Zhiwei Zheng

I am currently a first-year Ph.D student at University of Pennsylvania, focusing on multi-modality learning and generation, advised by Prof. Mingmin Zhao.

Before that, I obtained my Master's degree in EECS and Bachelor's degree in Automation from UC Berkeley and Huazhong University of Science and Technology, respectively. I also interned at Shanghai AI Lab, working on autonomous driving.

Feel free to reach out if you are interested in my research! I am also looking for motivated students to join me in working on exciting projects.

Email  /  Scholar  /  Github  

Research Interest

My research focuses on multi-modal learning, integrating point clouds, images, and wireless signals to advance robotic perception in areas like autonomous driving and quadruped robotics. In addition, I explore scene generation, scene-aware human motion and interactions with other agents, and signal simulation for diverse sensing modalities, with a broader goal of bridging the sim-to-real gap.

Selected Paper

We present MoScale, an autoregressive framework for hierarchical motion generation that integrates cross-scale hierarchical refinement and in-scale temporal refinement. Experiments demonstrate that MoScale achieves new SOTA performance and generalizes zero-shot to various motion generation and editing tasks.

We leverage LLM for diverse scene generation with corresponding human motion, and propose a phase-coherent ray tracing for RF signal simulation in a scalable way. It enables performance gain in both data-limited and data-adequate scenarios.

We propose an uncertainty quantification specifically designed for wireless signals and probabilistic learning for occupancy field training to enable high-resolution SLAM relying on RF signals only.

We present a frequency domain analysis framework that interprets the momentum method as a time-variant filter for gradients and propose a heuristic optimizer.

We propose acoustic volume rendering in the frequency domain to construct neural impulse response fields.

We explore whether LLMs could reduce the need for large-scale annotations for clinical reports.

We achieve velocity estimation with only one frame of point clouds.

Miscellaneous

• LEGO fan: Spent 7 hours building Batmobile.

Design and source code from Jonathon Barron's website.