Zhiyuan Robotics officially released its new generation of embodied base large model, Genie Operator-2 (GO-2), marking a major breakthrough in the field of embodied intelligence. The core mission of this model is to bridge the semantic gap between "understanding instructions" and "precise execution," enabling robots not only to understand commands but also to perform tasks effectively.
First-of-its-kind Action Thought Chain Technology
GO-2 introduces a "action thought chain" mechanism, changing the traditional "see and act" blind mode of robots. After receiving an instruction, the model first generates a structured action sequence plan internally, similar to how humans visualize a basketball trajectory before shooting, ensuring a clear and logically coherent action path.
Complementing this mechanism, GO-2 employs an innovative asynchronous dual-system architecture. The slow system handles high-level planning, while the fast system continuously adjusts action details at a higher frequency in real time. This design allows the robot to complete tasks stably even when faced with disturbances such as desk height deviations or object sliding.
Breaking Multiple Industry Benchmark Records
In terms of performance, GO-2 achieved first place in several global mainstream embodied intelligence benchmark tests, such as LIBERO, with an average success rate of 98.5%. Even in complex environments with lighting interference and background noise, the model still demonstrated strong generalization and robustness.
Currently, Zhiyuan Robotics has connected the entire loop from model training to real-world deployment through its Genie Studio development platform. This means that GO-2 is no longer just a demonstration model in the laboratory, but a productive system capable of continuous learning and evolution through cloud data, paving the way for the practical application of embodied intelligence.
