X Square Robot Opens Its Embodied AI Stack: Data, World Model, Action
X Square Robot is betting that general-purpose robots need a full stack, not a single model: high-quality interaction data, a world model that thinks in events, and an action model that can run before task-specific fine-tuning. The Shenzhen company has been publishing that stack in the open—and pairing it with wheeled bimanual platforms such as Quanta X2.
Three layers, one bet
In a July 2026 IEEE Spectrum deep dive authored with the company, X Square lays out the stack:
- Data — robot-free collection via wearable dual-gripper rigs (Universal Manipulation Interface / QUANXTA Zero Series), with physical playback so trajectories only count if they complete the task on a real robot. The company reports roughly a 20× lower collection cost versus all-robot teleop for comparable performance, and cites an about 85% data-validity rate under its inspection loop.
- World model (WALL-WM) — models action-grounded events (reach, grasp, place) rather than only fixed-time chunks; dual modes for long-horizon reasoning and real-time control.
- Action (Wall-OSS-0.5) — a vision-language-action model trained so the pretrained checkpoint should run on a real robot before task fine-tuning, with an X-Tokenizer that aims for semantic action codes reusable across embodiments.
Related open work includes XRZero-G0, a hardware/software framework for robot-free data collection, policy generation, and real-robot evaluation (MIT-licensed code on GitHub; G0-Dataset on Hugging Face). The Robot Report described a scaled 2,000-hour open dataset built with that pipeline, with paper arXiv:2604.13001.
Hardware context
On the company site, Quanta X2 is pitched as a next-generation wheeled humanoid with high-DoF arms and optional dexterous hands, aimed at home services, education, commercial cleaning, and logistics sorting. The software story is explicitly end-to-end: perception, decision, and high-precision manipulation under the WALL family of models.
IEEE Spectrum notes that much of the quantitative evidence still comes from X Square’s own robots and benchmarks. The company is inviting external reproduction as code and datasets land in public repos.
A Human’s Take
Open data with physical playback is the part I trust most. A trajectory that looks like a grasp on video but shoves the object is garbage; replaying on hardware and discarding failures is how you stop teaching the model lies.
I still want third-party numbers on cross-robot transfer and long-horizon home tasks. Until someone else runs Wall-OSS on a non-X Square arm and publishes the failure modes, treat the stack as a serious open research platform with strong claims—not a solved home robot.