Speaker: Chaoyi Pan Carnegie Mellon University
Time: 2024-12-17 10:00-2024-12-17 11:00
Venue: Seminar Room 2, 19th Floor, Tower C, TusPark (https://meeting.tencent.com/dm/00uCfUm7ZoV5)

Abstract:

The development of machine learning algorithms for decision-making has been a collaboration between **offline learning** and **online planning**. Both paradigms have scaled remarkably well with advancements in computing power. For example, in games like chess and Go, systems such as AlphaGo learn a value function offline and then utilize Monte Carlo Tree Search (MCTS) for online decision-making during gameplay. Similarly, in the development of large language models (LLMs), approaches like OpenAI’s GPT models combine extensive offline training with online reasoning capabilities refined through reinforcement learning (RL). In robotics, however, the integration of offline learning and online planning is less cohesive, with these two domains often evolving independently. While large-scale RL has demonstrated its applicability to real-world tasks through extensive offline training, the online search component remains underexplored. This talk will introduce **MBD** (Model-Based Diffusion) and **DIAL-MPC** (Diffusion-Inspired Annealing for Legged MPC), a sampling-based Model Predictive Control (MPC) framework that incorporates a novel diffusion-style annealing process to enhance online search scalability. The proposed annealing technique is underpinned by a theoretical landscape analysis of Model Predictive Path Integral Control (MPPI), establishing a connection between MPPI and single-step diffusion. Algorithmically, the method iteratively refines solutions online, achieving both global coverage and local convergence. To the best of our knowledge, DIAL-MPC is the first training-free method capable of real-time optimization for full-order quadrupedal dynamics.

For more details, please visit our website:

https://lecar-lab.github.io/dial-mpc

https://lecar-lab.github.io/mbd

Short Bio:

Chaoyi Pan is a second-year Ph.D. candidate in Electrical and Computer Engineering at Carnegie Mellon University, co-advised by Prof. Guanya Shi (LeCAR Lab) and Prof. Guannan Qu. His research focuses on bridging learning and control in robotics, leveraging tools from generative model. He received his B.S. degree with highest honors from the Department of Electronic Engineering at Tsinghua University in 2023.