AgiBot's GO-3 Preview model took a store robot from 28.6% to 85.7% success in five days, and opens up at year-end

AgiBot's new robot foundation model is being sold on what it learns after it starts work. GO-3 Preview, which co-founder and co-president Yao Maoqing unveiled at the company's Innovation Day on Oct. 29, was put to work in a Shanghai convenience store: a G2 robot running the new stack, with the Genie Master agent framework, packed deli food, fetched items and restocked shelves. AgiBot says the robot's success rate on the full task flow rose from 28.6% on its first day to a steady 85.7% after five days, with no human demonstrations or manual tuning.

That loop is the product. AgiBot pairs GO-3 with Genie Evolver, a post-training system that draws on about 100 real robots plus physical simulation and world models, and Genie Master, which lets a deployed robot organize long tasks and keep what it learns on site. "A demo tests the ceiling of a single capability; productivity tests the floor of the system," Yao said, according to Sina Tech. AgiBot plans to open-source both GO-3 Preview and Genie Master by the end of 2026.

Both store figures are sevenths

28.6% and 85.7% are exactly 2/7 and 6/7. That fits seven full task runs a day, or a multiple of seven, so the five-day gain may rest on a few dozen attempts. If the curve holds over hundreds of runs, how fast a model improves on the job becomes the number to compare between robot models; if it doesn't, the store result is a demo with a curve attached.

AgiBot's other learning demos point the same way. On a single task, the company says, online learning beat conventional supervised fine-tuning by 17.7 percentage points. After four hours of online practice, a robot outscored the average tested human at tabletop curling, and it pulled five blocks in a row from a tower without toppling it.

More data and a bigger model both moved the zero-shot curve

AgiBot also showed scaling results from its own experiments. More training data lifted the zero-shot average success rate on 100 tasks across 20 skill types from 13.4% to 56.4%, and a larger model lifted it from 56.2% to 73.2%, Sina Tech reported. The architecture, which AgiBot calls WRAM (World Reasoning Action Model), models understanding, prediction and action together and learns from first-person human video alongside robot trajectories, according to Interesting Engineering. AgiBot says it cuts visual input to one-sixteenth of DINOv3's tokens and lowers action-recovery error by 35%, and that GO-3 Preview tops the RoboDojo, RoboTwin and RoboColiseum leaderboards.

The larger count comes next. AgiBot has scheduled six G2 robots to work a real convenience store for 24 hours on Nov. 2-3, livestreamed across four shifts. A day of live shifts will run far more than seven attempts.