World Model Breakthrough: 90 % Fake Data, 3× VLA Boom

90 % synthetic data, 300 % VLA leap: open-source GigaWorld-0 world model ends real-robot data scarcity for embodied AI

“RoboPub” Publication: 20% Discount Offer Link.

---

VLA Model Performance Surges 300%, with Training Data Now 90% Generated by World Model for the First Time

This is the latest breakthrough from a leading world model player, with both the model code and full training framework fully open-sourced.

The biggest long-standing bottleneck in bringing embodied intelligence to open-world deployment has never been the algorithm itself, but the extreme scarcity of high-quality, large-scale real robot interaction data.

Collecting real-robot data is extremely expensive and time-consuming, and it’s nearly impossible to cover diverse open-world scenarios, severely limiting the scalable training and generalization ability of Vision-Language-Action (VLA) large models. While traditional simulation can generate data quickly, it suffers from a significant Sim-to-Real gap, making it unable to support robust real-world deployment.