WHAMS Explained: When Machines Learn to Simulate the Physical World — World Action Models Deep Dive

Forum topic · 小凯 · 2026-05-22

Summary

This post provides a deep-dive analysis of WHAMS (World Action Models), an embodied AI architecture presented by Google Research with Fudan University and the National University of Singapore (announced May 2026). Unlike traditional Vision-Language-Action (VLA) models that react only to current observations, WHAMS introduces a predictive World Action Model (WAM) paradigm: before acting, a robot mentally simulates future visual states and selects the best action. The core technique is a diffusion world model that jointly denoises future frames and action sequences, ensuring generated actions are spatiotemporally aligned with predicted physical outcomes. Training draws on four data sources: human teleoperation demonstrations, sub-millimeter tactile simulation, large-scale internet video pretraining, and causal physics understanding to bridge the sim-to-real gap. A Fast-WAM inference architecture skips costly rendering at deployment, reducing response latency to about 190 milliseconds. The article frames WHAMS as a shift from reactive to anticipatory robotics, where machines internalize an internal physics engine to predict consequences before acting.

Foreword: From "Feeling an Elephant Blindfolded" to "Foreseeing the Future"

The author opens with a classical-style argument: past robots were essentially "blind men touching an elephant" — they executed instructions step by step, but when the road held debris or the water rippled, they faltered because they did not know the *consequences* of their actions.

In the summer of 2026, Google Research and several academic partners introduced WHAMS (World Action Models). With it, robots no longer merely *react* — they learn to *foresee*. Before extending an arm, the robot has already rehearsed thousands of possible futures in its "mind" — a shift from passive contact to active anticipation.

> Note: WHAMS (World Action Models) > An embodied AI architecture that tightly couples physical world simulation with action generation. It gives robots an "internal physics engine" that guides current actions by predicting future visual states.

1. The Paradigm Leap: From VLA to WAM

Previous agents used the VLA (Vision-Language-Action) paradigm: see the object, understand the language, act. The flaw: they know the current action, but not what happens *after* it.

WHAMS opens the WAM (World Action Model) era. It asks not only "what should I do?" but "what will the world look like after I do it?"

| Model Paradigm | Core Logic | Intelligence Profile | | :--- | :--- | :--- | | VLA (Reactive) | $p(a \| o, l)$ | Reactive: handles each situation as it comes, ignores consequences. | | WAM (Predictive) | $p(o', a \| o, l)$ | Anticipatory: rehearses the future state $$o'$$ mentally, then acts optimally. |

> Note: VLA (Vision-Language-Action) > The traditional embodied AI paradigm, directly mapping multimodal inputs to action instructions without explicitly modeling physical consequences.

2. Under the Hood: Joint Denoising in a Diffusion World Model

How can a machine foresee the future? The secret lies in the Diffusion World Model.

WHAMS does not treat images and actions separately — it weaves them into a single "joint tensor." During its "thinking" process, like drawing silk from a cocoon, it simultaneously reconstructs a high-definition future scene and the optimal action trajectory from a blob of noise. This is Joint Denoising:

$T_{denoise}: \epsilon \to [o'_{t+1:t+k}, a_{t:t+k}]$

> Note: Joint Denoising > Simultaneously denoising image latent variables and action sequences in the diffusion model, ensuring generated actions and predicted future physical states are fully spatiotemporally aligned.

3. A Four-Layer Data Ecology

Where Co-Scientist excels at "thinking," WHAMS excels at "acting." To build its physical intuition, it draws from four sources:

1. Teleoperation foundation: precise demonstrations from human expert operators. 2. Simulation wings: millions of trials in sub-millimeter-level tactile simulation. 3. Open-world vision: watching millions of hours of short videos to absorb "gravity" and "occlusion" from everyday scenes. 4. Causal soul: learning not just the form but the physics logic behind actions.

> Note: Sim-to-Real > Transferring algorithms trained in virtual simulation to the real physical world. WHAMS narrows the "physical gap" through large-scale video pretraining.

4. Lightning Fast: The Fast-WAM Secret

Diffusion models of the past were accurate but slow. WHAMS introduces the Fast-WAM architecture: "exhaustively complex at training time, lightning-fast at deployment."

During training, high-definition video acts as a strict teacher of physical laws. At inference, the model skips expensive rendering and strikes straight at the physical essence, compressing response latency to 190 milliseconds — "still as a maiden, swift as a hare."

5. Conclusion: The Ultimate Symphony of Code and Matter

The birth of WHAMS means code is no longer cold logic — it has acquired "weight" and "touch," marking the point where AI genuinely begins to understand our physical world. When a machine can foresee the tilt angle of the coffee cup you hand it, or predict the trajectory of a falling tissue, the boundary between us and robots dissolves a little further. The symphony of code and matter has begun.

References and Paper Information

Core paper:

Title: World Action Models: The Next Frontier in Embodied AI
Release date: May 2026
Key contributors: Google Research, Fudan University, National University of Singapore (NUS)
Technical framework: WHAMS (WAMs)
Key techniques: Joint Diffusion, Spatiotemporal Priors, Fast-WAM Inference

Technical keywords:

Embodied AI
Diffusion World Model
Joint Denoising
Flow Matching constraints