AGIBOT unveils GO-3 Preview, an embodied foundation model for learning across the full robot lifecycle

SUMMARY

AGIBOT has unveiled GO-3 Preview, an embodied foundation model that it says supports learning at scale through pre-training, post-training and post-deployment. The company reports first place on three embodied AI benchmarks, RoboDojo, RoboTwin and RoboColiseum, and says the model will be fully open-sourced by the end of 2026.

The model is built on WRAM, the World Reasoning Action Model, an embodied-native architecture that joins task understanding, future-state prediction and action generation. Three components carry this: CoAE, a control-oriented autoencoder for visual representations; SVP, which predicts a complete future state in one step; and KASO, which selects future predictions more consistent with the actions taken in training. Pre-training draws on human first-person videos and robot trajectories. AGIBOT says that in zero-shot settings the model can apply learned capabilities to new tasks without task-specific supervised fine-tuning on physical robots.

For post-training, AGIBOT built Genie Evolver, which combines a fleet of 100 physical robots with physics simulation and world models. Its Collective Experience Learning component aggregates execution experience across multiple robots in real-world environments. AGIBOT illustrates the approach with tabletop curling, a task in which a robot must control both the force and the angle of each shot. The company says that after just a few hours of practice, combining real-world interaction with world-model learning, the robot achieved a higher win rate than typical human players.

After deployment, the Genie Master agent framework uses the Embodied Execution Protocol, which standardises interfaces for capabilities, guidance, feedback and continuity. AGIBOT says its Genie G2 improved across multiple complex manipulation tasks in a real convenience store without human demonstrations or manual tuning.

Written with AI assistance from the source; every sentence was checked against it automatically before publication. Not reviewed by an editor. Follow the source link for the full text.