Spatial-UMI-Dream: Scaling World Action Models with Metric Geometry for UMI Manipulation
5B world action model pretrained on 605 h of RGB-D UMI data that predicts RGB, metric depth and actions. Adds RoboTwin-UMI; reports 85% average real success, +24.2 points over the best baseline.在 605 小时 RGB-D UMI 数据上预训练的 5B 世界动作模型，同时预测 RGB、度量深度和动作。并提出 RoboTwin-UMI 基准；报告真机平均成功率 85%，比最强基线高 24.2 个百分点。
BibTeX
@misc{yuan2026spatial,
title = {{Spatial-UMI-Dream: Scaling World Action Models with Metric Geometry for UMI Manipulation}},
author = {Weiduo Yuan and Ze Yuan and Sihan Liu and Jionghao Wang and Liang Heng and Zhenyu Zhao and Yue Wang},
year = {2026},
eprint = {2611.07229},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2611.07229},
}Manipulation & Dexterous Hands操作与灵巧手World Models世界模型Datasets & Benchmarks数据集与基准umiwamrgb-dmetric-depthbenchmarkrealsim