World Action Models are Zero-shot Policies

Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, Jiannan Xiang, Ayaan Naveed Malik, Kyungmin Lee, William Liang, Nadun Ranawaka Arachchige, Jiasheng Gu, Yinzhen Xu, Guanzhi Wang, Fengyuan Hu, Avnish Narayan, Johan Bjorck et al. (16 additional authors not shown)

Keywords: Robot Learning: Imitation Learning, Robot Learning: Model Learning, Robot Learning: World Model

TL;DR: DreamZero is a World Action Model for robot policy that shows state-of-the-art task generalization, with real-time control, and strong cross-robot transfer.

Abstract: State-of-the-art Vision-Language-Action (VLA) models excel at semantic generalization but struggle to generalize to unseen physical motions in novel environments. We introduce DREAMZERO, a World Action Model (WAM) built on a pretrained video diffusion backbone. Unlike VLAs, WAMs learn physical dynamics by predicting future world states and actions, using video as a dense representation of how the world evolves. By jointly modeling video and action, DREAMZERO effectively learns diverse skills from heterogeneous robot data without relying on repetitive demonstrations, resulting in over 2× improvement in generalization to new tasks and environments compared to state-of-the-art VLAs in real-robot experiments. Crucially, through model and system optimizations, we enable a 14B autoregressive video diffusion model to perform real-time closed-loop control at 7 Hz. Finally, we demonstrate cross-embodiment transfer in both directions: (1) video-only demonstrations from other robots or humans improve unseen task performance by over 40% with just 10–20 minutes of data, and (2) DREAMZERO adapts to entirely new embodiments, achieving zero-shot generalization on the YAM robot with only 30 minutes of play data.

Submission Number: 110