Architecture

WUM

2026ActivePublished: 23 July 2026Updated: 23 July 2026Published

Key innovation

Training perception, language, action and physical prediction within a single unified network, instead of separate vision, language and action modules.

How it works

Instead of combining separately trained modules, the World Unified Model trains a single network that jointly represents: perception (processing sensory data), language (instruction understanding and reasoning), action (robot motion control) and physical prediction (anticipating the effects of actions in the world). Through joint training, the model builds a coherent representation in which reasoning and acting are grounded in the same space. The architecture is realized in the WALL models (e.g. WALL-B) at a billion-parameter scale, supporting generalization across tasks and scenarios.

Problem solved

Classical robotic systems combine separate vision, language and control modules, which makes coherent information flow between them and generalization difficult. WUM integrates these functions into a single network, enabling natively multimodal, end-to-end control and incorporating physical prediction (the consequences of actions) within one model.

Components

PerceptionSensory input

Processing sensory data (vision and other modalities) within the shared network.

Language and reasoningUnderstanding and decisions

Understanding natural-language instructions and reasoning (including spatial and causal).

Action controlControl output

Generating precise robot control actions, including force control and positioning.

Physical predictionWorld modeling

Predicting the effects of actions in the world (a world-modeling element) trained in the same network.

Unified end-to-end networkIntegration

A single jointly-trained network replacing separate vision/language/action modules.

Implementation

Reference implementations

X Square Robot (GitHub)

Python · X Square Robot

Official

X Square Robot (Hugging Face)

Python · X Square Robot

Official

Implementation pitfalls

Harder interpretability and debuggingMedium

A unified end-to-end network is harder to analyze and fix than a modular architecture with separated components.

Fix:Interpretability methods, explicit reasoning stages (e.g. embodied chain-of-thought).

High data and compute costHigh

Joint multimodal training requires large, diverse embodied data and significant compute.

Fix:Efficient action tokenizers, pretraining, data reuse.

Results mainly vendor-reportedMedium

Many results and claims come from the vendor and require independent verification.

Fix:Await independent benchmarks and replications.

Evolution

2025

WALL-OSS

X Square Robot’s open-source vision-language-action model (September 2025) — an earlier stage of the WALL series.

2026

WALL-B and the World Unified Model architecture

Inflection point

The WALL-B model (April 2026, Beijing) introduces the World Unified Model architecture, training perception, language, action and physical prediction in one network.

Hyperparameters (configurable axes)

Parameter scaleHigh

Model size; WALL models operate at a billion-parameter scale.

miliardy parametrów / billion-level

Unified modalitiesHigh

Which functions are trained jointly in one network.

percepcja + język + akcja + predykcja fizyczna

Compute bottleneck

Joint end-to-end multimodal training

Jointly training perception, language, action and physical prediction in one billion-parameter network is compute-intensive and requires large, diverse embodied data.