Title: Bridging Models and Learning towards Humanoid Robustness and Versatility

Date: Friday, September 26th, 2025

Time: 12:00 - 2:00 pm ET

Location: Erskine Love Building 184, or  Zoom link

Zhaoyuan Gu

Robotics Ph.D. Student

Woodruff School of Mechanical Engineering

Georgia Institute of Technology

Committee:

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

Dr. Maegan Tucker – School of Electrical and Computer Engineering and Woodruff School of Mechanical Engineering, Georgia Institute of Technology

Dr. Sehoon Ha – School of Interactive Computing, Georgia Institute of Technology

Dr. Patrick Wensing  – Aerospace and Mechanical Engineering, University of Notre Dame

Dr. Guanya Shi – Robotics Institute and the School of Computer Science, Carnegie Mellon University

Abstract:

Humanoid robotics is in a transformative era. The ever-more reliable and accessible hardware offers a unique opportunity to address the pressing demand for general-purpose humanoid agents. Yet, today’s humanoids still face persistent challenges of robustness against unexpected perturbation, and most can perform only a single task. Our key insight is that models provide strong, robust, and safe guarantees of humanoid locomotion. To extend beyond model-based methods, we propose to improve a pre-trained policy through reinforcement learning fine-tuning. This proposal advances the foundations of robust locomotion and extends toward versatile loco-manipulation through three integrated contributions: (1) Real-time Model Prediction Control (MPC) with formal robustness guarantees for perturbation-resilient locomotion; (2) Reinforcement Learning (RL) motion imitation that creates versatile and precise loco-manipulation from expert demonstration; and (3) A proposed fine-tuning approach that exceeds the limits of the expert demonstrations, gaining versatility and success rates critical for real-world deployment. Together, these efforts bring the rigor of formal models with the flexibility of learning, advancing toward the overarching goal of robust and versatile humanoid robots.