Large language models handed AI a clean recipe: pretrain a big model on broad data, and general capability follows. Robotics never got that recipe. For years, robots have been stitched together from separate perception, planning, and control modules that rarely add up to intelligence a machine can carry from one task to the next. X Square Robot, a Chinese embodied-AI company, thinks it has found the missing formula — and it wants to release it in the open.
The company’s bet is an integrated stack built on three layers: the data a robot learns from, a world model that predicts how the physical world changes, and an action model that fuses perception, planning, reasoning, and decision-making into executable behavior. Both the world model and action model are described as complementary but independent families sharing a code base, sitting inside a broader World Unified Model that trains vision, language, action, and physical prediction together.
The through-line is a handful of principles. The basic unit of robot data is an interaction, not a trajectory — a demonstration counts only if it changes the world as intended, not just because joints moved. Pretraining should yield usable capability, not merely an initialization. And behavior should be modeled around physical events rather than fixed time slices.
That thinking shows up most clearly in data collection. Instead of teleoperating a robot, X Square’s QUANXTA Zero Series — built on its Universal Manipulation Interface (UMI) system — captures demonstrations from people wearing a dual-gripper rig. The clever part is quality control:
- A sample of recorded trajectories is physically replayed on a real robot, and only those that actually complete the task count as valid — turning validity into a measured number rather than an assumption.
- Cheap human data is used for pretraining, then anchored with a small amount of real-robot data. The company reports this matches an all-robot dataset at roughly a 20-fold lower collection cost.
The world model, WALL-WM, treats an action-grounded semantic event — reaching, grasping, placing — as its unit, rather than chopping motion into fixed-duration windows. It couples a text-to-video model to a freshly initialized action network that reads video features without overwriting them, preserving the visual prior. It runs in two modes: variable-length events for long-horizon reasoning, and fixed-length output for real-time control.
The action layer pairs Wall-OSS-0.5, a vision-language-action model that must run on a real robot before any task-specific fine-tuning, with X-Tokenizer. The latter reframes tokenization as a semantic interface, so the top-level code carries a motion’s intent while lower codes handle detail — stable enough to reuse one tokenizer across robots without re-tuning.
Investors seem convinced: X Square Robot’s valuation has climbed above 20 billion yuan (about US $2.9 billion). The bigger test comes next. Most current evidence rests on the company’s own robots and benchmarks — but with the world model code now public, the wider community can finally try to reproduce it.
