Zhiyuan Unveils GO-3 Preview Foundation Model, Genie Master Framework, and Genie G2 Max Robot

Related topics top start Related topics top end Futures promotion begin Futures promotion end Miaopai begin Miaopai end Video player start Video player end Market chart begin Market chart end

Sina Tech News, evening of October 30: At the Embodied Work Productivity sub-forum of 2026 Zhiyuan Innovation Day, Zhiyuan unveiled achievements including the embodied foundation model GO-3 Preview, the embodied Agent framework Genie Master, and the heavy-duty embodied industrial robot Genie G2 Max, and demonstrated a deployable, operable, and continuously optimizable Physical AI productivity technology loop.

"A demo tests the upper limit of single-point capability, but productivity tests the lower limit of the system." In the view of Yao Maoqing, co-founder, co-president, and president of the embodied business system of Zhiyuan, robot work productivity is a set of strict engineering standards: measurable—success rate, cycle time, and MTBF (mean time between failures) are all quantified; replicable—not one demonstration line, but hundreds or thousands of production lines; sustainable—not delivery as the endpoint, but getting better with use. Only when embodied intelligence possesses all three capabilities at the same time can it truly enter industrial sites.

It is reported that the GO-3 Preview released this time, through the unified WRAM architecture for understanding, prediction, and action generation, can achieve cross-scenario generalization and fine manipulation capabilities. Experiments announced on site show that in the training data scale expansion experiment, the model's Zero-Shot average success rate on 100 tasks across 20 skill categories increased from 13.4% to 56.4%; in the model scale expansion experiment, the Zero-Shot average success rate increased from 56.2% to 73.2%. The two sets of experiments respectively demonstrate the capability improvement trends brought by data scale and model scale expansion.

This means that the capability improvement of physical intelligence is beginning to show a verifiable Scaling law. Recently, GO-3 Preview has topped three major embodied intelligence evaluation leaderboards: RoboDojo, RoboTwin, and RoboColiseum.

In addition, targeting the uncertainty in real production tasks, Zhiyuan Genie Evolver can combine real machines, simulation, and world models: real machines provide real feedback, simulation expands coverage, and world models help robots explore the consequences of different decisions in key states. In the mobile phone packaging task, a collective experience learning system composed of 96 real machines achieved a task success rate of 98.3%; in the mobile phone charging task, the success rate increased from 60.0% to 97.5%.

After robots are deployed, the continuous evolution of their capabilities is equally important. Based on Zhiyuan's newly launched embodied Agent framework Genie Master, it can organize long-horizon tasks, maintain progress, check execution results, and support interruption and resumption after exceptions. In a convenience store in Shanghai, Genie G2 equipped with Genie Master has achieved continuous improvement in complex and fine tasks such as prepared food packaging, picking, and restocking, with the success rate of completing task processes rising from 28.6% on the first day to a stable 85.7% by the fifth day. At the same time, deployment trajectories can further be used to optimize action strategies. (Wen Meng) news_keyword_pub,stock,sz300024 iframe below main text iframe below main text General statement General statement QR code at end start QR code at end start

QR code at end start QR code at end start Editor name and work code

Editor name and work code