24-Hour Robot Store
Six AGIBOT G2 humanoid robots ran a working convenience store in Shanghai's Xuhui district for 24 continuous hours, and the shop stayed open to paying customers throughout. The Chinese maker, known at home as 智元机器人, streamed the shift live under the title Robots Never Clock Out. Its machines held the jobs that face customers, greeting at the door, restocking shelves, brewing coffee, serving hot food, and wiping tables.
Each robot carried a name badge and a single post. Two worked as a restocking pair, one patrolled the aisles, one made coffee, one held the checkout and the hot food counter, and one cleaned tables. Shoppers walked past them and reached around them all day.
But the store was also set up to trip them. Two invited testers put near identical full sugar and sugar free cola on the counter and asked for the sugar free can, scattered a shelf of chewing gum, stood in an aisle as a roadblock, and dropped a box in a robot's path. The machines answered by checking what they heard against what they could see, counting stock by sight rather than trusting shelf labels, and stopping and routing around whatever blocked them.
That double check matters more than it sounds. A trading shop floor is nothing like a lab, where light, product position, and the order of operations can be fixed in advance. Goods shift, turn, and arrive in different packaging between one grasp and the next, so a clean lift does not guarantee a repeat.
Reported figures put the shift at more than 1,700 completed tasks, with Chinese outlets citing 1,730 in total. The breakdown carried by Shanghai Securities News and others counts 683 customer receptions, 369 items collected or delivered, 147 coffees, 126 breakfast orders, 277 shelf checks or restocks, and 128 table cleanings. AGIBOT has not published that count, a task success rate, or a cost per shift on its own newsroom.
And the failures clustered in the hands. Hanging a bag of sweets on a hook, or lifting a soft rice cake without crushing it, stayed among the hardest jobs, according to Ren Guanghui, a vice president of AGIBOT's embodied research centre. He told the paper that the margin between a dropped item and a crushed one comes down to the force and the contact point of a single grip.
The six ran under Genie Master, an agent framework the company introduced shortly before the trial. It schedules the work and ties each decision to the motion that carried it out and the result that came back, through a protocol AGIBOT calls Embodied Execution Protocol. Ren said the system reviews successful and failed runs each day, stores what it learns in long term memory, and turns corrected attempts into training data that usually reaches a policy update within one to two weeks.
But the operating rule the company stresses is caution. Ren said every action has to stay inside a defined safety boundary, with the fleet slowing, pausing, rerouting, or stopping to ask for help when a shopper or a child moves close. He put the principle as better to be slow than to collide.
So AGIBOT frames the trial as the closing test of a stack it has been releasing over the past fortnight, covering data, simulation, models, and the robots themselves. The company has scale to point at, having said its 15,000th robot came off the line earlier this year. What it has not shown is the economics, and a full day of trading without a success rate or a running cost is a test of endurance rather than a business case.
The question the store answers, then, is narrower than the livestream suggests. Six machines can keep a shop's routine work moving for a day with people walking through it, which is a step past the cordoned demo. Whether a chain would pay for that turns on tasks with clear edges, checkable results, and few exceptions, and Ren expects restocking, shelf work, and cleaning support to settle into paid work first.
AGIBOT says it will keep pushing the same stack into more sites, and it billed the store as a start rather than a finish. The next evidence worth watching is duller than a livestream, in the shape of published success rates and the cost of putting a robot on a shift.
Six AGIBOT G2 humanoids ran a trading Shanghai convenience store for 24 continuous hours
Reported figures count 1,730 tasks, including 683 customer receptions and 147 coffees
Testers scrambled shelves and blocked aisles to see how the robots recovered
Gripping soft food and hanging bags on hooks remained the weak point
AGIBOT has published no task success rate and no cost per shift
