GO-3 Preview
AgiBot launched GO-3 Preview, an embodied model built on a World Reasoning Action Model architecture, which it says ranks first on RoboDojo, RoboTwin and RoboColiseum.
GO-3 Preview links task reasoning, prediction of what happens next and action generation in one WRAM architecture. It is trained on first-person human videos and robot trajectories, then post-trained with Genie Evolver, a system using about a hundred real robots plus simulation and world models. After deployment the Genie Master agent framework lets robots diagnose failures and keep improving.
- Date
- Thursday 29 October 2026
- Lab
- AgiBot
- Kind
- model
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Convenience-store full task flow success rate, G2 robot | 28.6% on day 1 to 85.7% after 5 days without human demonstrations or tuning | company |
| Single-task success gain over conventional supervised fine-tuning (hanging small hooks) | +17.7 percentage points | company |
AgiBot plans to open-source the full model by the end of 2026; leaderboard first places are company claims.
Sources
This record was checked against its sources on 30 October 2026. How we check