Kaifeng Zhang

I am a Ph.D. student in computer science at Columbia University, advised by Prof. Yunzhu Li. Prior to this, I obtained my Bachelors degree from Tsinghua University (Yao Class). I am fortunate to receive mentorship from Prof. Kris Hauser during my Ph.D. study, and Prof. Xiaolong Wang, Prof. Yang Gao, Prof. Li Yi during my undergrad.

My research sits at the intersection of robotics, 3D vision, physics simulation, and machine learning. I am interested in bridging the gap between robotic simulation and the real world for robust and scalable robot manipulation.

If you'd like to discuss research opportunity, collaboration, Ph.D. application, or anything related, feel free to reach out via email: kaifeng dot z at columbia dot edu.

Google Scholar  /  Github  /  Twitter  /  LinkedIn  /  CV

Photo credit: Bingjie Tang

We propose a framework for robot policy evaluation in simulation environments, using Gaussian Splatting for rendering and soft-body digital twin for dynamics.

We optimize a spring-mass physics model of deformable objects and integrate the model with 3D Gaussian Splatting for real-time re-simulation with rendering.

We propose a neural particle-grid model for training dynamics model with real-world sparse-view RGB-D videos, enabling high-quality future prediction and rendering.

We learn neural dynamics models of objects from real perception data and combine the learned model with 3D Gaussian Splatting for action-conditioned predictive rendering.

We learn a material-conditioned neural dynamics model using graph neural network to enable predictive modeling of diverse real-world objects and achieve efficient manipulation via model-based planning.

We achieve 4D neural implicit reconstruction from only a single-view scan using deformation and topology regularizations.

We propose a fully self-supervised method for category-level 6D object pose estimation by learning dense 2D-3D geometric correspondences. Our method can train on image collections without any 3D annotations.

We show that fusing fine-grained features learned with low-level contrastive objectives and semantic features from image-level objectives can improve SSL pretraining.

Template borrowed from Jon Barron