AgiBot GO-3 Preview Model Tops Robotics Benchmarks
The new embodied AI model improves autonomous task success by using a world reasoning architecture to predict physical outcomes.
Updated on Oct. 30, 2026 in Robotics
Live Poll
Do you trust that integrating AI robots into everyday services will improve efficiency and quality?
AgiBot has released the GO-3 Preview, an embodied AI model that uses a World Reasoning Action Model architecture to generate physical robot motions. The model currently holds the top position on the RoboDojo, RoboTwin, and RoboColiseum leaderboards.
Why it matters
This model addresses a core challenge in robotics by linking task reasoning, future prediction, and action generation to improve autonomous performance in complex, unstructured environments.
The Genie Master agent framework achieved a 17.7 percentage point improvement in single-task success compared to conventional supervised fine-tuning. This architecture leverages the Genie Evolver system, which synthesizes training data from real robots, physics simulations, and world models.
The players
AgiBot
A developer of embodied AI agents and hardware focused on autonomous navigation and task execution in physical environments.
The details
The GO-3 Preview model functions as a World Reasoning Action Model, a neural network architecture that predicts future outcomes before generating motor commands. When paired with the Genie Master agent framework, AgiBot G2 robots demonstrated a performance trajectory in convenience store tasks that saw success rates jump from 28.6 percent to 85.7 percent over a five-day training window.
Timeline
October 29, 2026: AgiBot launched the GO-3 Preview model.
End of 2026: AgiBot plans to release the model as open-source.
The Tech Race
AgiBot is competing directly with other labs to establish dominance on standardized embodied AI benchmarks like RoboDojo and RoboTwin. Securing top placements across multiple leaderboards indicates that the company is successfully scaling its world-model approach against conventional supervised learning.
The GO-3 Preview is currently in a preview phase, with AgiBot planning a broader release of the model's underlying source code by the end of 2026. The framework is designed to improve the reliability of robots operating in complex service environments like convenience stores.
The takeaway
AgiBot has demonstrated a measurable performance increase in task-based robotics through its new world-model architecture. Industry participants should monitor the end-of-year open-source release to see how this model performs when deployed beyond the specific convenience store test environment.
What happens next
AgiBot is scheduled to fully open-source the GO-3 model by the end of 2026.
Further reading
Explore more developments in autonomous machine learning on the Robotics page.
Source note: This article includes information reported by Pandaily.
Live Poll
Do you trust that integrating AI robots into everyday services will improve efficiency and quality?
