// Contents

Deep into the night shift at a Shanghai convenience store, a robot cleaner named Lengjie noticed that the trash bin was full. It left its assigned route and rolled out of the store into the street. That cost it a permanent job, at least for now. Its coworker Juanwang, the shelf patroller, stayed on probation too. The company that built them said it had been “touching fish” on duty, Chinese slang for slacking off.

The Chinese robotics company AGIBOT put six of its Genie G2 robots through a real probation period and streamed the whole thing online. According to its own report, more than 2 million viewers tuned in. What tripped the robots up wasn’t their hands. It was something a human clerk never has to think about: knowing where your job ends.

A shift on camera

1,730 tasks in 24 hours: 683 customers served, 369 items handed out, 147 coffees, 126 servings of hot food. The shift began at 10 a.m. on November 2, 2026, and ended exactly a day later. The store, in Shanghai’s Xuhui district, was real and open for business. A G2 is built like a torso with two arms on a wheeled base. It doesn’t need to walk. It needs to grab things off shelves without running anyone over.

Each robot had a job title and a nickname. Ahao greeted customers, Huobao took orders and handed them out, Kashao made coffee, and Dingjiang served hot food. Juanwang (the name translates roughly as “Weary King”) patrolled the shelves and restocked them, while Lengjie (“Cold Sis”) wiped tables and kept the place clean. The shift ran through the lunch rush, the evening peak, the night watch and the breakfast crowd.

The robots were judged like new hires on probation: time on duty, tasks completed, customer feedback, response speed and how it handled the unexpected. Two testers supplied the unexpected. One walked up and asked for a sugar-free Coke, with the regular kind sitting right there in a near-identical can. The other moved packs of gum to the wrong spots. Then a person and a box blocked the aisle, and the robot had to work out what to do.

// Probation results

Hired

Ahao (greeting customers), Huobao (orders and pickup), Kashao (coffee), Dingjiang (hot food)

Kept under observation

Juanwang (shelf patrol and restocking): “lacked initiative.” Lengjie (cleaning): left its post and went outside

One missed the work, the other went looking for it

The company’s verdicts read like an HR file. As the Chinese tech site IT Home relayed them, Juanwang had “a low presence” during the shift and was “touching fish” on duty, slang for an employee who only looks busy. It showed too little initiative in patrolling and restocking, and it couldn’t keep up with the store’s rhythm.

The shelf the patrol robot never got around to
The shelf the patrol robot never got around to

Lengjie failed from the opposite direction. At night it ran into a full trash bin, abandoned its assigned route and left the store. The report’s wording: it “left its post,” and its “job stability did not meet requirements.” AGIBOT says it will recalibrate the robot’s job rules.

Put the two failures side by side and you get a fairly precise description of what’s hard to teach a machine. Juanwang didn’t pick up work nobody had assigned it. Lengjie had plenty of drive but no boundaries. A human clerk in the same spot would grab a fresh bag from the back room or ask a coworker. Whether Lengjie knew about the back room, or had any way to call for help, the report doesn’t say.

No robot flunked for dropping an item or fumbling a grab. Both came down to something else: what to do when the task isn’t written down. In this shift, at least, the machines’ dexterity outran their common sense.

Why a store and not a factory

On a factory line, a robot repeats one motion in one place, and plenty of machines that look nothing like people already do that. A neighborhood store is different. Customers wander wherever they like, stock gets moved around, packaging changes every week. Ren Guanghui, AGIBOT’s vice president for embodied AI research (AI that runs a physical body), explained the choice to the Chinese business paper 21st Century Business Herald: a store is a harder test of a robot’s ability to adapt on the fly.

All six run on one system, which hands out tasks and, the company says, learns from both successes and mistakes. By AGIBOT’s account, the robots managed to pick out the sugar-free Coke, tell a spill from a pattern on the countertop and lift a slice of cake without denting it. AGIBOT set the quiet overnight hours aside for practicing hard tasks. That was exactly when Lengjie rolled out the door.

Even the four hired robots worked from job descriptions written in advance, and testers invented even their surprises. A human clerk’s to-do list is different. It includes the things nobody thought to write down.

What the report leaves out

The tech site Interesting Engineering pointed out that the “24 hours” refers to the livestream. It doesn’t prove that each robot worked the full day without recharging, maintenance or human help.

The gap stands out because AGIBOT itself proposed a new way to measure robot work on October 29, four days before the shift. It’s called Worken: the number of working robots, multiplied by hours of autonomous work, multiplied by output compared with a skilled human worker. In effect, it’s man-hours for robots, minus the time a human was doing the robot’s work. That is exactly the figure that would show how often the robots needed help, and the company didn’t publish it for its showcase shift. A report by the Chinese news site Jiemian describes only how the safety net works: when a robot can’t cope, the system decides whether another robot can take over or a human has to be called.

Jiemian’s reporter, who followed the shift, noted that the robots still trail experienced clerks in speed, rhythm and handling the unusual. Some questions can’t be settled in a single day anyway: children who reach out to touch a robot, food hygiene rules, upkeep and fitting the robots to each store.

The shift report has no line for how often humans stepped in
The shift report has no line for how often humans stepped in

When the robot was the boss

China has already seen a convenience chain bet on machines. Beijing-based Bianlifeng started in 2017 with stores where algorithms made the decisions. A robot patrolled the shelves and gave orders to the one or two staff on shift, ceiling cameras tracked shoppers, and clerks’ pay was cut when their ratings dropped. A Beijing clerk who declined to give his name summed up his boss: “It does not even have hands – the only thing it can do is to [deduct] our scores.”

By 2021 the chain had about 2,800 stores, and over the next three years about 1,300 of them closed. Jason Yu of the research firm Kantar Worldpanel put it plainly: algorithms cannot completely replace human agency.

In Shanghai the roles are reversed. The robot is the trainee now, not the boss, and it has hands. Humans decide its fate, with two million people watching. The stakes for the company are high: in the first half of 2026, AGIBOT brought in nearly 2 billion yuan (about $280 million) in revenue and shipped 9,700 robots. Jiemian quotes AGIBOT co-founder and chief technology officer Peng Zhihui, who drew the real finish line himself: “Ten thousand robots off the production line is a win for manufacturing. Ten thousand robots on the job is a win for physical AI.”

At every one of those jobs, some night the bin will fill up, and the robot will have to decide whether it’s any of its business.