X Square Robot opens its embodied AI stack: data, world model, action

X Square Robot, a Shenzhen-based Chinese company valued at over 20 billion yuan (about $2.9 billion), has described and released its embodied AI stack for general-purpose robots. Instead of one model, the company bets on three coupled layers: cheap data collection, the WALL-WM world model, and the Wall-OSS action model. The write-up describing the approach appeared in mid-July 2026.

Key takeaways

The stack has three layers: data, the WALL-WM world model, and the Wall-OSS-0.5 action model.
Data is collected without a robot, via a wearable VR rig with dual grippers (QUANXTA Zero) — roughly 20x cheaper than all-robot datasets.
Validation by replay on a real robot yields about 85 percent valid samples.
WALL-WM models events (reaching, grasping, placing) rather than fixed time windows.
Wall-OSS-0.5 is a vision-language-action model usable before any task-specific fine-tuning.

Three layers of one stack

Founded in 2023, the company builds its own WALL model family around a single thesis: the edge in general-purpose robotics will not come from scaling the model alone, but from a coherent stack of data, world model, and action model. The basic unit here is not a motion trajectory but an interaction — a successful change of world state.

Data, not model size, is the real bottleneck for general robots.

A thesis from X Square Robot's embodied AI stack description.

Robot-free data at 1/20 of the cost

The first layer is data collection. Instead of tedious robot teleoperation, X Square Robot uses a wearable VR rig with two grippers (the QUANXTA Zero system, based on the Universal Manipulation Interface idea). A person performs the tasks directly, and the recordings feed training without a robot in the loop.

Quality control is central. Collected data is replayed on a real robot, which filters out physically infeasible samples and yields roughly 85 percent valid demonstrations. The company says this collection method is about 20 times cheaper than building a dataset purely on robots. The strategy is to pretrain on robot-free demonstrations and then anchor with a small dataset from real hardware.

WALL-WM: a world model counted in events

The second layer is the WALL-WM world model. Its main difference from typical approaches is event-level modeling rather than fixed time windows. Instead of splitting a recording into equal time slices, the model describes coherent behaviors: reaching, grasping, placing. That lets it reason over a long action horizon.

Architecturally, WALL-WM is a text-to-video model coupled with a freshly initialized action network, which preserves the visual priorsvisual priors: knowledge about how the world looks, learned earlier from a large video corpus and preserved during further training. from video pretraining. It runs in two modes: an event mode, with variable-length segments for long-horizon reasoning, and a fixed-length mode for real-time robot control.

Wall-OSS and X-Tokenizer: action ready before tuning

The third layer is the Wall-OSS-0.5 action model, a vision-language-action system. Its premise is usefulness even before task-specific fine-tuning. The model is trained on three objectives at once: discrete action tokens, language grounding, and continuous action generation.

Tying it together is the X-Tokenizer, a semantic interface with hierarchical intent and detail codes. An intermediate layer describes action in terms of approach, contact, force, and recovery, which is meant to let the model transfer across different robots without re-tuning. It is an answer to the classic cross-embodiment problem: differences in control frequency, latency, compliance, and sensing.

Why it matters

Most teams working on VLA models focus on the scale and architecture of the model itself. X Square Robot shifts the center of gravity to data and how it is represented, arguing that this is where the real bottleneck lies. If the thesis holds, cheaper robot-free data collection could meaningfully lower the barrier to entry for the whole field.

Event-level modeling instead of time windows is also an argument in the ongoing debate about what a world model for a robot should even be. Finally, the emphasis on recovery from failure, not just raw success, matches the reality of homes and warehouses, which do not reset after a failed attempt. Releasing the code openly lets independent teams verify these claims, which matters because the company so far tests mainly on its own robots and benchmarks.

What's next?

The stack's code has been released for community testing and reproduction — the first test will be independent results outside the company's own benchmarks.
No third-party benchmarks yet — the key gap to close when assessing real generalization.
X Square Robot builds its own robots (including Quanta X2 and Quanta X1 Pro) — deployments on that hardware will show whether cross-platform transfer works in practice.