Six Humanoids Ran a Shanghai Store for 24 Hours Straight
Six AgiBot G2 humanoid robots worked a 24 hour shift in a Shanghai convenience store that stayed open to paying customers, and the company counted 1,730 tasks completed by the end.
The store in Xuhui district ran from ten in the morning to ten the next morning with the shift livestreamed, and AgiBot says more than two million people watched it happen in real time rather than in an edited film. Each robot had one role, with one greeting at the door, two handling picking and shelf patrols, one making coffee, one on the checkout and hot food counter, and one cleaning tables.
And the task list is more revealing than the headline number. The six logged 683 customer receptions, 369 item fetches and deliveries, 147 coffees, 126 breakfast orders, 277 restocking and shelf checks and 128 table cleanings, which is a day of real retail work rather than a demonstration of one trick repeated.
But the totals hide the rhythm of the day. Spread across 24 hours, 683 customer receptions works out at about 28 an hour and 147 coffees at about six, which describes a quiet store rather than a rush, and that is the honest setting for a first shift of this kind.
But the testers were invited to make it hard. They mixed regular and sugar free cola and asked only for the sugar free can, scrambled a shelf of chewing gum, stood in aisles and dropped boxes in the robots' paths, and the robots answered by checking voice requests against what they could see and by counting items on a shelf rather than trusting price tag positions.
And the failures were in the hands. Hanging a bag of sweets on a hook or lifting a soft cake without crushing it stayed among the hardest tasks, which is the familiar problem for humanoids, where navigation and recognition have improved faster than fine manipulation. A hand that can grip a can is not yet a hand that can hook a bag onto a rail at the first attempt.
Genie Master coordinated the shift, linking task decisions, motion and feedback through what AgiBot calls an Embodied Execution Protocol, and the system reviews successful and failed runs, stores experience in long term memory and feeds corrected attempts back into training data. The company says no human was in the loop, meaning no demonstrations and no on site tuning during the run.
So the experiment is best read as a test of the learning loop rather than of six robots. Four of the six passed a probation review that weighed completion, response speed, exception handling and customer feedback, which implies two did not, and the store kept trading throughout the full shift with customers walking in.
So the next test is repetition. A single 24 hour run shows the loop can close once, and what a retailer will ask is whether six machines can do it again on an ordinary weekday months later, with the store manager rather than a robotics researcher in charge.
Six AgiBot G2 robots worked a 24 hour shift in a Shanghai convenience store open to customers
The company counted 1,730 completed tasks, including 683 customer receptions and 147 coffees
Four of the six passed a probation review based on completion, speed, exceptions and customer feedback
Fine manipulation such as hanging bags on hooks remained the hardest category
683 customer receptions across the day is roughly 28 an hour
Source: Pandaily
