YuZhou Wu

YuZhou Wu

MSc Robotics student at the University of Sheffield. Research interests in Vision-Language-Action models, 4D perception, and efficient Transformers.

YuZhou Wu

Hi! I am YuZhou Wu, a Master’s student in Robotics at the University of Sheffield (expected Sept 2025).

My research interests lie at the intersection of Robotics, Vision-Language-Action (VLA) models, and efficient spatiotemporal representation learning, with a particular focus on 4D perception, multi-agent reasoning, and Transformer acceleration for real-time embodied systems.

I am actively preparing for PhD applications and am interested in research-oriented positions related to robot learning, embodied AI, and large-scale multimodal models.

🔍 Research Interests

My current research focuses on:

  • Vision-Language-Action (VLA) Models
    • Policy learning for embodied agents (e.g., LIBERO, DROID-style datasets)
    • Action tokenization, latent action modeling, and policy generalization
    • Streaming inference and real-time deployment constraints
  • 4D Point Cloud Perception & Transformers
    • PointTransformer / SpUNet architectures for dynamic scenes
    • Cross-frame attention and KV-Cache enhanced Transformers
    • World-coordinate alignment and temporal token reuse
  • Efficient & Interpretable Deep Models
    • Token pruning, attention sparsification, and quantization-aware training
    • KV cache reuse for streaming perception
    • Model compression for deployment on resource-constrained systems

🧪 Selected Research Projects

KV-PT: Key-Value Cache Enhanced Point Transformer for 4D Space Analysis

  • Proposed a temporal KV-cache mechanism for PointTransformer-style models
  • Enables cross-frame attention with nearest-neighbor and max-attention token reuse
  • Applied to SemanticKITTI sequential LiDAR data

IPCV: Information-Preserving Compression for Vision Encoders

  • Proposed an information-preserving compression framework for vision encoders, enabling token reduction without sacrificing semantic fidelity
  • Designed structure-aware token pruning and feature reweighting mechanisms, guided by representation consistency constraints
  • Achieved strong performance–efficiency trade-offs across multiple vision backbones, improving inference speed and memory usage

Multi-Agent Motion Prediction & Decision-Making Survey

  • Systematic taxonomy of multi-agent prediction and planning methods
  • Comparative analysis on Waymo Open Motion & Argoverse2
  • Covers game-theoretic, MARL, intent-sharing, and V2X-based approaches

🛠 Technical Skills

  • Programming: Python, C++, MATLAB
  • Deep Learning: PyTorch, TorchScript, ONNX
  • Robotics: ROS2, Nav2, SLAM, TurtleBot3
  • 3D / 4D Vision: PointTransformer, SpConv, SemanticKITTI
  • Tools: Git, Docker, Conda, VSCode Remote, tmux

📌 Current Goals

  • Submitting papers to CVPR / ECCV / NeurIPS / ACL
  • Applying for PhD programs in Robotics / Embodied AI
  • Exploring efficient VLA models for real-time robotic control

Feel free to explore my publications, projects, and code repositories, or reach out if you are interested in collaboration.