relying on models of varying fidelity. Layered control provides a powerful framework for organizing these components and coordinating decisions across different levels of an autonomy stack.
A central challenge in this paradigm lies in preserving the desired specifications across layers. My research explores principled approaches to layered control architectures that tackle this challenge across heterogeneous dynamics, timescales, and constraints.
Data collection from physical systems can often be expensive or risky in practice, limiting the capabilities and scalability of learning-enabled controllers. At the same time, related systems and control objectives may exhibit shared structure that can be exploited during learning. Multitask reinforcement learning and control aim to exploit this structure by learning a common controller across a distribution of dynamics and objectives.
A fundamental challenge that remains lies in understanding when a common controller can effectively transfer across heterogeneous tasks. My research explores system-theoretic approaches to characterizing task heterogeneity and its implications for multitask control learning.
Learning-enabled controllers involve uncertain components, and such uncertainty can propagate through the control loop and compromise safety and performance. This motivates approaches for quantifying uncertainty and leveraging uncertainty estimates to establish theoretical guarantees.
Statistical learning provides a foundation for analyzing the reliability of learned components in dynamical systems and control. My research explores statistical foundations for learning dynamics and control, with a focus on characterizing uncertainty in the identification and safe control of autonomous decision making agents.