AgiBot's six robots logged 1,730 tasks in a 24-hour store shift, and this time no success rate

Six AgiBot G2 robots worked a full day in a Shanghai convenience store that stayed open for business, and AgiBot's final tally is 1,730 completed tasks. The livestreamed run went from 10 a.m. on Nov. 2 to 10 a.m. on Nov. 3 Beijing time, Securities Times reported, citing People's Finance News: 683 customer greetings, 369 items fetched and delivered, 277 restocking and inspection rounds, 147 coffees, 128 table cleanings and 126 breakfast orders, watched by more than 2 million people online.

AgiBot closed this run with a count. Its June test ended with a rate: eight G2s on the quality-inspection stage of Longcheer Technology's tablet line in Nanchang logged 64,828 operations in 64 hours with a 99.99% success rate, PEdaily reported. In the store, AgiBot says it records task success rate, first-attempt success, human takeover frequency and system availability, among other measures. None of them is in the final tally. Until they are, the store shows that six robots can hold six jobs for a day, and leaves open how often they got them right.

The factory robots did ten times the work per hour

On AgiBot's own numbers, each factory G2 averaged about 127 operations an hour (64,828 over 64 hours across eight robots, assuming all eight ran the whole time). Each store robot averaged about 12 tasks an hour (1,730 over 24 hours across six). The strongest case for the count is that store work is paced by customers: a greeter waits for someone to walk in, a barista for an order. Ren Guanghui, a vice president at AgiBot's embodied AI research center, told reporters that six robots were used to cover the store's main roles and to test how those roles work together. The six jobs were greeter, pick-and-pack, stocker, cleaner, barista, and hot food plus checkout.

The robots did miss. They occasionally failed tasks during trading hours, Wenhui Daily reported. AgiBot describes the run as "no human in the loop": no human demonstrations and no on-site manual tuning, with the robots finding the cause of a failure and adjusting on their own.

The store data goes back into training within two weeks

That loop is the same pitch AgiBot made for GO-3 Preview, whose earlier store trial went from 28.6% to 85.7% success in five days. AgiBot told reporters that lessons at the agent layer are summarized the same day and used the next, that policy updates take a few days a round, and that data from the livestream, once filtered, de-identified and checked, normally shows up in a policy update within one to two weeks. On that timeline, what the six robots got wrong on Nov. 2 should be in their policy by mid-November.