A Chinese robotics startup wants to build the 'recipe' that gives robots general intelligence
X Square Robot is open-sourcing an AI stack that aims to do for physical robots what large language models did for text. The approach is interesting. The independent evidence is still thin.
Key points
- X Square Robot, a Chinese embodied-AI company, is releasing its full robotics AI stack as open-source software in 2025.
- The company claims its data-collection method reaches performance comparable to all-robot datasets at roughly 20 times lower cost.
- Its world model, WALL-WM, organises robot behaviour around named actions like "reach" or "grasp" rather than fixed time windows.
- Its action model, Wall-OSS-0.5, is designed to work on a real robot before any task-specific fine-tuning, which is rare in the field.
- All headline results so far come from the company's own benchmarks, meaning independent testing has not yet confirmed them.
Teach an AI system enough text, and something useful emerges. That insight, called pretraining on broad data, turned large language models, the technology behind chatbots like ChatGPT, into general tools that millions of people use every day.
Robotics has no equivalent shortcut. A robot that folds laundry cannot simply transfer that skill to loading a dishwasher. The systems that run inside most robots today are built from separate, specialist parts that do not add up to anything general.
X Square Robot, a Chinese company focused on what researchers call embodied AI (meaning AI that lives in a physical body and must act on the real world), thinks it has found the recipe. And, unusually for a Chinese AI startup, it is releasing the core of it for anyone to use and inspect.
The company's bet rests on three layers working together: how robots collect learning data, a world model that predicts what will happen next in the physical world, and an action model that turns perception and reasoning into actual movement.
What makes this different from other robot AI projects?
The most distinctive piece is how X Square Robot handles data, the examples a robot learns from. Collecting robot demonstrations is expensive because you need a real robot, a trained operator, and hours of careful supervision. X Square Robot built a wearable rig with two grippers that a person wears on their hands. No robot required. That cuts collection costs sharply.
But cheaper data is worthless if it is wrong. The company adds a quality check that is genuinely unusual: a sample of recorded movements is physically replayed on a real robot, and only movements that complete the task are counted as valid. A gripper that closes a fraction of a second too early might look like a successful grasp in the recording, but the robot has actually pushed the object away. That trajectory is thrown out.
The company reports that combining this cheaper human-demonstration data with a small amount of real-robot data reaches performance comparable to datasets collected entirely on robots, at about one-twentieth of the cost. That is a striking claim. The caveat, as IEEE Spectrum AI noted in its coverage, is that these numbers come from X Square Robot's own robots and tests. Outside researchers have not yet independently confirmed them.
The world model, called WALL-WM, takes a different approach to time. Most robot AI systems break movement into fixed-length chunks, the way a film is cut into frames of equal length. WALL-WM instead organises behaviour around meaningful events: reach, grasp, place. Each event ends when the action is complete, not when a clock ticks over. The idea is that boundaries should match what is actually happening, not what is convenient to compute.
The action model, Wall-OSS-0.5, is a vision-language-action model (software that reads camera images, understands language instructions, and produces physical movement). The company's stated requirement for it is that the model must do something useful on a real robot before anyone trains it further for a specific task. Pre-trained general ability first, specialisation second. That ordering is the goal of the whole project.
For ordinary people, none of this changes anything today. General-purpose home robots remain years away. But if independent researchers can reproduce and extend these results, the open-source release could accelerate the field in ways that a closed, proprietary system cannot.
