Date: Tuesday, October 6th, 2026

Time: 3:00 PM - 4:30 PM EDT

Location: MRDC 3515

Zoom: https://gatech.zoom.us/j/96846452330?pwd=oGfmTGbPoLo2zyp5ulwGl9PjECLl0Y.1 

Zhigen Zhao

Ph.D. Student

Institute for Robotics & Intelligent Machines

Georgia Institute of Technology

Committee members

Dr. Ye Zhao (advisor): Woodruff School of Mechanical Engineering, Georgia Institute of Technology

Dr. Jiachen Li: School of Industrial and Systems Engineering and Woodruff School of Mechanical Engineering, Georgia Institute of Technology

Dr. Shreyas Kousik: Woodruff School of Mechanical Engineering, Georgia Institute of Technology

Dr. Sonia Chernova: School of Interactive Computing, Georgia Institute of Technology

Dr. Shiqi Zhang: School of Computing, Binghamton University, State University of New York

Abstract

Robots deployed in unstructured, human-centric environments must carry out long-horizon tasks that couple discrete decisions with contact-rich continuous motion. Task and Motion Planning (TAMP) addresses this by decomposing the problem into a discrete task plan and a continuous motion plan, but classical TAMP requires re-engineering for each new task or environment, scales poorly with problem size and horizon, and accumulates no experience across deployments.

This proposal builds toward generalizable, scalable, and experience-driven robot autonomy by progressively replacing the hand-engineered layers of classical TAMP with learning-based counterparts. First, we formulate TAMP as a single bilevel optimization that couples symbolic search with dynamics-consistent motion and remains scalable by exploiting task structure. Second, we move this optimization offline and learn fast, robust motion policies by imitation, supported by cross-platform teleoperation and egocentric data infrastructure and by a discrete action representation that makes multi-task policies steerable at inference time. Finally, we propose an embodied agent system and an agent memory evaluation benchmark in which an LLM/VLM orchestrates VLA skills and a self-evolving multimodal memory accumulates verified experience across deployments, enabling the robot to continually self-improve from its own experience without retraining.