Six AgiBot G2 humanoid robots completed a 24-hour shift at a convenience store in Shanghai's Xuhui district, running from 10 a.m. on November 2 to 10 a.m. on November 33. The livestreamed event, branded "Robots Never Clock Out," drew more than 2 million online viewers5.
The six robots logged 1,730 tasks across the shift2. The breakdown: 683 customer receptions, 369 item fetches and deliveries, 147 coffees, 126 breakfast orders, 277 restocking and shelf checks, and 128 table cleanings.
Fixed roles, four periods
Each robot held a fixed assignment: one greeted customers at the door, two handled picking, restocking and shelf patrols, one made coffee, one ran checkout and hot food, and one cleaned tables. AGIBOT divided the challenge into four operating periods covering the lunch rush, evening rush, late-night duty, and breakfast preparation1.
AgiBot's embodied agent framework, Genie Master, coordinated all six robots. The company said the test included no human demonstrations or manual on-site tuning.
Adversarial testing and probation results
Invited testers attempted to break the system: they mixed regular and sugar-free cola and then requested the sugar-free version, scrambled a full shelf of chewing gum, and blocked aisles or dropped boxes in the robots' paths. The robots confirmed drinks by combining voice requests with visual checks, restored shelves by counting items by sight rather than relying on price-tag positions, and rerouted around obstacles.
Four of the six robots passed the event's probation review, which weighed task completion, response speed, exception handling, and customer feedback.
Gaps in disclosure
AgiBot released task counts but withheld success rates and human-takeover figures. The livestream offered a longer observation window than a short demonstration but did not establish that each robot operated independently without interruption or human assistance, according to a Humanoids Daily report cited by Vending Times. Gadget Review separately noted that no task log or failure rate had been released4.
ANALYSIS The 1,730-task figure and the four-of-six probation pass rate give a partial picture of throughput but leave reliability unmeasured; without success rates or intervention counts, the demonstration functions more as a marketing event than a controlled evaluation.
The store challenge followed AGIBOT's recent AI Day announcements, which included new AI models, the G2 Max robot, and a proposed productivity metric called Worken. Earlier on November 3, AGIBOT also unveiled GO-3 Preview, an embodied foundation model that topped three major embodied AI benchmarks and is slated for open-source release by year-end[2].