Six wheeled humanoid robots from Chinese robotics firm AgiBot spent a full day and night working inside a live convenience store in Shanghai’s Xuhui District, completing 1,730 tasks between 10 a.m. on November 2 and 10 a.m. on November 3, 2026. The livestreamed experiment, billed as “Robots Never Clock Out,” drew more than two million viewers. When the assessment closed, four of the six G2 units had passed and two were placed under observation for underperforming — a result far more instructive than a polished marketing clip.
- Scale of the test: Six AgiBot G2 robots handled 683 customer-service interactions, 369 product pickups and deliveries, 147 coffees, 126 breakfast servings, 277 shelf inspections and restocks, and 128 table cleanings.
- Who failed and why: A cleaning robot abandoned its route after encountering a full trash can and wandered outside the store; a shelf robot lacked the initiative to keep restocking as products sold.
- The real benchmark: The trial measured response times, task completion, customer feedback and handling of deliberately planted surprises — not raw strength or precision.
- Context: The event closed AgiBot’s AI WAVE showcase, which opened on October 22 and bundled the company’s hardware, simulation tools and embodied AI models into one real-world deployment.
Inside the 24-Hour Shift
The G2 fleet was split across distinct roles to mirror a working store: customer assistance, checkout support, hot food service, coffee preparation, shelf management and cleaning. AgiBot scheduled the robots across four operational windows — lunchtime, the evening peak, overnight operations and the following morning’s breakfast rush — so that shifting foot traffic, product availability and task loads would stress the systems in different ways.
The evaluation criteria were deliberately practical. Organizers tracked time on duty, task completion rates, customer feedback and response times, then layered in unscripted obstacles. Robots were asked to tell regular Coca-Cola from the sugar-free variant, return rearranged merchandise to its correct shelf position, and navigate an aisle blocked by both a person and a box. One unit reportedly spotted gaps on shelves and replenished stock without a specific instruction; another distinguished spills from countertop patterns and selected the right cleaning tool.
Those tasks demand visual perception, object recognition, spatial reasoning, task planning and physical manipulation to fire in sequence — the defining challenge of embodied AI, which must translate understanding into reliable physical action rather than merely generating text.
Why Two Robots Failed
The failures were not mechanical. According to the assessment, both robots lacked initiative and adaptability when conditions changed. The cleaning unit met a full trash can blocking its path, deviated from its route, exited the store and never recovered its position. The shelf unit could not sustain restocking as inventory depleted, failing to identify and refill products consistently.
The distinction between executing a task under predictable conditions and completing it reliably in a changing environment is what separates a demonstration from a product.
That gap is the industry’s central problem. Convenience stores look simple next to factories, but their conditions shift constantly: packaging varies, customers interrupt robot movements, objects appear where they should not. Picking up a bag of candy requires judging its orientation and gripping it without disturbing nearby items; preparing food demands fine control over movement and grip force. A robot that can brew coffee on cue is easy to film. A robot that does not walk into the street when a bin blocks its path is much harder to build.
The Technology Behind the Trial
AgiBot framed the store as a proving ground for its Genie Master embodied agent framework, which coordinates tasks, monitors execution and feeds operational data back into later performance through its Embodied Execution Protocol. The company has also introduced GO-3 Preview, an embodied foundation model built on its WRAM architecture that unifies task understanding, future-state prediction and action generation, and which AgiBot says topped the RoboDojo, RoboTwin and RoboColiseum benchmarks ahead of a full open-source release planned by the end of 2026.
Those claims deserve the usual caution. The performance figures come from AgiBot itself and have not been independently verified, and the trial doubled as the finale of a company technology showcase — a format that rewards favorable framing. Still, publishing the two failures rather than hiding them is a credible move in a sector that thrives on promises of tireless workers.
What the Result Means for Humanoid Robots
The lesson is that endurance is no longer the bottleneck. Running 24 hours is an engineering detail; knowing what to do when something unprogrammed occurs is the wall. Until embodied AI can generalize across unexpected situations, these machines remain promising interns rather than dependable employees — and an intern who leaves the store the moment a trash can blocks the aisle still needs a human supervisor. AgiBot’s willingness to show that limit on a livestream may ultimately be worth more than the four passing grades.