Zoom Link: https://upenn.zoom.us/j/99241403942?pwd=utmb7mJnRmc4z4whwVTCLr1f7Hi0UU.1
Meeting ID: 992 4140 3942
Passcode: 051736

Abstract: The transition of robot deployment beyond structured factories into homes, offices and busy urban spaces introduce compounded challenges of shifting task demands, unpredictable human motion, partial and noisy perception. Relying on a single monolithic planner to interpret observations, reason over high-level objectives, and compute safe low-level motions all at once is computationally intractable. This thesis presents multi-principled hierarchical frameworks across three problem domains that decompose the planning problem across different granularities: high-frequency low-level planners with formal safety and stability guarantees, coupled with low-frequency high-level planners that reason over task structure and environmental state. For the first domain of interactive manipulation, I augment Neural ODE models with online correction terms for stability and safety, and integrate a reactive temporal-logic task planner, thereby enabling a Franka robot arm to seamlessly adapt to disturbances and human interactions for wiping and stirring tasks. Second, for autonomous parking under partial observability, I propose a time-indexed Hybrid A star algorithm and a probabilistic strategy planner which improves parking efficiency and safety in simulated parking lots. Finally, for a mobile manipulation platform, I develop a stochastic sampling-based planner that incorporates multifaceted perception uncertainty from human motion prediction and object detection, alongside a high-frequency safety filter. Overall, by decoupling complex task planning from mathematically guaranteed safe control, this thesis establishes hierarchical architectures critical for achieving safe, reactive, and robust autonomy in diverse, dynamic, human-centric environments.