X Square opens a shortcut for robot training

X Square Robot has announced the open-source release of XRZero-G0, a hardware and software framework designed to generate robot-learning data without relying constantly on real robots. The company is also releasing G0-Dataset, a validated multimodal dataset of more than 2,000 hours for embodied AI and robotics research.

The notable point is not just that code and data are being opened. XRZero-G0 targets a practical bottleneck in robot manipulation: training a robot usually requires demonstrations collected through teleoperation, which means expensive machine time, skilled operators, and a limited number of trials per day. Robot-free collection shifts part of that work to humans wearing sensors, then converts their demonstrations into usable examples for robot policies, the models that choose the robot’s actions.

X Square Robot says its framework addresses a common weakness in those methods: the gap between what a human sees during data collection and what a robot will perceive when deployed. XRZero-G0 combines a head-mounted camera, two wrist cameras, a VR interface, and interchangeable grippers. The observations are aligned with the robot’s perception space. The pipeline also includes automated inspection of trajectories, with collision and joint-limit constraints, followed by validation through real-robot replay.

The company gives two figures that should be read as experimental results rather than industrial guarantees: about 85% effective data yield under controlled conditions, and up to a 20x reduction in the amount of data that must be collected on real robots. The core mechanism is a mix of roughly ten robot-free episodes for each real-robot episode, which the company says achieved performance comparable to training only on real-robot data in the evaluated tasks.

For robotics teams, the implication is straightforward. If the results hold up outside controlled settings, researchers could iterate faster on manipulation tasks and transfer behaviors more easily across different robot bodies. This is not a claim that a general-purpose robot is suddenly ready. It is a more modest but important infrastructure move: reducing the cost of physical data, one of the slowest inputs in embodied AI.