World Action Models are Zero-shot Policies

arXiv 26.2 2026 64.4 method, application

TLDR

DreamZero uses a video diffusion world action model for zero-shot policy learning, achieving 2x generalization improvement and real-time control at 7Hz.

Reasoning

The paper introduces a novel World Action Model (WAM) that outperforms state-of-the-art VLAs in generalization and enables real-time closed-loop control. Strengths include cross-embodiment transfer and use of heterogeneous data, but the abstract lacks detailed ablation studies or comparisons to other world model approaches.

Read-first score

Read-first score 64.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 53.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 42%

75.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=experiment,result

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=github

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 295.

Keyword Scores

world model

video world model

world dynamics prediction

generative world model

interactive world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

DreamZero as a World Action Model (WAM) built upon a pretrained video diffusion backbone, jointly modeling video and action to learn diverse skills from heterogeneous robot data without repetitive demonstrations.
Enables real-time closed-loop control at 7Hz for a 14B autoregressive video diffusion model through model and system optimizations.
Demonstrates cross-embodiment transfer: video-only demonstrations from other robots or humans yield over 42% relative improvement on unseen task performance with just 10-20 minutes of data.
Enables few-shot embodiment adaptation with only 30 minutes of play data while retaining zero-shot generalization.

Methodology

DreamZero is a World Action Model built upon a pretrained video diffusion backbone. It learns physical dynamics by predicting future world states and actions, using video as a dense representation of how the world evolves. The model jointly models video and action, trained on heterogeneous robot data, and employs model and system optimizations to enable real-time closed-loop control at 7Hz for a 14B autoregressive video diffusion model.

Key Results

DreamZero achieves over 2x improvement in generalization to new tasks and environments compared to state-of-the-art VLAs in real robot experiments. Cross-embodiment transfer with video-only demonstrations yields a relative improvement of over 42% on unseen task performance with just 10-20 minutes of data, and few-shot embodiment adaptation with 30 minutes of play data retains zero-shot generalization.

Resources

Paper Project