I am Xinghao Chen, a master’s student at the University of Washington.

My research interests focus on computer vision and generative AI, and I have worked on projects involving 3D Gaussian splatting (3DGS), diffusion models, LLM agents, image segmentation, and reinforcement learning.

I am actively looking for PhD positions (Spring & Fall 2026)! If you are interested in working together or have potential PhD opportunities, please feel free to contact me. Edu email

News

• [04.2025] Started the Developing Immersive Experiences for AR/VR course, where I am building interactive 3DGS experiences for VR headsets.
• [01.2025] Joined the TACO Group at Texas A&M University as a remote research assistant under Dr. Zhengzhong Tu.
• [12.2024] Started working on the “Machine Learning for Community-Driven Coastal Erosion Monitoring and Management” project at the UW Applied Physics Laboratory with advisers Dr. Roxanne Carini, Dr. Morteza Derakhti and Dr. Arindam Das.
• [12.2024] Enrolled in the Large Language Models course taught by Dr. Karthik Mohan. Course web page.
• [09.2024] Enrolled in the Computer Vision course taught by Dr. Stan Birchfield.
• [09.2024] Began the M.S. program in Electrical and Computer Engineering at the University of Washington.

Ongoing Research

RL-guided image restoration agent (multi-metric NR-IQA rewards)

We are building a reinforcement learning agent that sequences specialist restoration tools under mixed, coupled degradations. Unlike JarvisIR’s MRRHF ranking stage, we use Best-of-K GRPO/PPO: for each image the policy samples multiple plans, executes them end to end, and learns from group-relative advantages computed from a multi-metric NR-IQA reward (NIQE, BRISQUE, MUSIQ, ImageReward). The objective directly maximizes expected quality - cost, with an explicit stop action, small step/time penalties, and priors that discourage over-sharpening or over-smoothing.

The agent conditions on visual embeddings, NR-IQA signals, and recent actions, and acts over a toolbox (denoise, deblur, derain/snow, dehaze/low-light, super-resolution, colorization, face restoration). Training can warm-start from CleanBench-Synthetic imitation, then optimize on CleanBench-Real and broader non-driving data; we randomize metric weightings to avoid gaming and can target CVaR-style worst-case improvements. We are also exploring 4K-Agent ideas (Q-MoE routing, rollback) to further boost robustness and efficiency.

Open-source NuRec-style scene reconstruction

Re-creating NVIDIA’s recent Omniverse NuRec + 3DGUT workflow with an open toolchain so multi-sensor logs (RGB, LiDAR, IMU) can be converted into photorealistic 3D Gaussian scenes for Isaac Sim and CARLA. The roadmap covers Gaussian-based reconstruction, USD export, and lightweight viewers that mirror NuRec features like Physically Accurate Dataset streaming and Sensor RTX-quality ray tracing.

The study leans on NVIDIA’s August 2025 developer demo and Newsroom briefing-where NuRec shipped with CARLA integrations, Foretellix toolchain support, and Voxel51 FiftyOne dataset hooks-to map required components before swapping in open kernels (e.g., gsplat, Nerfstudio, differentiable rendering).

Education

University of Washington

September 2024 ~ January 2026 (Expected)

M.S. Electrical & Computer Engineering

Henan University

September 2020 ~ June 2024

B.E. Automation

Contact

Personal email: xhc42@outlook.com

Edu email: cxh42@uw.edu

WeChat: ICXH42