Zhiyuan has released the GO-3 Preview embodied intelligence model, which achieves full-chain scalable learning across pre-training, post-training, and post-deployment stages. The model ranked first in all three evaluations: RoboDojo, RoboTwin, and RoboColiseum. During pre-training, WRAM is employed to absorb approximately 90% of human first-person perspective data. In the post-training phase, Genie Evolver is used for mixed training involving real robots, simulation, and world models. After deployment, Genie Master is applied to accumulate success and failure experience in real-world scenarios such as convenience stores.