Mengdi Xu

Research Abstract:

In an open-ended world, robots are bound to face tasks they've never seen before. For example, a household robot must manipulate unseen objects or complete long-horizon cooking tasks with novel compositions. Similarly, an autonomous vehicle must interact safely with human drivers with substantially different behaviors unknown beforehand. Such realistic and highly uncertain scenarios cast significant challenges to robot learning algorithms, and require powerful generalization capability before the massive robot deployment. In light of those real-world problems, my research vision is to build robots that can generalize to challenging new environments with strong efficiency and robustness. To enable the robot's efficient generalization with minimal environment interactions and computational efforts, I developed few-shot robot learning methods to acquire new low-level skills through several demonstrations and solve complex physical puzzles with the help of large language models. To ensure robustness when facing uncertain tasks, especially human-in-the-loop high-stake scenarios, I proposed to model the possible multiple tasks with a hierarchical structure, learn the structure in an unsupervised manner, and exploit the structure to boost the robustness. I theoretically proved that the hierarchical architecture strikes a balance between distributional robustness and expected policy performance. My future objectives are two-fold: first, to develop efficient generalists that go beyond statistical generalization to compositional generalization by relying on the rich prior knowledge of the robot foundation models; and second, to build reliable autonomy consisting of certifiable distributionally robust agents with satisfactory generalization performance and robustness guarantees.

Bio:

Mengdi Xu is a Ph.D. student at Carnegie Mellon University, advised by Prof. Ding Zhao. Her research aims to build robots that efficiently and robustly generalize to challenging new scenarios akin to everyday life. Her research centers around (1) laying theoretical foundations for generalizable robot learning, (2) designing efficient and robust robot learning algorithms with statistical guarantees, and (3) advancing robots with novel capabilities. She has interned at Google DeepMind, MIT-IBM Watson AI Lab, and Toyota Research Institute. She obtained her Master's in Machine Learning from Carnegie Mellon University, Master's in Robotics from Johns Hopkins University, Bachelor's in Vehicle Engineering, and Bachelor's in Management from Tsinghua University. She was selected as an RSS Pioneer 2023 and a Computational and Data Science Rising Star 2023.