Action Images: End-to-End Policy Learning via Multiview Video Generation

TL;DR

This paper introduces Action Images, a unified approach that uses multiview video generation with pixel-grounded action images for improved robot policy learning and transfer across environments.

Contribution

It proposes translating robot actions into interpretable multi-view videos, enabling zero-shot policy learning directly from video backbones without separate control modules.

Findings

Achieves the strongest zero-shot success rates on RLBench and real-world tasks.

Improves video-action joint generation quality over prior models.

Supports multiple functionalities like control, joint generation, and action labeling within a shared model.

Abstract

World action models (WAMs) have emerged as a promising direction for robot policy learning, as they can leverage powerful video backbones to model the future states. However, existing approaches often rely on separate action modules, or use action representations that are not pixel-grounded, making it difficult to fully exploit the pretrained knowledge of video models and limiting transfer across viewpoints and environments. In this work, we present Action Images, a unified world action model that formulates policy learning as multiview video generation. Instead of encoding control as low-dimensional tokens, we translate 7-DoF robot actions into interpretable action images: multi-view action videos that are grounded in 2D pixels and explicitly track robot-arm motion. This pixel-grounded action representation allows the video backbone itself to act as a zero-shot policy, without a separate policy head or action module. Beyond control, the same unified model supports video-action joint generation, action-conditioned video generation, and action labeling under a shared representation. On RLBench and real-world evaluations, our model achieves the strongest zero-shot success rates and improves video-action joint generation quality over prior video-space world models, suggesting that interpretable action images are a promising route to policy learning.