Zhiyuan unveils GO-3 Preview to scale robot learning across pre-training, post-training and deployment
Zhiyuan on Oct. 29 introduced GO-3 Preview, an embodied AI model the company says brings scalable learning to the full robot lifecycle, spanning pre-training, post-training and deployment. The release was paired with claims that the model ranked first in three benchmark evaluations: RoboDojo, RoboTwin and RoboColiseum. The system is built around WRAM, short for World Reasoning Action Model, which links task understanding, future-state prediction and action generation. In pre-training, Zhiyuan said roughly 90% of GO-3 Preview’s data came from human first-person video, with the rest used alongside robot trajectories to connect observed state changes to executable control. The company also introduced KASO, a method designed to select among multiple plausible future predictions and align them with real actions during training. For post-training, Zhiyuan described Genie Evolver, a distributed hybrid setup that combines real robots, physics simulation and world models. In one experiment, the company said scaling the number of real robots improved phone-packaging success rates from 50.8% to 98.3%, while failures in a 120-run long-horizon test fell from 25 to 2. Zhiyuan also said GO-3 Preview achieved a win rate above ordinary human players in tabletop curling after only a few hours of practice in real and model-based environments. The deployment layer is handled by Genie Master, an embodied agent framework aimed at turning real store operations into new training data. Zhiyuan said its Genie G2 robot has already been tested in a real convenience store, where it kept improving on prepared-food packing, item retrieval and restocking without relying on manual demonstrations or tuning.








