Building a Foundation Stack for General-Purpose Robots

Robotics has never had its GPT moment. LLMs got a clean recipe—pretrain big, capability emerges—but robots have mostly been stitched together from separate perception, planning, and control modules that don't transfer well from one task to the next, let alone one machine to another. X Square Robot, a Chinese embodied-AI startup now valued above 20 billion yuan (roughly $2.9 billion), thinks it has found a workable version of that recipe, and it's releasing the pieces in the open.

The core insight isn't a bigger model. It's that robot data has been garbage for too long. X Square built a data-collection rig called QUANXTA Zero Series that captures human demonstrations with dual grippers instead of teleoperating an actual robot arm. That's not new by itself. What is new is the quality-control loop: the company replays a sample of recorded trajectories on a real robot and only counts the ones that physically complete the task. A gripper that snaps shut a fraction of a second too early still looks like a successful grasp on video, but it's actually shoved the object out of reach. Catching that distinction, X Square claims, gets them robot-grade performance at roughly one-twentieth the collection cost of an all-robot dataset—mostly because a wearable rig is dramatically cheaper than teleoperation gear.

On top of that data sits a world model called WALL-WM, which breaks from the industry habit of predicting motion in fixed time chunks. Instead it organizes behavior around events—reaching, grasping, placing—things you could describe in a sentence or watch unfold in video. A pretrained text-to-video model feeds a freshly trained action network without getting overwritten, so the system keeps whatever visual knowledge it already has while learning to act. It can run in a flexible event mode for long-horizon reasoning or snap into fixed-length chunks when a real controller needs steady, real-time output.

The action layer, Wall-OSS-0.5, sets an unusually blunt bar for itself: the pretrained model has to work on a real robot before any fine-tuning happens. It trains discrete action tokens, language grounding, and continuous motion generation all at once, with gradients flowing through the whole network rather than freezing parts of it. Paired with a tokenizer called X-Tokenizer that encodes the intent behind a motion rather than just its raw shape, the result is a system that reportedly stays stable even when noise gets added to an action—which is what lets the same tokenizer move across different robot bodies without retraining.

All of this is still measured mostly on X Square's own robots and its own benchmarks, which is the obvious asterisk. The company is releasing the world-model code publicly, so outside labs will finally get to poke holes in the claims or confirm them. Given how much of embodied AI research lives and dies by whether it generalizes beyond the lab that built it, that's the test that actually matters here.