Henghui Bao

I'm a master student majoring Computer Science at University of Southern California(USC), advised by Prof. Laurent Itti and Prof. Yue Wang, and I have also had the privilege of working under the guidance of Prof. Jiayuan Gu.

My research interests include: 3D Computer Vision, Vision for Robotics. I am interested in robust robot learning with minimal human supervision and generating photorealistic 3D environments from visual data to enhance simulation and real-world applications.

Email  /  Scholar  /  Github

Research

MoonSim is a versatile lunar simulator integrating physics-based texture synthesis, diverse topographical geometries, and a hybrid MuJoCo-Unreal Engine framework for high-fidelity robotic learning and research.

A zero-shot, multi-granularity pipeline that generates non-exclusive semantic labels by combining super points and 2D foundation models for precise part segmentation and meaningful naming, enabling seamless integration of fine-grained object segmentation with 3D language understanding.

USCILab3D is a densely annotated outdoor dataset of 10M images and 1.4M point clouds captured over USC's 229-acre campus, offering multi-view imagery, 3D reconstructions, and 267 semantic categories for advancing 3D vision and robotics research.

Value Explicit Pretraining (VEP) is a method for transfer reinforcement learning that pre-trains an objective-conditioned encoder invariant to environment changes, using contrastive loss and Bellman return estimates to create temporally smooth, task-focused representations, achieving superior generalization and sample efficiency on realistic navigation and Atari benchmarks.

Beogym is a data-driven simulator leveraging Gaussian Splatting and a large-scale outdoor dataset to render realistic first-person imagery with smooth transitions, enabling advanced training and evaluation of autonomous agents for visual navigation.

Services

Conference reviewer: CoRL 2024 Workshop XE