FlowWAM: Optical Flow as a Unified Action Representation for World Action Models

2026-07-14

AI Summary

Existing action representations struggle to simultaneously align with pretrained video generators and provide sufficient temporal motion information. This work proposes FlowWAM, which introduces optical flow as a unified, video-native action representation for the first time. FlowWAM employs a two-stream diffusion framework that jointly models RGB frames and optical flow, enabling large-scale pretraining on unlabeled videos without requiring action annotations. The approach naturally supports both policy mode (action prediction) and world model mode (future video generation). Experiments demonstrate that FlowWAM achieves success rates of 92.94% (Clean) and 92.14% (Random) on RoboTwin, significantly outperforming VLA and WAM baselines, and attains an EWM Score of 63.71 on WorldArena, improving trajectory accuracy by 18.4%.

Abstract

World Action Models (WAMs) are able to leverage pretrained video generators for both world modeling and action prediction. However, directly leveraging such video generators for control raises a new challenge: how to represent actions in a suitable form that aligns with pretrained video generators while carrying enough motion cues for accurate control. Existing numerical actions fail to satisfy the former, and prior visual action representations overlook the temporal motion structure across frames. We address this issue with FlowWAM, a dual-stream diffusion framework that adopts optical flow as a unified, video-native action representation. Flow videos share the same format as RGB videos and encode rich per-pixel displacement. By jointly modeling them within a shared pretrained video generator, FlowWAM can naturally implement two modes of WAMs. In policy mode, FlowWAM generates flow for action prediction, while in world-model mode, it uses target flow sequences to guide future video generation. Moreover, since flow can be easily extracted from raw videos without action labels, FlowWAM can leverage large-scale action-unlabeled video datasets for pretraining. We empirically find that our flow-based action representation delivers gains across both modes. On RoboTwin manipulation, FlowWAM raises the success rate to 92.94% on the Clean setting and 92.14% on Random, outperforming both VLA and WAM baselines. On WorldArena world modeling, it achieves the best overall EWMScore (63.71) with an 18.4% relative improvement in trajectory accuracy. More results can be found on our project website: https://flow-wam.github.io .

Problem

• World Action Models
• action representation
• optical flow
• video generation
• motion cues

Innovation

• optical flow
• world action models
• video-native action representation
• dual-stream diffusion
• action prediction