Six AgiBot G2 Humanoids Ran a Shanghai Convenience Store for 24 Hours: 1,730 Tasks, Four of Six Robots Passed

Chinese robotics firm AgiBot put six of its wheeled G2 humanoid robots to work in an operating convenience store in Shanghai's Xuhui District for a full 24 hours, running from 10 a.m. on November 2 to 10 a.m. on November 3, and says the machines completed a combined 1,730 tasks while more than two million people watched the livestream. The trial, announced in a GlobeNewswire release on November 8, was the finale of the company's AI WAVE technology showcase, which launched on October 22 with a string of releases spanning robotics data, simulation, AI models, hardware and operational software.

The task breakdown is the most granular part of the announcement. Over one day and one night, the six robots logged 683 customer-service interactions, 369 product pickups and deliveries, 147 coffee orders, 126 breakfast servings, 277 shelf inspections and restocking tasks, and 128 table-cleaning tasks. Each robot carried a different job description - customer assistance, shelf inspection, restocking, coffee and hot food, checkout, cleaning - and the trial deliberately covered four operating periods: the lunchtime rush, the evening peak, overnight operations and the following morning's breakfast rush.

The honest headline number is the pass rate: four of the six robots met the final assessment criteria, two did not. Evaluators also ran stress tests that had little to do with routine operation - ordering regular and sugar-free Coca-Cola to see whether the robots could tell near-identical bottles apart, rearranging packs of chewing gum on a shelf and asking the robots to restore the layout, and blocking an aisle with a person and a box to test obstacle recognition and rerouting. AgiBot argues that a convenience store is deceptively hard for robots precisely because it is ordinary: merchandise moves, customers return items to the wrong shelf, and a bag of candy hanging at a slightly different angle can defeat a grasp tuned on preset coordinates.

The orchestration layer drew as much attention as the robots themselves. AgiBot says the trial tested Genie Master, its embodied AI agent framework, which uses an Embodied Execution Protocol to link task planning, execution and feedback across the six machines, assigning work according to each robot's capabilities and the conditions in the store. Execution results flow back for analysis, and the framework can retain experience from successful and failed attempts in long-term memory for reuse on similar tasks - the difference, in the company's framing, between a demo that loops and a system that accumulates skill.

All of this comes with the usual asterisks. The trial was company-hosted, the numbers are self-reported, the two million viewers were watching a marketing event, and the release does not disclose labor economics - what the robots cost to run per hour, how many human interventions were needed overnight, or why the two failing robots fell short. The G2 is also wheeled rather than bipedal, which is precisely why it can survive a 24-hour shift, and a nod to how much of "humanoid robotics" in commercial service is really mobile manipulation on a stable platform.

Still, the shift in emphasis is notable. Chinese embodied-AI companies - AgiBot, Unitree, UBTech among them - have spent the past year moving from stage demos to staged real environments, and an operating convenience store with paying customers is a more demanding test than a showroom. The competitive question now is not whether a humanoid can pour 147 coffees in a day; it is whether any operator will pay for the privilege at scale, and whether the failure rate of two out of six over 24 hours improves fast enough to make the answer yes. AgiBot's next data point should be less cinematic and more useful: a multi-store deployment with independently observed numbers.