Xinyuan Li

I am a junior student at Zhejiang University, majoring in Control Science and Engineering. In addition, I am enrolled in the ACEE Honors Program at Chu Kochen Honors College and have completed extensive coursework in Computer Science.

Research Interests

• Foundation Models: Exploring VLMs and UMMs to enhance reasoning capabilities.
• Embodied AI & Robotics: Developing VLAs models and generalist agents capable of executing complex tasks in open-world environments.
• World Models & Video Generation: Leveraging Video Generation as scalable World Models to serve as interactive environments for VLA training and RL reward. I am particularly interested in:
  • Physical Grounding: Improve the physical consistency of video generative models to reflect real-world dynamics accurately.
  • Multiview Consistency: Ensuring 3D-aware spatial coherence across camera perspectives to build robust scene representations.
  • Efficiency: Optimizing Inference acceleration for video generation to enable high-throughput simulation for reinforcement learning.

Research Projects

• Cross-view Spatial Consistency for UMMs | ongoing research
  • Co-developed the research framework and iterated on methodology with senior PhDs.
  • Built data processing pipeline and conducted model training process.
  • Established a comprehensive evaluation suite by integrating benchmarks (e.g., MindCube, MMSI, CrossPoint) to rigorously assess model performance.
  • Collaborating with Prof. Manling Li. Targeting ECCV 2026!