When Code Learns to Dream: World Action Models (WAMs) and the Genesis of Embodied AI
Preface: The Robot's "Brain in a Vat" and Imagination
If you've ever watched a Boston Dynamics robot dog (Spot) navigate rough terrain, you may have marveled at its balance. But ask it to brew a perfect cup of English tea, and it would likely fall into a logical dead end. The reason is simple: it can "see" the cup and "hear" the command, but it doesn't understand what "brewing tea" physically means.
For a long time, we trained "reactive" robots—puppets on strings that capture external signals through complex sensors and produce a table-lookup response. This paradigm is known as VLA (Vision-Language-Action).
However, in May 2026, with the landmark survey *World Action Models: The Next Frontier in Embodied AI* (arXiv:2605.12090) from Fudan University and the Shanghai AI Laboratory, a realization crystallized: if an AI cannot mentally "rehearse" the future, it will never truly take over reality.
This is the birth of World Action Models (WAMs). It's not merely a technical iteration in embodied AI—it's a kind of "awakening" for silicon-based intelligence: they have finally learned to "dream."
---
1. Paradigm Shift: From "Blind Execution" to "Intervention Modeling"
A traditional AI is like a diligent apprentice lacking common sense. Tell it "walk forward" and it faithfully steps ahead. But if a cliff lies before it, it has no concept of "falling" until it drops.
WAMs fundamentally change this underlying logic. They no longer care only about "what should I do"—they care about "if I do this, what will the world become?"
#### Core Formula: A Joint Distribution with a God's-Eye View In the WAM universe, everything is distilled into one elegant expression:
> Note: Joint Probability Distribution > This formula means the model predicts two things simultaneously: the future observation \(o'\) and the corresponding action \(a\). It fuses "physical evolution" and "decision intent" together. In this view, an action is no longer an isolated command but a perturbation that reshapes the spacetime trajectory.
Compared with the old paradigm \(P(a|o, l)\), WAMs add an \(o'\). That tiny prime symbol represents the "vision" of one second later inside the AI's mind.
---
2. The Spine of the Gods: Three Pillars of WAMs
Building a system that can move, see, and "imagine" requires three modules—independent yet deeply coupled.
#### Representation: The Foundation of a Digital Twin How does an AI understand the test tube, robotic arm, or petri dish before it? It can't rely on raw pixels. WAMs tokenize video streams, converting chaotic color signals into semantic units machines can understand.
> Note: Tokenization > Imagine dismantling a fine oil painting into tens of thousands of Lego bricks. Each brick (token) represents a specific semantic or physical feature. This way, the AI isn't processing images—it's processing a language of physics.
#### Prediction: Alchemy on the Timeline This is the most fascinating part. Based on current tokens, the AI runs a miniature physics simulation internally: if the robotic arm drifts 5 cm left, how will the liquid slosh? Will the centrifuge's speed induce vibration?
#### Action: From Illusion Back to Reality Once the AI confirms the optimal future scenario internally, it converts that "dream" into real electrical signals—spinning motors, opening pneumatic valves.
---
3. The Architecture Debate: Cascaded vs. Joint WAMs
AI researchers are split into two camps, like "prefabricated parts" versus "monolithic pouring" in architecture.
#### Cascaded WAMs — The Modular Aesthetic Cascaded architectures work through decoupling. A representative work is UniPi.
- Step one: A video generation model (the world model) acts like a director, first sketching the "storyboard" of the next 5 seconds.
- Step two: A control model, like an actor, watches the storyboard and infers how to move.
- Strength: Extremely flexible—you can train the "director" on hundreds of millions of hours of internet cooking videos, giving it broad common sense.
- Pain point: Director and actor can lack coordination, causing prediction–execution inconsistency.
- Core representative: DreamZero.
- Strength: Actions and physical evolution are perfectly synchronized—no "brain says left, hand goes right."
- Pain point: Extremely hard to train; signals of different dimensions (pixels vs. vectors) often clash in the loss function.
#### Joint WAMs — End-to-End Brute Force A joint architecture is more like an all-rounder: within a single Transformer, it predicts the next video frame while emitting action commands.
4. The Data Flood: The Awakening of the First-Person View
Training WAMs requires massive data. But where do you find so many videos of robotic arm manipulation?
The survey points out that the 2026 breakthrough lies in large-scale egocentric video.
> Note: Egocentric Video > Think of GoPro-style footage. Though it shows a human doing experiments or cooking, the AI can quietly absorb humanity's tens of thousands of years of accumulated physical intuition by watching how hands move and objects respond.
This is why systems like Genie can learn gravity, collision, and friction purely from watching video, without human annotations.
---
5. The Dark Forest: WAMs' Last Frontiers
Despite the dawn WAMs offer toward AGI, three mountains still stand:
#### 1. The Ghost of Causality AIs often confuse "because" and "therefore." If every hand movement in a video coincides with a light turning on, it may conclude the hand is luminous. Teaching AI true causal logic rather than mere statistical correlation remains the thorniest problem.
#### 2. The "Clipping" Problem of Physical Common Sense Despite their erudition, large models often produce "physical hallucinations"—a bottle passing through a table, liquid flowing upward. In a high-precision lab, a 0.1 mm hallucination can destroy millions of dollars of equipment.
#### 3. The Life-or-Death Race of Real-Time Performance World models are heavy. Generating a video clip takes tens of seconds, but robot control loops demand responses within 10 milliseconds (100 Hz). It's as if you must finish a dream *and* adjust your breathing based on it in the blink of an eye.
---
Conclusion: When AI Becomes the "Chief Soul Architect"
The survey's closing words are stirring: "The significance of WAMs lies in transforming robots from prisoners of physical law into co-creators of the physical world."
The top scientist of the future will no longer be the one shaking flasks by hand. We will have thousands of WAM-equipped "silicon assistants"—operating in reality, rehearsing in illusion, continuously refining their understanding of truth through \(P(o', a | o, l)\).
When code learns to dream, reality is no longer a cold constraint but a laboratory of infinite extension.
---
References
1. arXiv:2605.12090: *World Action Models: The Next Frontier in Embodied AI* (2026). 2. UniPi Paper: *Learning to Act via Video Generation and Inverse Dynamics* (2025/26 Updated). 3. DreamZero: *Zero-shot Policy Generation via World Model Pre-training* (2026). 4. Genie: *Generative Interactive Environments from Unlabelled Video* (2024/25 Heritage). 5. VLA Survey: *From Reaction to Prediction: The Evolution of Embodied Foundations*.