X Square Robot released XRZero-G0, an open-source framework that collects robot training data from human operators using VR headsets and wearable sensors rather than operating physical robots. The system synchronizes head-mounted and wrist-mounted cameras to match how robots perceive tasks, generating demonstrations that transfer across different robot platforms. The code and accompanying G0-Dataset are available on GitHub under an MIT license.

The framework addresses a persistent bottleneck in embodied AI development: gathering training data from physical robots is slow and expensive, typically producing only a handful of usable demonstrations per day. XRZero-G0 operators reached collection speeds of 93.2 episodes per hour — roughly 10x faster than traditional methods. The system runs quality control at three levels: multi-view geometric consistency, kinematic validation with collision detection, and final playback testing on physical robots.

The company reports that combining approximately 10 human-collected episodes with 1 real-robot episode matches the performance of datasets built entirely from physical robot operations. This 10-to-1 ratio reduces real-robot data requirements by up to 20x in tested scenarios. Human demonstrations provide behavioral coverage, while the small portion of real-robot data anchors physical factors like motor latency and friction that VR cannot replicate.

The accompanying G0-Dataset contains over 2,000 hours of validated demonstrations covering 3,000 distinct manipulation tasks across vision, tactile, and audio modalities. Policies trained on the dataset showed zero-shot transfer to robot platforms outside the training set, performing tasks without platform-specific fine-tuning. The framework enables labs and companies to build training datasets without needing large fleets of expensive physical robots during initial data collection phases.