2026-07-25
Written by Lena Kaplan
Researchers have established a common foundation stack for general-purpose robots, including perception, manipulation, and locomotion capabilities. This framework enables the development of robust and adaptable robots capable of navigating complex environments and performing a wide range of tasks.
In recent years, artificial intelligence (AI) has made significant strides in various fields, with large language models playing a pivotal role in this progress. These models have enabled AI systems to learn from broad data and generalize capabilities across different tasks and domains. However, robotics, a field that has long been hindered by the lack of a unified approach, is now beginning to catch up.
X Square Robot, a Chinese embodied-AI company, has taken an ambitious step in this direction. The company's foundation stack, which combines data, world modeling, and action modeling, aims to provide robots with general-purpose intelligence that can be applied across various tasks and environments. In this article, we will delve into the details of X Square Robot's approach and explore its potential for revolutionizing robotics.
X Square Robot's foundation stack is built around three interconnected layers: data, world modeling, and action modeling. The data layer consists of high-quality interaction data, which is collected using a custom-built Universal Manipulation Interface (UMI) system. This system uses wearable VR rigs and dual grippers to capture demonstrations from humans, allowing for more diverse and realistic data.
The world model, called WALL-WM, is designed to predict changes in the physical world. It treats an action-grounded semantic event as its unit, rather than fixed slices of time. This approach allows the world model to understand the ways actions change objects, contacts, and task states. The wall-wm design reflects a specific concern about not discarding what large video models already know.
The action model, called Wall-OSS-0.5, is responsible for generating executable robot behavior. It trains three objectives together: discrete action tokens, language grounding, and continuous action generation. The model runs on a real robot before any task-specific fine-tuning, allowing it to produce capability that can be applied across different tasks.
The data layer provides the foundation for the world model and action model. The interaction data is structured in a way that feeds both models, making them interdependent. This approach allows the world model to understand the changes in the physical world and the action model to generate executable behavior.
The principle of treating interactions as units, rather than trajectories, also plays a crucial role in this framework. Demonstrations are successful only if they change the world as intended, not simply because the joints moved. This emphasis on interaction enables the system to learn from diverse data and generalize capabilities across different tasks.
While X Square Robot's approach holds great promise, there are still challenges to be addressed. The most significant one is the cost and quality of interaction data, which remains a bottleneck for general-purpose robots.
The company's Universal Manipulation Interface (UMI) system addresses this issue by collecting demonstrations from humans wearing wearable VR rigs and custom grippers. This approach breaks the expensive scaling law of teleoperation, allowing people to generate rich data independently of any robot.
Data quality control is also crucial, as errors in robot data can be far more expensive than in language data. X Square Robot's pipeline has a remarkable 85 percent data-validity rate, which ensures that only successful demonstrations are counted.
X Square Robot's foundation stack has the potential to revolutionize robotics by providing robots with general-purpose intelligence that can be applied across various tasks and environments. The company's approach combines three interconnected layers: data, world modeling, and action modeling.
While there is still much work to be done, X Square Robot's efforts have already garnered significant attention from investors, with the company's valuation climbing above 20 billion yuan (approximately US $2.9 billion). The release of its open-source components will enable researchers to test, reproduce, and build on the work, driving progress in embodied AI.
X Square Robot's foundation stack represents a significant breakthrough in robotics, providing robots with general-purpose intelligence that can be applied across various tasks and environments. By combining data, world modeling, and action modeling, the company has created a unified approach that addresses some of the most pressing challenges in robotics.
As researchers begin to test, reproduce, and build on X Square Robot's work, we can expect significant progress in embodied AI. The company's commitment to open-source components will enable the broader community to contribute to this effort, driving innovation and advancing our understanding of robot intelligence.
