X Square Robot Develops Open Stack for Universal Robots

Photo: IEEE Spectrum
Quick answer
X Square Robot has developed an open stack for universal robots, integrating data, the WALL-WM world model, and the Wall-OSS-0.5 action system.
Chinese startup X Square Robot has introduced an open technology stack for building universal robots, integrating data, a world model, and an action control system. Unlike traditional approaches, where perception, planning, and control systems operate in isolation, the new stack is built on three key principles: interaction with the physical world as the foundation for learning, the applicability of pre-trained models for real-world use without additional tuning, and behavior modeling based on physical events rather than fixed time intervals.
For data collection, the company developed the QUANXTA Zero Series system, which uses wearable devices with dual grippers instead of teleoperated robots. This reduces data collection costs by 20 times by leveraging more affordable equipment. Special attention is given to data quality: trajectories are validated on real robots, and only successful attempts are counted as valid. This approach enables the creation of cleaner and more efficient datasets, which is critical for training universal robots.
The WALL-WM world model, developed by X Square Robot, is built on events rather than fixed time intervals. This allows robots to better understand action sequences, such as grasping or moving objects, and adapt to new tasks. The model combines video prediction and action control while maintaining real-time performance. In experiments, WALL-WM demonstrated superior results on long-horizon task tests compared to traditional approaches.
The Wall-OSS-0.5 action system, in turn, enables robots to perform tasks without prior fine-tuning. It is trained on three objectives simultaneously: discrete action tokens, language grounding, and continuous motion generation. A key feature of the approach is the use of X-Tokenizer, which converts movements into semantically meaningful tokens understandable by the language model. This ensures stability and allows trained models to be transferred between different robots without additional tuning.
X Square Robot has already attracted investor attention, with the company's valuation exceeding $2.9 billion. The open-source world model and plans to expand testing on other platforms could accelerate the development of the universal robotics industry, making robots more accessible and adaptive.
Common questions
- What is X Square Robot's stack for robots?
- It is an integrated system comprising training data, the WALL-WM world model, and the Wall-OSS-0.5 action model. It allows robots to learn from interactions with the physical world and adapt to new tasks without extensive reconfiguration.
- What advantages does X Square Robot's approach offer?
- Key benefits include reduced data collection costs by using wearable devices instead of teleoperated robots, as well as improved training quality through trajectory validation on real devices. The WALL-WM world model also helps robots better understand action sequences.
- Why is X Square Robot making its stack open?
- The company believes openness will accelerate the development of the robotics industry by allowing the community to test, replicate, and improve the proposed solutions across different platforms and tasks.
Dzen feed: /feed/dzen.xml · RSS: /feed.xml