Chubin Zhang
About Me
I’m currently a third-year Ph.D. student at Tsinghua University, supervised by Prof. Yansong Tang and Prof. Jiwen Lu. I aim to build generalist embodied intelligence that learns autonomously and solves complex tasks in the open world. My research explores how scaling training-time and inference-time compute can enable robots to learn generalizable capabilities from an embodied data pyramid spanning Internet knowledge, human behavior, simulation, and real-robot experience, and continually acquire new skills through autonomous exploration and interaction.
Prior to my PhD, I received a bachelor from the Department of Automation, Tsinghua University. I am also a member of Class of General Artificial Intelligence of Tsinghua University. In my undergraduate research, I am fortunate to work with Yueqi Duan at the Department of Electronic Engineering of Tsinghua University.
Publications & Preprints
-
Preprint
arXiv preprint, 2026TL;DR Scales robot policy learning with noisy egocentric actions by accounting for label reliability and distilling action-relevant knowledge from demonstrated future outcomes.
-
Preprint
arXiv preprint, 2026TL;DR Unifies world prediction and action generation in one causal model, showing that learning to predict action outcomes and learning to act improve each other.
-
Preprint
arXiv preprint, 2026TL;DR Improves compositional manipulation by decomposing demonstrations into reusable latent skills and retrieving relevant skill memories to guide action prediction.
-
ECCV
European Conference on Computer Vision (ECCV), 2026TL;DR Accelerates vision-language-action inference by forecasting deep-layer token importance, preserving task-critical visual information while pruning redundant tokens.
-
ICML
International Conference on Machine Learning (ICML), 2026TL;DR Improves robustness to unseen visual disturbances with a lightweight information bottleneck adapter that filters visual noise without extra data or augmentation.
-
Preprint
arXiv preprint, 2026TL;DR Transfers manipulation skills from human videos to robots by contrastively aligning visual transitions with an executable latent action vocabulary.
-
ICML
International Conference on Machine Learning (ICML), 2026TL;DR Jointly denoises continuous representations and discrete tokens to combine expressive latent reasoning with effective language-model training and generation.
-
ICRA
IEEE International Conference on Robotics and Automation (ICRA), 2026TL;DR Combines multi-view semantic features with 3D spatial perception and a dense action head to improve generalization and generate coherent whole-body manipulation.
-
CVPR
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026TL;DR Teaches video models to retrieve relevant frames while reasoning, using tool-augmented reinforcement learning for long-video question answering and temporal grounding.
-
IROS
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026TL;DR Improves pretrained vision-language-action models through scalable online reinforcement learning, using a robotic process reward model to guide exploration.
-
TIP
IEEE Transactions on Image Processing (TIP), 2025TL;DR Learns semantic 3D occupancy from multi-camera images without LiDAR or 3D supervision, combining neural rendering, photometric consistency, and 2D semantic cues.
-
NeurIPS
Advances in Neural Information Processing Systems (NeurIPS), 2024TL;DR Reconstructs detailed 3D Gaussian assets from multiple images using sparse 3D anchor points and geometry-aware attention, enabling efficient reconstruction with dense input views.
-
CVPR
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023TL;DR Learns generalizable semantic fields by aggregating context across reprojected image rays, using efficient attention to transfer semantic understanding to previously unseen scenes.
Powered by Jekyll and Minimal Light theme.