Ph.D. Student in Artificial Intelligence Xi'an Jiaotong University & Zhongguancun Academy

I am a Ph.D. student in Artificial Intelligence at Xi'an Jiaotong University and Zhongguancun Academy.

My research focuses on computer vision, multimodal fusion, and embodied perception, with a particular interest in building structured scene representations for reconstruction, simulation-to-reality transfer, and robotic manipulation.

I am currently working on Gaussian Splatting-based representations, including transparent surface reconstruction, semantic 2D Gaussian fields, and dynamic 3D Gaussian fields for robot action prediction.

ACM Multimedia · Second Author

Transparent objects are difficult for 3D Gaussian Splatting because low-opacity Gaussian particles often make depth rendering incorrectly capture the background.

In this work, we improve the depth rasterization process by introducing a sliding-window strategy in the CUDA rendering pipeline. The method searches for local regions with maximum opacity and uses them to recover more reliable transparent surface depth.

Under Submission

Simulation and real-world environments often contain a significant domain gap, which limits the direct deployment of robot policies trained in simulation.

We use semantic 2D Gaussian Splatting as a bridge between the robot and the environment. The constructed semantic 2DGS field provides cross-scene geometric and semantic representations for robot models. During manipulation, the field can be dynamically updated through real-time environmental inputs, enabling more effective sim-to-real transfer while reducing the dependence on real-world robot data.

Ongoing Work

Accurate future scene prediction is important for safe and reliable robotic manipulation.

In this project, we explore a feed-forward dynamic 3D Gaussian field for predicting next-frame scene changes. By incorporating optical-flow priors and rigid-body constraints of robotic arms, we aim to make Gaussian particle motion physically meaningful and directly infer robot actions from particle dynamics.

• Gaussian Splatting for scene representation

  • Transparent surface reconstruction
  • Depth rendering and CUDA rasterization
  • 2D/3D Gaussian scene fields

• Embodied perception for robotics

  • Dynamic scene prediction
  • Robot manipulation from visual-geometric representations
  • Action inference from Gaussian particle motion

• Sim-to-real transfer

  • Semantic 2D Gaussian fields
  • Cross-domain geometric representations
  • Real-time environment-aware field updates

• Multimodal and geometric perception

  • Computer vision
  • Multimodal fusion
  • High-level scene understanding for embodied agents

• First Prize, Brain-Controlled Robotics Track, Brain-Computer Interface Competition, 2025

I am currently focusing on a compact but connected research direction:

Gaussian-based scene representations for embodied perception and robotic manipulation.

In practice, this means building dynamic, semantic, and physically meaningful representations that can support reconstruction, prediction, sim-to-real transfer, and robot action generation.

Some research repositories are being organized before public release.

I prefer to release projects with clear documentation, reproducible instructions, and concrete demos rather than raw experiment code.