Shenzhen based X Square Robot launched a $140M backed hardware kit on Monday to sell the synchronized, pre labeled training footage that AI controlled robots need to learn physical tasks.
X Square Robot wants to sell the training footage that AI-controlled robots need to learn. The Shenzhen-based startup launched its QUANXTA Zero Series on Monday, a three-product hardware and software kit that turns physical-world manipulation into a closed-loop data production line. It is the first concrete bid to become the default data-collection layer for embodied AI, the branch of artificial intelligence that trains robots to act in the real world. The launch lands at a moment when most labs still gather that data by hand.
The bet is that embodied AI's bottleneck is not model size or compute, but the volume, quality, and synchronization of physical training data. The field has spent the last two years scaling model parameters and GPU clusters while the data side of the pipeline has stayed artisanal.
QUANXTA Zero comes in three configurations. The flagship QUANXTA Zero-G1 pairs a single headband-mounted camera with dual grippers for hands-only teleoperation. QUANXTA Zero-G0 straps onto a PICO 4 VR headset with dual grippers and a backpack computer for whole-body mobile capture, and the corresponding XRZero-G0 codebase is public on GitHub. QUANXTA Zero-E0 carries a six-camera first-person array for environments where mobility matters more than dexterity.
All three feed into a single pipeline the company says handles data collection, sub-millisecond sensor synchronization (within 1ms across devices, with 100% frame-level video alignment), automated cleaning, intelligent annotation, model training, and evaluation in one loop. The integration is the pitch: traditional embodied-data stacks stitch together human teleoperators, separate annotation vendors, and offline training jobs, and the seams are where quality breaks down.
Independent validation is missing. The launch arrived as a company-issued press release distributed through PRNewswire, and the technical specifications are issuer figures, not third-party measurements. Cross-coverage from Robotics and Automation News and the Khaleej Times repeats the same talking points without adding outside assessment. The VIR writeup on the company's CVPR 2026 ManipArena benchmark paper follows the same pattern. The official product page confirms the product family but discloses no pricing, regional availability, or named pilots.
The capital base is real but modest. X Square Robot announced a $140 million Series A++ round before the QUANXTA launch, enough to manufacture the kits and stand up a support organization but not enough, on its own, to displace internal data collection at the largest U.S. labs. That structure explains why the company is publishing a GitHub codebase and pushing an academic benchmark at the same time as the product launch: it is recruiting researchers as customers and validators before it has enterprise buyers.
Embodied AI labs including Tesla, Figure, Physical Intelligence, and Google DeepMind's robotics group each collect their own training data through internal teleoperation farms. None has publicly committed to buying a kit from a Shenzhen startup. The closest independent artifact is the open-source XRZero-G0 repository, which lets researchers audit at least the data-collection layer. The field also has competitive open datasets such as Open X-Embodiment, RT-X, and DROID that labs can extend for free.
The plumbing underneath matters more than any single spec. Whoever standardizes the data-collection, synchronization, and labeling infrastructure for embodied AI captures upstream leverage on every downstream model, the same way TSMC's fabrication standards set the floor for every chip design. X Square Robot is making the earliest, most concrete claim on that layer with hardware it can ship today, an open codebase researchers can inspect, and a CVPR paper for academic credibility.
The kit is shipping, the open codebase is live, and the academic paper is published. What is not yet on the page is a third-party result trained on QUANXTA-collected data, and that is what would convert the bid into the field's default infrastructure.