Zhenghao Peng

Chinese University of Hong Kong, University of California, Los Angeles

Papers

Total Citations

H-Index

About

Zhenghao Peng is a researcher at the forefront of autonomous driving and human-AI interaction. His work centers on bridging the critical gap between simulation and real-world deployment for intelligent agents. Peng’s most impactful contribution is **“Learning to Drive by Watching YouTube Videos,”** which introduces an action-conditioned contrastive policy pretraining method. This pioneering approach allows autonomous driving models to learn complex behaviors directly from raw, unlabeled human driving footage, bypassing the need for expensive, manually annotated datasets. With 18 citations, this work has become a key reference for data-efficient policy learning. Peng further addresses the persistent sim-to-real challenge with **“Vid2Sim,”** a framework that transforms real-world video into realistic, interactive simulations for urban navigation. This work tackles the limitations of domain randomization by creating high-fidelity digital twins from video, enabling safer and more effective robot policy training. Additionally, his research on **“Human-AI Shared Control via Policy Dissection”** explores how to decompose learned policies for intuitive human oversight, allowing non-experts to collaborate with and guide AI agents in complex tasks. Through these contributions, Peng is shaping a future where autonomous systems learn more naturally and operate more reliably in the real world.