AgiBot releases GO-3 Preview embodied AI model
The AMW Read
GO-3 Preview extends a known robotics player's model strategy into deployment feedback, with potential segment-level impact if its reported gains generalize.
AgiBot releases GO-3 Preview embodied AI model
AgiBot released GO-3 Preview on October 29, describing it as an embodied AI model that connects task reasoning, prediction, and action generation. The company says it ranked first on the RoboDojo, RoboTwin, and RoboColiseum benchmarks. It plans to open-source the model fully by the end of 2026. The reported rankings and release plan have not been independently verified here.
The release extends AgiBot's work on general-purpose robot models from initial training toward learning after deployment. Its Genie Evolver system combines experience from roughly 100 physical robots with simulation and a world model. Its Genie Master agent framework records successes and failures to guide later tasks and produces data for further action-policy training. This approach puts AgiBot among companies seeking to build a reusable model layer for physical AI, rather than a separate controller for every task.
For builders and investors, the key test is whether these learning gains hold across more tasks and deployments. AgiBot reports that a robot's complete-task success rate in a convenience store rose from 28.6% on day one to 85.7% after five days. The source does not give enough detail to assess the result independently or establish how well it transfers to other settings. AgiBot's earlier GE-Act 2.0 release focused on scaling training hours; GO-3 Preview adds a stated approach to learning from deployment experience.
