Chinese company X Square Robot has shown an integrated software stack for general-purpose robots and open-sourced its key components. The system combines three layers: data collection, a world model and an action model, all built on physical events rather than fixed time slices. The code for the WALL-WM and Wall-OSS-0.5 models has been released to an open repository, and the architecture was detailed by IEEE Spectrum in 2026.
Key takeaways
- Three layers: data (the QUANXTA Zero system), the WALL-WM world model, the Wall-OSS-0.5 action model
- The basic unit of data is an interaction that actually changes the world, not merely a trajectory
- Data is collected with a wearable VR rig with dual grippers, not classic teleoperation
- 85 percent of collected trajectories count as valid, at roughly 20-fold lower collection cost
- The weights and code for WALL-WM and Wall-OSS-0.5 are released as open source
Three layers of one stack
The company, founded in late 2023, is now valued at over 20 billion yuan (about $2.9 billion), according to IEEE Spectrum. Its stack has three interdependent layers. The first is training data gathered by a system named QUANXTA Zero. The second is the WALL-WM world model?world model: A model that predicts future states of the environment from the current one, letting a robot 'imagine' the outcome of an action before performing it., which predicts what will happen in the environment. The third is the Wall-OSS-0.5 action model — a vision-language-action model that ties together image, text command and motion. The company sets a strict rule: after pretraining, the model must run on a real robot before any task-specific fine-tuning.
Events instead of time slices
The key difference is how the system understands a robot's behavior. Rather than splitting motion into equal time windows, WALL-WM works on events — coherent pieces of behavior such as reaching, grasping or placing an object. The model has two modes: an event mode for planning long sequences and a fixed-length mode for real-time control. The layers are tied together by X-Tokenizer, which turns motion into intent codes readable by a language model. Adding noise to an action barely moves the intent code, so the same tokenizer can be reused across different robots without retuning.
Data from a real robot
Instead of classic teleoperation, X Square Robot collects demonstrations with a wearable VR rig with dual grippers. Quality control is central: a sample of recorded trajectories is replayed on a real robot, and only those that actually complete the task count. The company reports 85 percent valid trajectories and performance comparable to all-robot datasets at roughly 20-fold lower collection cost. The model code and data-infrastructure documentation are public because, as the company stresses, embodied intelligence cannot be solved by a single organization.
Why it matters
The humanoid robot market is full of flashy demonstrations but lacks a shared software layer to build on. X Square Robot's proposal targets exactly that gap: instead of a single model, it offers a full chain from data to action. Opening the code is a bet that the winning standard will belong to whoever builds a community around it, not whoever locks the technology down. The focus on events and cheap data collection answers two real bottlenecks in robotics: the cost of data and transferring skills between different machines.
What's next
- The company names three questions for the community to test, including whether event-level representations generalize to new tasks and objects
- An open question remains whether the training transfers to robots built differently from those in the data
- X Square Robot wants real-robot evaluation to become a shared industry standard
Sources
- IEEE Spectrum — Building a Foundation Stack for General-Purpose Robots
- X Square Robot — About
- GitHub — X-Square-Robot/wall-x
