Zhiyuan unveils embodied model go-3 preview with real-world task capabilities

On October 29, 2026, Zhiyuan released the GO-3 Preview, an embodied model capable of full-chain learning through pre-training, post-training, and deployment. The model uses the WRAM architecture to connect task reasoning, future prediction, and action generation, enabling robots to understand instructions, predict outcomes, and execute actions. It has topped three industry benchmarks: RoboDojo, RoboTwin, and RoboColiseum, and will be fully open-sourced by year-end. GO-3 Preview leverages WRAM to scale learning through data and model improvements. Human-first video data and robot trajectories provide rich interaction and task insights, while enhanced learning capabilities organize diverse experiences into transferable skills. The model excels in zero-shot and few-shot tasks, performing complex actions like making jianbing and assembling LEGO without additional fine-tuning. In real-world deployment, the model demonstrates continuous self-improvement. The Genie Master framework enables robots to adapt and refine strategies autonomously. A deployed robot in a convenience store improved its task success rate from 28.6% to 85.7% within five days, showcasing its ability to self-correct and learn from failure.