Cycle WAM

World action models Video ↔ Action

Cycle WAMCoupling Action and Video
through Cycle Consistency

* Corresponding authors.

Paper arXivWatch demoCode Coming soonModels Coming soon

Learning the correspondence between visual change and executable action.
Music: “Cylinder Five” by Chris Zabriskie · CC BY 4.0 · Credits & edits

Download video

Two modalities.
One interaction.

Actions change the world. Video reveals those changes. Cycle WAM learns the correspondence in both directions.

Building on a video-pretrained world action model, we augment joint video–action generation with bidirectional conversion to learn forward and inverse dynamics. A differentiable cycle-reconstruction objective then connects the two mappings: infer actions from video, and use those actions to reconstruct the video.

The resulting model supports action-conditioned visual prediction, action inference from video, and robot control within a shared framework.

Joint video-action generation, bidirectional conversion, and video-action-video cycle reconstruction
Joint generation → bidirectional conversion → cycle reconstruction. Action trajectories and reconstruction in this conceptual figure are schematic.

Convert. Reconstruct.
Couple the two.

One shared WAM learns how an action looks,
and which actions explain a video.

Observed futureVideoVisual changes
Inverse dynamicsVideo to action
Inferred commandsActionRobot motion
Forward dynamicsAction to video
Predicted consequenceVideoReconstruction
↖

Cycle consistency compares the reconstruction with the source

↗
1

Joint generation

Co-train future video and actions using a pretrained video backbone.

2

Bidirectional conversion

Train video-to-action and action-to-video mappings with paired supervision.

3

Cycle reconstruction

Reconstruct the video through inferred actions, with gradients through both queries.

Training architecture

The co-training, bidirectional-conversion, and cycle objectives share a video-pretrained WAM backbone
The diagram shows modality-level tasks. Reconstruction uses the native video flow-matching target; a supervised inverse pass retains action supervision.

From visual change
to physical motion.

Approach, grasp, transport, release.
Recorded tabletop manipulation.

4.75× recorded speed

Place the remote control in the basket

Coordinated arm and gripper motion through grasping and placement.

4.75× recorded speed

Pick up the lemon and put it in the basket

The sequence shows transport and release above the basket; the landing is occluded.

Selected qualitative demonstrations. Across five evaluated real-world scenes, the paper reports task completion in 3/5 scenes for Cycle WAM and 0/5 for DreamZero. These counts do not establish repeated-trial reliability.

Predict the visual
consequence of an action.

Recorded simulator interaction and model predictions.
LIBERO Spatial, task 0.

Simulator interaction & predictions

Download video ↓

Left: recorded simulator interaction. Right: GT at the top, UWM at bottom left, Cycle WAM at bottom right.

Matched prediction clips

Four matched segments · Two camera views per panel

Ground truthUWMCycle WAM

Cycle WAM retains sharper visual boundaries in these examples. Object placement and view composition still differ from ground truth. Each panel contains two camera views; original frames are preserved without interpolation.

Evidence across
control and prediction.

Reported values from the accompanying paper.
Read the PDF for protocols and full ablations.

17.8%

Average success on four selected conditions

Spatial-T, Spatial-S, Goal-T, and Goal-S · 100 episodes per condition

Success percentages on four selected LIBERO-Pro conditions
MethodSpatial-TSpatial-SGoal-TGoal-SAverage
π0.51.020.02.038.015.3
UWM †21.01.03.00.06.3
DreamZero †27.013.012.02.013.5
Cycle WAM24.025.019.03.017.8

Selected rows from Table I. † LIBERO-130-finetuned checkpoints evaluated alongside Cycle WAM. The average covers only these four conditions. Gains are not uniform across conditions or aggregate control settings.

Acknowledgments

We thank Yifei Huang from Nanjing University for assistance with real-robot deployment.

Read the full study.

The PDF includes the method, theoretical analysis, evaluation protocols, and limitations.

Read the paper

Code and model weights are coming soon.

BibTeX arXiv · 2026
@misc{yang2026cyclewam,
  title = {Cycle WAM: Coupling Action and
           Video through Cycle Consistency},
  author = {Qinglin Yang and Jiepeng Wang and
            Yuxing Huang and Haibin Huang and
            Kai Xu and Chi Zhang and
            Ruizhen Hu and Xuelong Li},
  year = {2026},
  eprint = {2611.04711},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2611.04711}
}

arXiv:2611.04711 · cs.CV