DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation
DreamX-Phi is a geometry-aware video world model built on Wan2.2-TI2V-5B. Given an initial observation and a prescribed robot action sequence, it predicts the corresponding future video while preserving robot motion, scene geometry, and the state of the manipulated object.
- Geometry-aware action conditioning for end-effector poses and gripper states.
- Interaction-aware learning for coherent scene and object evolution.
- DMD-based few-step generation for efficient inference.
- First place on WorldArena 2.0 Track 1 and tied for second on Track 2.
Model weights and inference code will be made publicly available after the WorldArena 2.0 IROS Challenge concludes.
Read the technical report on arXiv. If you find this work useful, please cite:
@article{team2026dreamx,
title = {DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation},
author = {Team, DreamX and Chen, Rui and Chu, Xiangxiang and Li, Geng and Li, Jifan and Shi, Qingfeng and Tang, Datao and Tang, Jing and Wang, Jun and Zhang, Pengfei},
journal = {arXiv preprint arXiv:2608.13489},
year = {2026}
}This repository is released under the MIT License.