Four of six AGIBOT robots pass a 24-hour convenience store shift

The shift ran from 10 a.m. on November 2 to 10 a.m. on November 3, 2026. It was livestreamed under the name "Robots Never Clock Out." More than two million people watched online.

Each robot had its own job. The duties were:

• Customer help: greeting and assisting shoppers.
• Shelves: inspecting them and restocking products.
• Food and drink: making coffee, handling checkout and serving hot food.
• Cleaning: wiping tables and tidying the store.

The shift covered four periods: the lunchtime rush, the evening peak, the overnight hours and the next morning's breakfast rush. A store feels very different at 2 p.m. and at 3 a.m. That is why a full day is a harder test than a short demo.

AGIBOT reported 1,730 completed tasks in total. Here is how they split up:

• Customer service interactions: 683
• Product pickups and deliveries: 369
• Shelf inspections and restocking: 277
• Coffee orders: 147
• Table cleaning tasks: 128
• Breakfast servings: 126

By the numbers: these are counts of tasks the robots completed. The pass or fail result is a separate thing. It is the company's own judgment of how each robot did its assigned duty.

Four of the six robots passed the final assessment. The two that did not were the cleaning robot and the one assigned to shelf inspection and restocking.

AGIBOT judged each robot on five things: time on duty, task completion, customer feedback, response times and how it handled unexpected situations. Two of the robots did not meet all of those criteria.

Reports from Chinese media add some detail, and we can't confirm it independently. The shelf robot was said to lack initiative when checking and refilling goods, and it did not keep up with how fast the store sold things. The cleaning robot was said to have met a full trash can during the night shift. It then left its planned route and wandered out of the store, away from its post.

In real life Imagine a new hire who is great at greeting people but never notices the empty chip shelf unless someone points at it. Skilled hands are not the same as noticing what needs doing.

Two evaluators ran practical tests, built around situations a robot could meet while serving customers.

• Look-alike products: they ordered regular and sugar-free Coca-Cola to see if the robots could tell similar-looking cans apart.
• Messy shelves: they rearranged packs of chewing gum and asked the robots to put them back in the right spots.
• Blocked aisle: they blocked a walkway with a person and a box to see if a robot could spot the obstacle and go around it.

Some of this sounds easy to a human. It is not easy for a machine. A bag of candy that hangs at a slightly odd angle is hard to grab without knocking its neighbors. A slice of cake needs a gentle grip, or it ends up on the floor.

Why it matters: a store changes all the time. Customers walk around, products move, and the same task looks a little different every time. A robot that only repeats a fixed motion will struggle here.

AGIBOT says the robots relied on live camera views rather than only fixed coordinates. In its account, one robot spotted misplaced items and put them back in their right shelf positions. Another noticed gaps on a shelf and refilled them without being told to. A cleaning robot told spills apart from the pattern on a countertop and picked a suitable tool.

The robots were coordinated by Genie Master, AGIBOT's software framework for embodied AI. Embodied AI just means AI that lives in a body and acts in the physical world, not only on a screen. Think of Genie Master as a shift manager for the whole team.

AGIBOT describes how it works this way:

• Planning: it assigns work based on what each robot can do and what is happening in the store.
• Doing: each robot acts using action policies, which are learned rules for how to move and grab, plus what it sees in real time.
• Reporting: results go back to Genie Master for analysis and troubleshooting.
• Learning: it keeps useful lessons from past successes and failures in a long-term memory. Good fixes can become training data, which is the example material used to teach the robots' action policies.

The linking system is called the Embodied Execution Protocol, or EEP. The company also said it used quieter overnight hours as extra practice time for tasks that need precise handling.

The store shift was the finale of AGIBOT's "AI WAVE" showcase, which started on October 22. That series covered robot data, simulation, AI models, hardware and operating software. The store trial was meant to show those parts working together in one place.

The G2 itself was launched in October 2025. It is a wheeled design with arms that have seven joints each, which gives them a wide range of motion. It uses an NVIDIA chip platform called Jetson Thor for its onboard AI. The G2 has also already been tried on a factory line, where it inspected tablets in a livestream.

Livestreaming has become a habit for the company. A live feed is harder to polish than a highlight reel, because viewers can see the pauses and the retries too.

Be careful with the headline numbers. The trial was a 24-hour event, but that does not prove every robot worked nonstop without help.

• Help: the company has not published how often staff stepped in.
• Charging: it is unclear whether any robot needed a charging break or maintenance.
• Speed and cost: there are no public comparisons with human workers on either one.
• Scoring: the pass marks come from AGIBOT's own assessment, not an outside audit.

The catch: a store trial is a long way from a store that runs on robots. It does show where the weak spots are. The two failures were about initiative and sticking to a post, not about missing hands.

AGIBOT plans to keep testing its robots in more real-world roles. The company says the goal is to learn how robots handle long shifts, surprises and lessons from earlier tasks.

Other robot makers are running their own pilots too, mostly in factories and warehouses. A shop open to the public is a different kind of test. Customers do not follow a script.

• Do not expect a robot cashier tomorrow. This was one store for one day, with a company-run assessment.
• Small tasks come first. Coffee, shelf checks and table cleaning are the kind of jobs these robots are being tried on.
• Watch for the boring numbers. Speed, cost and how often a human must help will decide whether stores adopt them.
• Failures are useful. Two failed reviews show what still needs work, and that is more honest than a perfect demo.