AGIBOT unveils GO-3 Preview, an embodied foundation model designed to learn across pre-training, post-training, and post-deployment, aiming to continuously improve through real-world experience.

A real-world demonstration in a convenience store using Genie Master shows autonomous improvement without human in-the-loop, illustrating Recursive Self-Improvement in the Physical World (RSI in the physical domain).

GO-3 Preview emphasizes Physical RSI, showcasing autonomous improvement from real-world deployment without constant human intervention.

Examples highlight manipulation and adaptation capabilities through LEGO-like assembly, tabletop curling, block extraction, and throwing tasks.

World models and physics simulations broaden validated interactions and enable comparison of alternative strategies in controlled settings.

GO-3 Preview ranks first on three embodied AI benchmarks—RoboDojo, RoboTwin, and RoboColiseum—and will be fully open-sourced by the end of 2026.

WRAM (World Reasoning Action Model) underpins GO-3 Preview, uniting task understanding, future-state prediction, and action generation to enable generalization across tasks.

Core WRAM components—CoAE (Control-Oriented Autoencoder), SVP (Single-step Future Prediction), and KASO (Kinetic Alignment of Successful Outcomes)—drive perception, forecasting, and action alignment for physically plausible outcomes.

In pre-training, GO-3 Preview uses WRAM with CoAE, SVP, and KASO to extract robotic representations, predict futures, and select plausible futures aligned with learned actions.

Pre-training uses heterogeneous data, including human first-person videos and robot trajectories, to develop transferable, generalizable skills.

Pre-training focuses on scalable data and embodied-native learning to build environment-agnostic skills.

GO-3 Preview enables zero-shot and few-shot task adaptation in real-world robotic tasks, demonstrated across diverse examples and capable of transferring general skills to unfamiliar tasks.