Zhengjie (Jerry) Xu

I am a third-year PhD student in Computer Science and Engineering at the University of Michigan, Ann Arbor. I am luckily advised by Prof. Stella Yu. I am broadly interested in computer vision and robotics.

Previously, I obtained my M.S.E. in Robotics from University of Pennsylvania, and B.S. in Computer Engineering and Applied Mathematics from UCSD. During my master, I was fortunate to work with Prof. Jianbo Shi on image generation and reconstruction. I also worked with Prof. Hao Su and Prof. Garrison W. Cottrell during my undergrad.

Email  /  GitHub  /  Google Scholar

Research

Unified Humanoid Fall-Safety Policy from a Few Demonstrations

Falling is an inherent risk of humanoid mobility. Maintaining stability is thus a primary safety focus in robot control and learning, yet no existing approach fully averts loss of balance.When instability does occur, prior work addresses only isolated aspects of falling: avoiding falls, choreographing a controlled descent, or standing up afterward. Consequently, humanoid robots lack integrated strategies for impact mitigation and prompt recovery when real falls defy these scripts. We aim to go beyond keeping balance to make the entire fall-and-recovery process safe and autonomous: prevent falls when possible, reduce impact when unavoidable, and stand up when fallen. By fusing sparse human demonstrations with reinforcement learning and an adaptive diffusion-based memory of safe reactions, we learn whole-body behaviors that unify fall prevention, impact mitigation, and rapid recovery in one policy. Experiments in simulation and on a Unitree G1 demonstrate robust sim-to-real transfer, lower impact forces, and consistently fast recovery across diverse disturbances, pointing toward safer, more resilient humanoids in real environments.

Perceptual Artifacts Localization for Image Synthesis Tasks

We generalize Perceptual Artifacts Localization to ten diverse image synthesis, and shows promising accuracy. We also show the effectiveness of automatic artifacts fixing and quality assessment as downstream applications.

Optimizing Algorithms From Pairwise User Preferences

We present HashEncoding, a novel autoencoding architecture that leverages a non-parametric multiscale coordinate hash function to facilitate a per-pixel decoder without convolutions. By leveraging the space-folding behaviour of hashing functions, HashEncoding allows for an inherently multiscale embedding space that remains much smaller than the original image.

Generalization in Cardiac Image Segmentation

We focus on the task of cardiac image semantic segmentation and synthesize five different real-world scenarios to find the optimal training approach for deep learning models to achieve good generalization across different datasets.

Other Projects

These include coursework, side projects and unpublished research work.

Super-pixel Sampling in 3D

We implemented a differentiable grouping module and combine it with off-the-shelf point cloud analysis frameworks (PointNet++). The module optimizes the embedding space to enforce the local feature consistency. The learnt mid-level abstraction can be used to discover geometry primitives as well as simplify the downstream tasks. With this module, we out-performed the original framework in point segmentation task.

Teaching

I was a TA/GSI for the following courses:

• UPenn CIS 6800: Advanced Topics in Machine Perception (Fall 2022)
• UM EECS 542: Advanced Topics in Computer Vision (Fall 2024)
• UM EECS 442: Computer Vision (Winter 2025)

Misc

I have a lovely ragdoll as you can see from my profile photo. His name is Dodo(代代) and he is now 4 years old!

Design and source code from Jon Barron's website