Chinese embodied-AI company X Square Robot presents a clear vision for building general-purpose robots capable of moving between tasks and platforms: an integrated stack that starts with interaction data, passes through a model of the physical world, and ends with an action model combining perception, planning, reasoning, and decision-making to produce executable behavior. The proposal comes at a time when robots still generally rely on separate components that do not necessarily produce a general capability transferable from one task to another or from one machine to another.
Unlike large language models, which benefited from a recipe based on pretraining on broad datasets, the robotics field has not yet reached agreement on a corresponding recipe for building general embodied intelligence. X Square Robot is betting that this recipe does not consist of one comprehensive model, but rather interconnected layers operating within a broader approach the company calls World Unified Model, which combines vision, language, action, and the prediction of physical changes.
Interaction Instead of Motion Trajectories
The company’s vision rests on three main principles. The first is that the basic unit of robot data should be interaction, not the motion trajectory alone; a demonstration is not considered successful merely because the joints moved, but when it changes the world in the intended way. The second is that pretraining should produce a usable capability, not merely a starting point requiring extensive fine-tuning. The third is to organize behavior around physical events rather than fixed-length temporal segments.
These principles make the stack’s layers interconnected. Robot-free data used to train the action model is also organized to feed the world model. However, the company explains that the world model and action model are two independent and complementary model families that share a software base within the broader World Unified Model architecture.
Lower-Cost, More Controlled Data
X Square Robot believes that the bottleneck facing general-purpose robots is related less to parameter count than to the cost and quality of interaction data. It therefore developed the QUANXTA Zero Series data-collection system, a robot-interaction interface in which people wear a platform equipped with two grippers instead of remotely controlling a robot while recording demonstrations.
The key difference is the inclusion of physical verification in the quality-control loop. The platform records trajectories, then a sample of them is replayed on a real robot, and only demonstrations that actually complete the task are counted. Closing the gripper a fraction of a second too early may appear in the data to be a successful grasp, but in reality it may push the object away. According to the company, the data-validity rate in its pipeline reached approximately 85%.
The company also combines a large quantity of recorded human demonstrations collected without a robot with a small amount of real-robot data to anchor the model to the dynamics of a specific machine. It says this approach achieves performance close to that of a dataset relying entirely on robot data, while reducing collection costs by approximately 20 times, primarily because the wearable platform costs less than teleoperation setups. However, the strongest reported results were measured on the company’s robots and data-collection pipelines, making broader independent testing necessary to verify generalizability.
An Event-Based World Model
The company’s world model is called WALL-WM and adopts action-related semantic events as its basic unit. This unit includes coherent behavior such as reaching for, grasping, or placing an object—actions that can be described in language, seen in video, and executed as motion.
This approach criticizes dividing behavior into fixed temporal windows, because a window boundary may fall in the middle of a single movement or combine two different movements. WALL-WM therefore operates in an event mode using variable-length segments suited to long-horizon task reasoning, while also providing a fixed-length mode for producing stable real-time outputs that the control unit can use.
To preserve the visual knowledge present in large video models, the design connects a text-to-video transformer to a newly adapted action network that reads video features without overwriting them. The company says its experiments included testing generalization to long-horizon tasks in settings unseen during training, and that the model outperformed baselines fine-tuned on relevant data within the company’s real-robot benchmark. However, according to the article, these results are still measured on an internal benchmark, while releasing the code allows the community to test and reproduce them.
A Deployable Action Model and Semantic Tokenization
Wall-OSS-0.5 represents the company’s vision-language-action model. X Square Robot sets a strict standard for itself: a pretrained model should operate on a real robot before any task-specific fine-tuning is performed.
The model trains three objectives together: discrete action tokens, language-action alignment, and continuous-action generation, while allowing gradients to pass through these components instead of freezing parts of the network. The company cites unprompted behavior that includes approaching, grasping, and recovery, including a task involving deformable objects that were not included in its training data.
The X-Tokenizer action interface redefines the conversion of continuous motion into tokens as learning a semantic interface. The higher-level token expresses the motion’s intent, while lower-level tokens carry fine-grained details, with these representations aligned to the language model’s features. According to the company’s description, adding noise to the action changes the intent token only slightly, helping reuse the tokenizer across different robots without retraining it.
What Still Needs to Be Proven?
The article argues that combining data-quality control through physical replay, event-based modeling, and a deployability-before-fine-tuning standard gives X Square Robot’s approach distinctive coherence. However, this does not yet settle the question of scalability beyond the company’s ecosystem. The company’s valuation rose above 20 billion yuan, or approximately 2.9 billion dollars, signaling investor confidence in the importance of data infrastructure, foundational models, and scalable training systems in embodied AI, but it is not in itself evidence of technical performance.
The company invites researchers to test three issues: how well event-based representations generalize across different tasks, scenes, objects, robots, and failure conditions; whether pretraining remains effective on robots the model did not see during training; and whether it is possible to build a shared evaluation for real robots that compares not only success rates but also identifies the causes of failure, such as misunderstanding instructions, perception breakdowns, or weak recovery.
According to X Square Robot’s vision, bringing robots into homes requires more than increasing the success rate on a single task. A reliable home robot must recognize its uncertainty, slow down when necessary, ask for help, and return the environment to a safe state after dropping an object or misunderstanding a request. It also needs cautious personalization that learns the household’s routines and preferences while making safety and trust a priority. In light of the current evidence’s heavy reliance on the company’s robots and benchmarks, this vision remains an important technical thesis awaiting broader validation from the research community.