Agibot Releases GO-3 Preview Embodied AI Model
Agibot has released GO-3 Preview, a new embodied AI model built on the WRAM (World Reasoning Action Model) architecture that tops three major industry benchmarks. The model recently demonstrated advanced multi-scenario capabilities, such as making pancakes and operating in convenience stores, and is scheduled for full open-source release by the end of 2026. This release marks a significant milestone in scaling robotic intelligence by seamlessly connecting task reasoning, world prediction, and action generation across a full lifecycle. Achieving autonomous post-deployment improvement allows robots to evolve without human intervention, bridging the gap toward practical commercial deployment in complex real-world environments. GO-3 Preview utilizes a hybrid post-training system called Genie Evolver combined with the Genie Master agent framework to facilitate zero-shot, few-shot, and continuous online learning. During testing in a real convenience store, the robot improved its overall task success rate from 28.6% to 85.7% over five days without human intervention.
## BACKGROUND
Embodied AI integrates machine learning and computer vision to enable physical agents like robots to perceive, reason, and act within their physical surroundings. Advanced frameworks often utilize Vision-Language-Action (VLA) models and world models to help robots understand physical dynamics rather than merely mapping observations to reactive motions.