Jia-Peng Zhang张嘉鹏

Education

Selected Publications

• SkinTokens: A Learned Compact Representation for Unified Autoregressive Rigging Jia-Peng Zhang, Cheng-Feng Pu, Meng-Hao Guo, Yan-Pei Cao, Shi-Min Hu ACM Transactions on Graphics (TOG), 2026, Accepted with Minor Revisions (zjp-shadow.github.io/works/SkinTokens/) Introduced SkinTokens, a learned discrete representation for skinning weights. By framing skinning as a sequence prediction task, the unified TokenRig framework jointly models skeletons and skin deformations, achieving substantial improvements in rigorous out-of-distribution geometry evaluations.
• One Model to Rig Them All: Diverse Skeleton Rigging with UniRig Jia-Peng Zhang, Cheng-Feng Pu, Meng-Hao Guo, Yan-Pei Cao, Shi-Min Hu ACM Transactions on Graphics (TOG) (SIGGRAPH 2025) (zjp-shadow.github.io/works/UniRig/) Proposed UniRig, a unified framework for automated skeletal rigging leveraging large autoregressive models and a novel Skeleton Tree Tokenization method. Trained on Rig-XL (a new large-scale dataset), it efficiently generates topologically valid structures, far outperforming existing state-of-the-art tools across diverse generic models.
• CharacterGen: Efficient 3D Character Generation from Single Images with Multi-View Pose Canonicalization Hao-Yang Peng, Jia-Peng Zhang, Meng-Hao Guo, Yan-Pei Cao, Shi-Min Hu ACM Transactions on Graphics (TOG) (SIGGRAPH 2024), 43(4): 1–13, 2024 (charactergen.github.io/) Presented a highly efficient framework for lifting single images into pose-unified 3D character meshes. It utilizes an image-conditioned multi-view diffusion model for pose calibration and a transformer-based sparse-view network, producing geometrically and texturally robust outputs ready for animation.
• Transferring the Intelligence of VLMs to Robotic Control Meng-Hao Guo, Zhe-Han Mo, Jia-Jun Wang, Yi Zhang, Kejin Wang, Yi-Xuan Deng, Jia-Peng Zhang, Yongming Rao, Shi-Min Hu arXiv preprint arXiv:2609.22966, 2026 (robodawn.top/) Introduced RoboDawn, a human-intuitive interface that lets an agentic VLM control robots in a closed loop through discrete translation, rotation and gripper commands, grounded by a few in-context demonstrations. Without task-specific robot training it achieves state-of-the-art success rates on RoboTwin 2.0 C2R and RoboDojo, and transfers to a real Franka arm.
• TopoCap: Learning Topology-Agnostic Motion Priors for Monocular Video-to-Animation Cheng-Feng Pu, Jia-Peng Zhang, Meng-Hao Guo, Yan-Pei Cao, Shi-Min Hu SIGGRAPH 2026 Conference Papers (arxiv.org/abs/2606.12153) Introduced TopoCap, the first unified framework that extracts motion from monocular video and retargets it onto characters with arbitrary, unseen skeletal topologies, from bipeds to hexapods and inanimate objects, without test-time optimization. A Graph CVAE learns a topology-agnostic motion manifold, and conditional flow matching predicts its codes from video; trained on Mobjaverse, a new dataset with over 5,000 skeletal topologies and 2 million frames.

Open Source Projects

Honors and Awards

• Rising Star Pioneer Award, Huawei Kunpeng & Ascend Developer Conference (KADC)2026
• Comprehensive Scholarship for Ph.D. Students, Tsinghua University2025
• Silver Medal, National Olympiad in Informatics (NOI)2019
• Gold Medal, China Team Selection for Informatics (CTS)2019

Internship Experience