What is a world action model?

Series 01 · Module 05 · Lesson 1. Builds on Module 03 (VLAs) and Module 04 (world models).

Three questions a robot model can answer

Write $o$ for what the robot sees now, $l$ for the instruction, $a$ for the action and $o'$ for what it will see next. The model families in this series differ in what they predict:

Family	Learns	Answers
VLA	$p(a \mid o, l)$	What should I do?
World model	$p(o' \mid o, a)$	What happens if I do this?
World action model (WAM)	$p(o', a \mid o, l)$	What should I do, and what will it look like?

The notation is conceptual. Real models also condition on history, several cameras and robot state, and V-JEPA 2 predicts $o'$ as an embedding rather than pixels.

The chain rule gives two designs, and a WAM learns both at once

The joint distribution factors two ways:

$$p(o', a \mid o, l) \;=\; \underbrace{p(a \mid o, l)}_{\text{VLA}}\;\underbrace{p(o' \mid o, a, l)}_{\text{world model}} \;=\; \underbrace{p(o' \mid o, l)}_{\text{video generator}}\;\underbrace{p(a \mid o, o', l)}_{\text{inverse dynamics}}$$

The first factorization is "act, then imagine the consequence": a VLA plus a world model. The second is "imagine the outcome, then work out the action". UniPi (2023) built a policy that way: generate a video of the task, then recover actions with an inverse-dynamics model. A world action model trains one network on the joint distribution, so action prediction and video prediction share every parameter.

Why predict video if you only need actions?

A demonstration gives a VLA a few numbers per timestep to learn from: the action targets. Predicting the next frames gives far denser supervision, because each frame shows how the whole scene responded. The DreamZero paper frames the trade-off this way: VLAs inherit semantic knowledge from their vision-language backbone but generalize poorly to unseen physical motions, while a model that predicts the world has to learn dynamics.

Video is also a representation every robot shares. That is what makes two of the reported results possible:

• Learning from video-only demonstrations. DreamZero reports that 10–20 minutes of video from humans or other robots, with no action labels, improved unseen-task performance by over 42%.
• New bodies. DreamZero adapted to the YAM robot with 30 minutes of play data. Dream-1 claims zero-shot control of robots it has never seen (company-reported).

Two instances

DreamZero (NVIDIA GEAR Lab, February 2026; arXiv:2602.15922). A 14B autoregressive video diffusion model built on Wan2.1-I2V-14B. It runs closed-loop at a reported 7 Hz after model and system optimizations. On AgiBot it reports 62.2% average task progress against 27.4% for the best pretrained VLA baseline. The code is Apache-2.0; the DROID checkpoint's weights are CC BY-NC 4.0 (non-commercial).

Dream-1 (Dream Labs, October 2026). A hosted API, founded by DreamZero co-authors. It takes a head camera and two wrist cameras and returns 1.6 seconds of actions per call. For robots it supports natively it outputs joint angles; for unseen robots it outputs end-effector poses that each robot's own inverse-kinematics solver turns into joint commands. No weights, architecture or training-data details are published. The company lists its own limitations: results depend on IK quality, it does not control the arm's path to a pose, and it is a research beta not validated for safety-critical use.

Questions to carry into the close readings

1. Does better video prediction mean better actions, or can a model imagine well and still act badly?
2. What does real-time control of a 14B video model cost, and how far is 7 Hz from what dexterous tasks need?
3. End-effector poses plus IK make a model robot-agnostic. What does that interface give up?
4. Which numbers are peer-reviewed, which are author-reported, and which are vendor claims? (Lesson 05-03 is about reading each kind.)

References

• Ye et al., World Action Models are Zero-shot Policies, arXiv:2602.15922, 2026. Project page · Code
• Dream Labs, Dream-1, 2026.
• Du et al., Learning Universal Policies via Text-Guided Video Generation (UniPi), arXiv:2302.00111, 2023.
• Brohan et al., RT-2, arXiv:2307.15818, 2023.
• Assran et al., V-JEPA 2, arXiv:2506.09985, 2025.