SE PhD Prospectus Defense of Ahmad Ahmad

TITLE: Planning, Control, Learning, and Monitoring for Robotic Systems Under Temporally Constrained Tasks

ADVISOR: Calin Belta, MechE; Roberto Tron, ME, SE

COMMITTEE: Roberto Tron, ME, SE; Wenchao Li, ECE; Alyssa Pierson, MechE; Gioele Zardini, MIT

ABSTRACT: This dissertation presents a unified framework for efficient, robust, and interpretable robotic control and planning under temporal constraints. Temporal logics (TLs) have been widely used to formalize interpretable tasks for cyber-physical systems. Time Window Temporal Logic (TWTL) has been recently proposed as a specification language for dynamical systems. In particular, it can easily express robotic tasks and allows for efficient, automata-based verification and synthesis of control policies for such systems. We leverage TWTL to express complex tasks and develop innovative approaches for planning, control, and learning. Our key contributions include (i) Efficient Planning: We introduce a sampling-based planning algorithm that ensures robust TWTL task satisfaction using novel quantitative semantics and incremental monitoring. (ii) Reinforcement Learning (RL): We propose enhancements to Proximal Policy Optimization for efficient learning, incorporating offline guidance using simple behavior cloning policies and TWTL-based reward shaping. (iii) RL-Optimized Sampling: We apply RL to optimize the sampling strategy in a safe sampling-based motion planning algorithm to improve planning efficiency. By combining these advancements, we provide methods with significant implications for various robotic applications.