AgiBot Launches GO-3 Preview Embodied Model, Tops Three Robotics Leaderboards

On Oct 29, AgiBot (智元) launched GO-3 Preview, an embodied model described as the culmination of its AI WAVE work. The system targets full-chain, large-scale learning across pre-training, post-training, and post-deployment, rather than treating training as a one-off stage.

The model rests on an embodied-native architecture that unifies understanding, prediction, and action. With Genie Evolver and Genie Master, GO-3 Preview extends learning into post-training and deployment, allowing robots to refine behavior through experience interaction and real-world operations.

AgiBot says GO-3 Preview has taken first place on three benchmarks: RoboDojo, RoboTwin, and RoboColiseum. The triple result suggests that scalable embodied learning can translate into strong benchmark performance, while the post-deployment learning loop points to more adaptive robots over time.

The company positions the release as part of a broader push toward embodied AI that learns continuously from deployment data. For the field, the leaderboard sweep provides an early signal that unified architectures and post-training loops may matter as much as pre-training scale in robotics.