Time: 3-4 pm Eastern Time, Feb 13 (Friday), 2026

Location: MSEE 112

Zoom Link: https://purdue-edu.zoom.us/j/98798335169

Coffee and snacks will be provide.

Bridging the Gap: How to Integrate MPC and RL for Legged Robots

Abstract:

Over the past five years, we have witnessed deep reinforcement learning (RL) methods solve impressive motion control tasks for quadrupeds and humanoids, often surpassing model-based methods such as Model Predictive Control (MPC). This mirrors the "bitter lesson” observed in fields like computer vision, where methods that scale with data eventually dominate. In this talk, I argue that MPC still has a vital role to play. I will present my early attempts at combining RL and MPC for quadrupedal locomotion, as well as my recent work using MPC to control a quadrupedal robot playing table tennis. Furthermore, I will discuss specific scenarios where MPC retains an edge over learning-based methods. Finally, I will share insights on how we can advance both MPC and RL by leveraging their complementary strengths to drive mutual progress.

Speaker:

Dr. Zhaoming Xie is a research scientist from the Robotics and AI Institute (formerly Boston Dynamics AI Institute). He obtained his Bachelor's degree from the Hong Kong University of Science and Technology, his Master's degree from the Georgia Institute of Technology, and his PhD from the University of British Columbia. He also spent time as a post-doc at Stanford University before joining the Robotics and AI Institute as a research scientist. His research focuses on applying machine learning and optimization techniques to achieve athletic intelligence in robotics and simulated characters. He was one of the first to successfully transfer policies trained with deep reinforcement learning in simulation to a physical bipedal robot.

Organizers: Ziran Wang (ziran@purdue.edu), Yan Gu (yangu@purdue.edu), Yu She (shey@purdue.edu)