Huaze (Patrick) Liu

Huaze Liu 刘铧泽

Hi, I'm Huaze, also go by Patrick. I'm a visiting student at Georgia Institute of Technology working on humanoid whole-body control and tactile sensing, advised by Prof. Ye Zhao.

Previously, I graduated from Harvey Mudd College with high distinction, where I worked with Prof. Adyasha Mohanty on sensor fusion for trustworthy autonomous navigation. I also worked on humanoid robot control and dexterous manipulation under the supervision of Prof. Michael Tolley and Prof. Hao Su at UC San Diego.

I work on robotics control, dexterous manipulation, and reinforcement learning. Broadly speaking, I'm interested in how systems perceive and acts in the real world robustly.

Email / Github / Linkedin / Google Scholar

Research

My research focuses on:

Dexterous Manipulation — integrating model-free learning with physics-based simulation, optimization, and control to solve contact-rich dexterous manipulation efficiently and reliably.
Whole-Body Manipulation — integrating tactile sensing with whole-body control to solve manipulation tasks beyond table-top settings.

Publications & Preprints
	Catch It All!: Generalizable Dynamic Catching with a Dexterous Hand Kehlani Fay, Arth Shukla, Huaze Liu, Darin Djapri, Alexiy Buynitsky, Nicklas Hansen 4th RSS Workshop on Dexterous Manipulation [pre-release project page] / [pre-Arxiv]
	Language-Driven Multimodal Semantic Change Detection in Urban Maps Huaze Liu , Zihao Gao , Adyasha Mohanty Proceedings of the Institute of Navigation GNSS+ conference (ION GNSS+ 2025) [Paper] / [Slides] / [Code]
	Conformal Prediction for Reliable Test-Time Uncertainty in Robotic Vision Huaze Liu, Adyasha Mohanty Southern California AI and Robotics Symposium 2025 [Paper] / [Paper (Long version)]

Featured Project
Beyond robotics, I'm also interested in AI agents, computer vision, and ML theory. I regularly maintain and build new projects here.
	ProxWeld for Vulcan OmniPro 220 [Code] Built a manual-grounded AI support agent for the Vulcan OmniPro 220 welder using retrieval over bundled PDF manuals and the Claude Agent SDK, returning page-level citations, manual visuals, and structured wiring/troubleshooting helpers through a Next.js chat UI.
Input Output	Automated Video Object Eraser Pipeline [Code] A reproducible workflow for removing dynamic objects from videos by integrating YOLO for detection, OSTrack for tracking, SAM2 for segmentation, and DiffuEraser for diffusion-based inpainting. The system ensures temporal consistency and high-quality background restoration.
	Vision-based Localization and Path Planning TurtleBot System [Report] / [Code] Developed a dual-robot system for autonomous emergency response, where an "explorer" robot uses vision-based detection and an EKF/UKF to locate targets and a "rescuer" robot uses the A* algorithm to navigate to them. The framework is validated in the Webots simulator.
	Robotics CAD Collection for low-cost robot platforms — Open-source SolidWorks Parts [Code] A growing collection of SolidWorks models and assemblies (grippers, robot arms, and more) aimed at enabling affordable, open-source robotic platforms. Future additions planned: small vehicles, mobile manipulators, multi-legged robots, and more.
	Transfer Learning of DETR for Traffic Object Detection [Report] / [Code] Fine-tuned DETR (End-to-End Object Detection with Transformers) for traffic object detection task using KITTI object detection dataset, achieving 71.1% mAP with 18.1ms latency.
	A Multi-Objective Acquisition Function of Bayesian Optimization [Report] We developed a Scalarized Upper Confidence Bound function for Bayesian Optimization in hyperparameter tuning. It was tested by benchmarking on SVM and CNN models, reducing latency by 8% at equivalent accuracy compared to standard methods.