ESE Ph.D. Dissertation Proposal Defense – “Layering and Task Generalization in Control Architectures”

Learning-based control architectures have unlocked remarkable capabilities in autonomous systems, from robotic manipulation to autonomous driving and traffic coordination. This progress has motivated the widespread use of layering to manage large-scale systems and multitask control to handle multiple systems and missions. The increasing integration of these architectures into society calls for a rigorous theory that guarantees safety in layered designs and clarifies when multitask controllers can generalize well across tasks.

In this talk, I will present my work on advancing the foundations of layering and task generalization in control architectures. I will first introduce a framework for the co-design of planning and tracking layers in multirate control of constrained linear systems. This framework provides computable tracking error bounds along with safety guarantees for any planner. I will then develop formal measures of task heterogeneity for multitask linear quadratic control and demonstrate how these measures bound the suboptimality of using a shared controller on tasks with distinct objectives, for both optimal and policy-gradient approaches. These results offer interpretable conditions under which a single controller can perform effectively across many tasks. I conclude with directions for extending these methodologies to nonlinear systems and distributed networks of multiple agents.