Zhiqi Huang

Zhiqi Huang

World Models for Physical AI

Zhiqi Huang

I am a Research Assistant at The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen) and an M.Phil. student in Information Architecture at Waseda University. I study world modeling and simulation for Physical AI. My research develops simulation-ready world representations, action-conditioned world simulation, and learning systems that turn model-generated experience into real-robot capability.

My background spans real-time graphics, controllable 3D generation, physically based appearance, robust semantic perception, and structured spatiotemporal prediction. From 2021 to 2024, I built cross-platform rendering and PBR pipelines at 4399 Games. I now apply these foundations to real-to-sim-to-real robot learning.

News

  • [Jul. 2026] Joined CUHK-Shenzhen as a Research Assistant.
  • [2026] Three manuscripts are under review, including two first-author works.
  • [2026] SIE3D was published at IEEE ICASSP 2026.

Research Direction

World Models for Physical AI: Representation, Simulation, and Action

My central question is: how can world models provide scalable and reliable experience for training, evaluating, and improving robots in the physical world?

I work at the interface of generative world modeling, physics-based simulation, and robot learning. My research agenda covers three connected directions:

  • Simulation-ready world representations: build structured and editable models of real environments from visual observations. These representations capture geometry, semantics, articulation, material appearance, and physical properties, with robust perception across sensor and distribution shifts.
  • Action-conditioned world simulation: generate multimodal rollouts of future states conditioned on robot actions. Neural dynamics and physical simulation capture contact, deformation, uncertainty, and counterfactual outcomes.
  • World-model-driven robot learning: use simulated and imagined trajectories for policy pretraining, planning, evaluation, and adaptation. Real robot data drives joint improvement of the world model and policy.

These directions form a closed-loop real-to-sim-to-real data engine across world reconstruction, counterfactual simulation, policy learning, deployment, and model update. Evaluation uses real-world task success, robustness, policy ranking, and sim-real correlation.

My long-term direction includes world-action models that jointly predict future world states and executable robot actions.

Publications

SIE3D: Single-Image Expressive 3D Avatar Generation via Semantic Embedding and Perceptual Expression Loss

IEEE ICASSP 2026 - First Author, Corresponding Author
Zhiqi Huang, Dulongkai Cui, Jinglu Hu

SIE3D generates an editable 3D Gaussian head avatar from one image, preserves identity, and supports language-level control over expressions and accessories. The work established my foundation in multimodal conditioning, structured 3D representations, and invariant-preserving editing.

Project Page IEEE Xplore arXiv Code

Manuscripts Under Review

  • Controllable PBR material generation from long-form descriptions - First author; manuscript under review. Generates geometry-aligned, relightable albedo, roughness, and metallic channels from detailed material specifications. The work develops structured appearance representations for relighting and controllable visual simulation.
  • Deform-then-edit forecasting for longitudinal 3D CT - First author; manuscript under review. Forecasts localized volumetric change and preserves stable anatomy. Its deform-then-edit transition model separates coherent transport from localized residual change. This structured transition view motivates my work on action-conditioned world simulation.
  • Robust pseudo-labeling under imaging noise and long-tailed data - Second author and corresponding author; manuscript under review. Develops feature-threshold dual calibration for rare-class semantic segmentation across sonar, underwater, and adverse-weather imagery. The work contributes robust semantic perception for simulation-ready world representations under sensor and distribution shifts.

Experience

The Chinese University of Hong Kong, Shenzhen

Research Assistant - Jul. 2026 - Present

  • Current focus: simulation-ready world representations and action-conditioned world simulation for robotic manipulation.

4399 Games

Graphics Engineer / Senior Graphics Engineer - 2021 - 2024

  • Promoted to Senior Graphics Engineer in 2023; led a rendering team of 3-5 engineers and owned the rendering roadmap for Era of Conquest on mobile and PC.
  • Built and optimized its real-time rendering pipeline, including cross-platform shaders, PBR material representation, and mobile performance.
  • Developed reusable asset, material, and rendering pipelines for interactive virtual worlds across mobile, PC, and web platforms.

Education

  • Waseda University, Fukuoka, Japan
    • M.Phil. in Information Architecture, Apr. 2025 - Mar. 2027 (expected)
    • English-taught program; current GPA: 3.8 / 4.0
  • Sun Yat-sen University, Guangzhou, China
    • B.E. in Software Engineering, 2017 - 2021
    • GPA: 3.7 / 4.0

Technical Background

  • Graphics and systems: C++, C#, GLSL/HLSL, Unity, Vulkan, OpenGL, real-time rendering, PBR, cross-platform optimization
  • Generative and 3D learning: Python, PyTorch, diffusion models, 3D Gaussian Splatting, CLIP/LongCLIP, mesh and material processing, multi-view evaluation, volumetric modeling, structured spatiotemporal prediction

Languages and Honors

  • Chinese: Native (Mandarin and Cantonese)
  • English: Professional working proficiency (TOEFL iBT: 90)
  • Outstanding Student Scholarship (Third Prize), Sun Yat-sen University, 2018-2019