China's AgiBot Releases Embodied Foundation Model GO-3 Preview, Plans Open Source by Year-End

en.Wedoany.com Reported - On October 29, at the 2026 AgiBot Innovation Day held in Shanghai, AgiBot officially released its next-generation embodied foundation model GO-3 Preview, built on its self-developed embodied native architecture WRAM, with plans for full open source by the end of 2026. Yao Maoqing, co-founder and co-president of AgiBot, presented the model architecture and its synergy with the post-training system and embodied agent framework.

GO-3 Preview adopts the WRAM (World Reasoning Action Model) architecture, integrating environment understanding, future state prediction, and robot action generation into a unified modeling system, and is trained for cross-task generalization and fine manipulation capabilities. The model recently ranked first in three robot manipulation benchmarks: RoboDojo, RoboTwin, and RoboColiseum.

The model training system incorporates the Genie Evolver post-training system, which combines real robot data, simulation environments, and world models for subsequent training and optimization of robot manipulation policies. The accompanying Genie Master embodied agent framework was released on October 27, enabling robots to identify causes of task failures during actual operations, adjust manipulation strategies, and accumulate experience. AgiBot has already applied this framework to convenience store operation validation with the Genie G2 robot, involving continuous improvement of complex manipulation tasks.

GO-3 Preview and the two aforementioned systems together form a technical system covering pre-training, post-training, and post-deployment learning. AgiBot has scheduled the open-source plans for the model and framework for the end of 2026, with GO-3 Preview being the new model released at this Innovation Day, while Genie Master was previously released as an independent technical achievement.