Six humanoid robots ran a working convenience store in Shanghai for 24 hours, from 10 a.m. on November 2, 2026, to 10 a.m. on November 3, and four of them passed their own company’s review. AGIBOT, the manufacturer, livestreamed the event, titled “Robots Never Clock Out”, to more than two million viewers, and reported that the six G2 machines completed 1,730 tasks. The two that failed are the more informative part of the result, because their failures were not about strength or dexterity.
Video: news report on the AGIBOT convenience store test, from the ShanghaiEye channel on YouTube. The task counts and assessment results in this article were reported by AGIBOT and relayed by the outlets listed below.
The setup
The store sits in Shanghai’s Xuhui district and stayed open to regular customers throughout, according to Interesting Engineering. AGIBOT split the day into four shifts: the lunch rush, an evening peak the company called a stress test, overnight watch, and a morning breakfast sprint. Each robot held one job. Gasgoo’s account lists a greeter named Ahao, a clerk named Huobao, a barista named Kasha, a hot-food manager named Dingjiang, a patrol and restocking robot named Juanwang, and a table-clearing robot named Lengjie.
The G2 is a wheeled humanoid, so the test said little about walking. It was a test of hands, planning and coordination in a space built for people. Gasgoo quoted Ren Guanghui, vice president of AGIBOT’s Embodied Intelligence Research Center, explaining the choice of venue with “Real enough, and hard enough.” Special inspectors were also invited to make things difficult on purpose, and the stress tests described in the coverage included telling regular Coca-Cola from sugar-free, re-shelving chewing gum that had been rearranged, and getting around an aisle blocked by a person and a box.
The task counts
AGIBOT’s totals, as reported by Interesting Engineering, add up to the headline figure of 1,730:
- Customer-service interactions: 683
- Product pickups and deliveries: 369
- Shelf inspection and restocking: 277
- Coffee orders: 147
- Table-cleaning assignments: 128
- Breakfast servings: 126
Gasgoo’s per-robot breakdown fits the same numbers: 369 pickup-and-delivery orders for the clerk, 147 cups of coffee for the barista, 126 hot meals for the food manager, 277 patrol and restocking actions for Juanwang and 128 table clearings for Lengjie. Early in the run, a progress report cited by Magica had counted 337 tasks after 4.5 hours on duty. That report established no success rate, and neither has the final tally: Gadget Review noted that no task log or failure rate had been released.
That is the main caveat on the numbers. A count of tasks completed is not a rate of tasks attempted, and every figure here comes from AGIBOT. Independent verification is absent from all of the coverage reviewed.
Why two robots failed
AGIBOT graded the robots on time on duty, task completion, customer feedback, response times and handling of unexpected situations. Four passed. Two went into an observation group for further improvement, a pass rate of 66.7 percent for the machines. Gasgoo gave the longer list of criteria that the company says would matter for deployment: task success rate, first-attempt success rate, human takeover frequency, anomaly recovery time and system availability.
The reasons for the two failures were behavioral. Juanwang, the patrol and restocking robot, lacked initiative, rarely acted without being prompted and did not keep pace with the store. Lengjie drifted from its planned route and walked out of the store, which Gasgoo read as poor positional stability. According to TechSpot, as relayed by Interesting Engineering, a full trash bin had blocked the cleaning robot’s route and it never returned to its post. The shelf robot, in the same account, did not consistently identify and refill products as stock sold.
Gasgoo drew the common thread: both failures involved the drive to start a task and the ability to respect boundaries when the rules do not cover a situation. Those are planning and judgment problems. A robot that can grip a coffee cup well but cannot decide that a shelf is running empty is not limited by its actuators.
What stayed hard
Ren was candid about the gaps in the Gasgoo interview. He gave two examples of tasks that look trivial and are not. One is the golden sponge cake, where a robot must grip tongs with enough force to hold the cake but not crush it, drop it into a bag, then switch grips to lift the bag. The other is a bag of Skittles hanging on a hook, where the hole in the packaging has to slide off without snagging the wrapper or knocking over nearby products. His explanation for why soft goods are difficult was that “Once a soft package deforms under force, the conditions for all subsequent actions change entirely.”
The article listed other weak spots: deformable objects, complex contacts and long multi-step operations, slower adaptation to on-site changes, and a fine-manipulation rhythm that trails a seasoned clerk. Early ordering-system glitches produced failed orders, missed status updates and positional conflicts between robots. Human takeover frequency was tracked as a metric, the system can request human help, and on-site adaptation and maintenance still rely on engineers, which Ren said keeps the cost of replicating the setup in another store high.
On deployment timing, Gasgoo quoted him as saying the distance depends on the task, not on a timeline.
What AGIBOT says would make it scale
Gasgoo set out the company’s own tests for when a task is ready to spread across stores: the task needs clear boundaries, results that can be verified, and an anomaly rate low enough for the system to recover by itself. On the full-stack approach AGIBOT is betting on, success would look like higher qualified output with less retry and takeover time, less debugging when items or store layouts change, improvements that do not degrade skills already learned, and costs that fall as the system matures.
Ren also described why a store resists tidy planning, saying that “Restocking might block a pickup spot; cleaning could clog an aisle.” Jobs that look separate on a roster compete for the same floor space, which is a coordination problem rather than a manipulation one. On safety his stated rule was blunt: better slow than sorry, never collide.
Taken together, those criteria describe a gap between a demonstration and a product. A 24-hour livestream with inspectors and four passing robots shows that a coordinated fleet can hold together for a day in real conditions. It does not show how often a human stepped in, how long recoveries took, or what a second store would cost to set up, and the company itself lists those as the numbers that matter.
The software behind it
Two layers coordinated the robots. The Genie Master framework links task planning, execution and feedback, and the Embodied Execution Protocol assigns work and collects the results. By AGIBOT’s account, it keeps both successful and failed attempts and turns effective adjustments into training data. At its Innovation Day on October 29, AGIBOT also proposed a measure it calls Worken, short for Effective Intelligent Work Hours, defined by Gasgoo as the number of active robots multiplied by autonomous work duration multiplied by an output-efficiency ratio.
The same event introduced the company’s GO-3 Preview embodied model, which, per Let’s Data Science, ranked first on two evaluations in company-presented results, and which Interesting Engineering reports AGIBOT plans to open-source by the end of 2026. None of the store coverage says whether GO-3 ran in the stores, so the link between the model and the trial is unestablished.
Where the market stands
Gasgoo cited Counterpoint data for the first half of 2026: global humanoid shipments above 22,000 units, up nearly 300 percent from a year earlier, with Chinese firms holding about 86 percent of the top five slots. AGIBOT shipped about 9,700 units, over 43 percent of the total, and Unitree more than 7,000, about 31 percent. Entertainment, commercial performance, education and data collection accounted for more than 60 percent of use, guided tours and retail greeting about 19 percent, smart manufacturing 13 percent and logistics 5 percent.
Set against those proportions, a convenience store is a harder job than most robots currently hold, which is the reason AGIBOT chose it. The same week brought a different kind of deployment news from the United States: Agility Robotics’ Digit humanoid is slated for a manufacturing pilot at an AGCO plant in Hesston, Kansas, beginning in 2027, according to High Plains Journal, with tasks such as sorting and stacking parts and placing totes in storage. For a look at one of the engineering trade-offs behind hands like these, the earlier article on why Boston Dynamics dropped the pinky on Atlas covers how a hand is simplified for practical work.
Sources
- Interesting Engineering, November 7, 2026
- Gasgoo, AgiBot’s Convenience Store Test: The Challenge Is Not Working, but Keeping It Going
- Gadget Review on the livestream and missing task log
- Magica on the early task count
- Let’s Data Science on the GO-3 Preview
- High Plains Journal, November 8, 2026
- ShanghaiEye on YouTube
