Dex4D: Task-Agnostic Point Track Policy for Zero-Shot Dexterous Manipulation
Task-agnostic 'any pose to any pose' dexterous policy trained in sim on thousands of objects and conditioned on 3D point tracks; prompted with tracks from generated videos, it runs real tasks zero-shot.在仿真中用数千种物体训练与任务无关的“任意位姿到任意位姿”灵巧策略，以 3D 点轨迹为条件；用生成视频提取的点轨迹作提示，即可零样本完成真实任务。
BibTeX
@misc{kuang2026dex,
title = {{Dex4D: Task-Agnostic Point Track Policy for Zero-Shot Dexterous Manipulation}},
author = {Yuxuan Kuang and Sungjae Park and Katerina Fragkiadaki and Shubham Tulsiani},
year = {2026},
eprint = {2602.15828},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2602.15828},
}