Yuntian Deng | 邓云天I am an undergraduate student in Electrical and Computer Engineering at Shanghai Jiao Tong University. I am a core member of the SJTU Supercomputer Team and an intern at Galbot. My interests lie at the intersection of Embodied AI and High-Performance AI Systems. I work on robot-learning algorithms, AI infrastructure, and efficient large-model inference—from research prototypes to system-level optimization, with the goal of building real-time-responsive, physics-aware embodied agents. |  |
Research Interests |
Embodied Intelligence. Vision-Language-Action models, world-action models, and learning methods that improve robotic manipulation in complex environments. Efficient AI Systems. Distributed inference, compilation, quantization, operator optimization, and CPU–GPU scheduling for scientific and foundation models. |
Publications |
| *: joint first author; †: project lead; ✉: corresponding author(s) |
 | DualWAM: Dual-System World Action Models for Asynchronous Global Planning and Local RefinementYixin Zheng*, Jiangran Lyu*, Yuntian Deng*, Kai Liu, Yizhou Zhou, Yizhou Wang, Xiaoguang Zhao, He Wang✉, Zhizheng Zhang✉Submitted to ICLR 2027 arXiv / project page An asynchronous dual-system world action model that combines global planning with wrist-guided local refinement, improving zero-shot manipulation success by 4.5 percentage points with a 16.6× critical-path speedup. |
 | Emerging Extrinsic Dexterity in Cluttered Scenes via Dynamics-aware Policy LearningYixin Zheng*, Jiangran Lyu*,†, Yifan Zhang, Jiayi Chen, Mi Yan, Yuntian Deng, Xuesong Shi, Xiaoguang Zhao, Yizhou Wang, Zhizheng Zhang✉, He Wang✉Robotics: Science and Systems (RSS), 2026 🏆 Best Paper Finalist arXiv / project page A dynamics-aware policy-learning framework that enables robots to exploit environmental contacts and object interactions in cluttered scenes. |
ExperienceFull CV → |
 | SJTU Supercomputer Team XflopsCore Member · 2024.10–PresentCompete in ASC and SCC International Supercomputer Contest. Optimize large-scale software through operator reconstruction, profiling, and multi-thread scheduling. Maintain the team’s open-source HPC Start Guide. Independently deployed DeepSeek-R1-671B on the SJTU "Zhiyuan-1" NPU Cluster. |
 | GalbotEmbodied AI Algorithm Research InternResearched World Action Model (WAM) for robotic manipulation. The work focused on improving robot-arm performance and generalization in complex real-world scenarios through embodied model and manipulation-pipeline research. |
Awards
|
Miscellaneous
|