Chinese robot startup X Square bets on unified stack for general-purpose machines

X Square Robot, a Chinese embodied-AI company, has made an unusually explicit argument about how to build general-purpose robots. The central insight is that robotics lacks the equivalent of the large-language-model recipe: the field has assembled systems from separate perception, planning, and control modules that rarely add up to intelligence a robot can carry from one task to another or one machine to another. X Square's answer is an integrated stack spanning data collection, a world model called WALL-WM, and an action model called Wall-OSS-0.5, held together by three core principles: the basic unit of robot data is an interaction that succeeds only if it changes the world as intended; pretraining should yield usable capability, not just initialization; and behavior should be modeled around physical events rather than fixed time slices.

The company's approach to data collection reflects a deliberate choice to prioritize quality over scale. Rather than teleoperating robots—a method that forces operators to work within each machine's kinematics, latency, and viewpoint, yielding slower and less diverse demonstrations—X Square built its Universal Manipulation Interface (UMI) data collection system, QUANXTA Zero Series, around capturing human skill. Operators wear a wearable rig with dual grippers, and demonstrations are recorded independently of any robot. The distinctive step is physical validation: a sample of trajectories is replayed on a real robot, and only those that actually complete the task count as valid data. This quality-control loop is uncommon; a gripper that closes a fraction of a second too early still looks like a grasp in the recording but has pushed the object away and should be rejected. The company reports that pretraining on a large volume of robot-free human data, then adding a small anchor of real-robot data, reaches performance comparable to an all-robot dataset at roughly a 20-fold lower cost of collection, primarily because the wearable rig is far cheaper than a teleoperation setup.

X Square's world model, WALL-WM, departs from mainstream approaches. Most action models predict fixed-length chunks of motion from the current image and instruction, which is computationally convenient but forces behavior into time-based windows rather than semantic boundaries. WALL-WM instead treats an action-grounded semantic event—reaching, grasping, or placing—as its unit: a coherent piece of behavior that can be named in language, seen in video, and executed as motion. The architecture couples a text-to-video model to a freshly initialized action network that reads from video features without overwriting them, preserving the visual prior learned by large video models. From this single process, WALL-WM offers two modes: an event mode running in variable-length segments for reasoning over long horizons, and a fixed-length mode for steady real-time control. In company-conducted experiments, models trained on limited datasets outscored baselines fine-tuned on related data when evaluated on long-horizon tasks in unseen settings on X Square's real-robot benchmark—a meaningful result if it holds, though it is currently measured only on the company's own hardware.

The action layer introduces two connected ideas. First, Wall-OSS-0.5, X Square's vision-language-action model, is designed to run on a real robot before any task-specific fine-tuning, a departure from designs that freeze parts of the network. The model trains three objectives together—discrete action tokens, language grounding, and continuous action generation—while keeping gradients flowing through all of them. The second idea is X-Tokenizer, which reframes tokenization as learning a semantic interface: the top-level code stands for the intent of a motion, while lower-level codes carry finer detail, all aligned with the language model's features. A practical consequence is stability: adding noise to an action barely moves the intent code, allowing one tokenizer to be reused across robots without re-tuning. Together, these design choices give the action layer the ability to transfer capability across machines.

X Square's valuation has climbed above 20 billion yuan (about US $2.9 billion（約4600億円）), reflecting investor confidence that data infrastructure, foundation models, and scalable training systems will be long-term differentiators in embodied AI. However, the company acknowledges that most of the current evidence comes from its own robots and benchmarks. With the world model code now being released to the public, the robotics community will have the opportunity to test, reproduce, and build on the work across more robots, tasks, and settings—a crucial next phase for validating whether the stack's principles generalize beyond X Square's controlled environment.

The robotics field has long struggled with a fundamental problem that large language models solved: how to build systems that generalize. Where LLMs benefit from pretraining on broad data and then fine-tuning for specific tasks, robotics has historically cobbled together separate perception, planning, and control modules that rarely transfer knowledge across different robots or tasks. X Square Robot's explicit wager is that the solution lies in an integrated stack where data infrastructure, world modeling, and action generation are tightly coupled from the ground up.

The company's diagnosis—that data quality, not model size, is the real bottleneck—reflects a pragmatic shift in how the embodied-AI field is thinking about scaling. By using inexpensive wearable rigs to capture human demonstrations, then validating them through physical playback on real hardware, X Square claims to have cracked a cost problem that has long made robot learning expensive. The reported 20-fold reduction in data collection cost, if confirmed independently, would reshape the economics of robot training. Equally important is the architectural choice to organize the world model around semantic events—coherent behaviors like grasping or placing—rather than fixed time windows. This design respects what large video models already know while still yielding executable motion, a middle ground between pure prediction and pure control.

The fact that X Square's valuation has exceeded US $2.9 billion（約4600億円）, and that the company is releasing code openly, signals confidence among investors that data infrastructure and foundation models will be long-term differentiators in embodied AI. However, much of the current evidence comes from the company's own robots and benchmarks. The next phase—independent testing and reproduction across a wider range of robots, tasks, and settings—will determine whether these principles truly generalize or remain specific to X Square's carefully controlled environment.

No comments yet. Be the first to share your thoughts!