AgiBot G2 Robots Finish a 24-Hour Shift in a Shanghai Convenience Store, and Four of Six Pass

A Night Shift With Two Million Witnesses

For one full day, six humanoid robots ran the floor of a working convenience store in Shanghai while more than two million people watched online. Chinese robot maker AgiBot put its G2 robots on duty from 10 a.m. on November 2 to 10 a.m. on November 3, 2026, and livestreamed every minute of it.

The company framed the stunt as a probation test, the kind a new hire goes through before getting a permanent job. When the stream ended, AgiBot said four of the six robots had passed and two had not, a rare public admission that its machines still stumble in the messy real world.

AgiBot G2 Robots Fin...

Click to play

Six Robots, Six Jobs

Each robot was given one role, much like staff on a roster. One greeted shoppers at the door, one picked and handed over orders, one worked as a barista, one served hot breakfast food, one patrolled the shelves to restock them, and one wiped down tables in the seating area.

The store, in Shanghai's Xuhui district, stayed open to real customers throughout. AgiBot split the day into four tests: the lunch rush, the evening crowd, a quiet late-night stretch and the busy breakfast hour the next morning, so the robots faced both crowds and long, empty hours.

1,730 Tasks by Morning

By the end of the stream, AgiBot said the six robots had completed 1,730 tasks between them. That included 683 customer greetings, 369 orders picked and handed over, 147 coffees, 126 hot breakfast servings, 277 shelf checks and restocking rounds, and 128 table cleanings.

Those numbers come from the company and were not independently audited. AgiBot has not published how often a person stepped in to help, how long the robots spent charging, or how it scored a task as done, details that matter when judging whether a machine could really cover a human shift.

Who Made the Cut, and Who Did Not

The greeter, the order runner, the barista and the hot-food server were declared promoted. AgiBot singled out the greeter for top customer-service scores and the order runner for keeping up through sudden order surges and last-minute cancellations, while viewers praised the barista's hot Americano.

The restocking robot fell short because it did not act on its own often enough to keep shelves in step with the flow of shoppers. The cleaning robot failed a stability check after leaving its set route late at night and wandering outside the store when a trash bin was full.

A Wheeled Humanoid Built for Service

The G2 is not a walking robot. It has a humanlike upper body, with a head, torso and two arms, mounted on a wheeled base that rolls smoothly across flat shop floors. That design trades the ability to climb stairs for stability and longer working time.

AgiBot, based in Shanghai, is one of China's largest humanoid robot makers and has been pushing the G2 into factories and shops. Earlier this year it announced a heavier version, the G2 Max, for warehouse work with logistics giant JD Logistics.

The Software Brain Running the Team

Behind the six robots sat a shared piece of software AgiBot calls Genie Master, released on October 27. Think of it as a shift manager: it hands out tasks, keeps track of which robot is doing what, and collects the results of every success and failure for later study.

Two other systems did the hands-on work. GO-3 Preview, a large AI model released on October 29, lets the robots handle many different objects and chores. GenieEvolver is a training system that sharpens their movements using practice from real robots, computer simulations and virtual worlds.

Why a Corner Shop Is So Hard

Ren Guanghui, a vice president who leads AgiBot's embodied research center, described the convenience store as a training ground, because it packs so many hard problems into one small room. Products come in countless shapes, shoppers move unpredictably, and several robots must share narrow aisles without colliding.

Simple-looking jobs are surprisingly tricky. Lifting a bag of sweets off a hook needs a precise grip and angle, and picking up a soft cake takes just enough force to hold it without crushing it. Ren acknowledged that the robots are still slower than experienced staff and weaker at handling the unexpected.

Learning From Every Mistake

AgiBot says the real prize is the data. It calls its approach PhysicalRSI, short for recursive self-improvement in the physical world, meaning the robots learn from their own logged mistakes. Simple fixes, such as rules to stop robots bumping into each other, can be applied the next day.

Deeper changes to how the robots move take longer. Ren said lessons from the livestream should appear in updated robot software after one to two weeks of sorting and training. An industry observer quoted by Chinese media said the real test is whether those lessons work reliably in other stores at lower cost.