Six AgiBot Genie G2 Units Staff a Convenience Store for 24 Hours, Completing 1,730 Tasks

From the concentrated debut of technical solutions to stepping into real stores for duty, the results of AgiBot's AI WAVE series face a real-world test. On November 2, six Genie G2 humanoid robots entered a normally operating convenience store in Xuhui, Shanghai, launching a 24-hour "probation test," with the entire process watched and supervised online by netizens. On November 3, this livestreamed challenge concluded. In an around-the-clock real commercial environment, the robots directly confronted fluctuations in customer flow and unexpected disruptions, completing continuous operations across multiple product categories and multiple positions, accumulating valuable frontline practical samples for embodied intelligence to land in the retail service sector and for exploring large-scale application of humanoid robots. Since the AI WAVE plan was launched on October 22, AgiBot has successively released a series of technical results including datasets, a simulation platform, embodied models, robot bodies, and an operating system, and this convenience store real-world operation integrates all the technologies, operating continuously and stably in a real environment with constantly changing customer flow. For embodied intelligence to achieve large-scale deployment, it is precisely rooted in repeated trials and continuous accumulation at each individual position. The six robots each have their own division of labor: one stationed at the entrance to receive customers and answer inquiries; two partnered to handle shelf restocking and inspection; one dedicated to making coffee; one stationed at the checkout counter, also handling prepared food service; and one responsible for cleaning tables in the store. The six positions cooperate with one another, covering the complete service chain of convenience store reception, product management, preparation, and cleaning. This 24-hour assessment advanced in step with the store's customer flow rhythm, with the lunch peak, evening customer flow, late-night duty, and the next morning's peak all being tests. Task completion rate, response speed, exception handling capability, and customer feedback are all indicators for judging whether they can smoothly "become regular employees." The convenience store scenario looks standardized in its processes, but poses numerous challenges for robots. Hanging a bag of Skittles on a hook—if the angle is slightly off, it may knock down other products; grabbing golden cake—if force control is slightly careless, the product will slip or break. In a laboratory environment, lighting and item positions are controllable, but in a real store, customers walk through at any time, and the angles at which products are placed and their packaging states change at any moment; the same pick-up action faces a different environment each time, and a single successful grasp does not mean it can be stably reproduced. In actual operation, the robots performed stably overall: products misplaced by customers were autonomously returned to their proper places relying on visual inventory; shelf vacancies could be autonomously identified and restocking initiated; they could distinguish stains from table patterns and start cleaning as needed. Of course, during their duty the robots did not succeed every time and occasionally made operational mistakes. But in real scenarios, a mistake does not equal failure; both humans and robots make errors, and the core lies in whether problems can be identified and lessons learned. The core criterion of this assessment is not the pursuit of zero errors, but the possession of self-correction capability. This is also the design philosophy of AgiBot's Genie Master embodied agent framework: it is the scheduling brain shared by the six robots, coordinating task allocation, collecting operation results, and supporting continuous learning. Genie Master connects the closed loop of task judgment, action execution, and result feedback. The agent allocates tasks based on on-site conditions, the robots adjust actions according to real-time perception, and execution results are sent back to the system, providing a basis for subsequent handling of exceptions. In addition to immediate correction, the system also aggregates all successful and failed operation records, precipitating them into long-term memory so that similar tasks can reuse past experience; effective adjustment strategies are converted into training data to feed back into optimization of the action models. The robots accumulate skills while on duty, and during idle periods they also practice on their own. This learning mechanism has already begun to show results, with Genie G2 continuously optimizing on fine manipulation tasks. Throughout the test, the six Genie G2 robots completed a total of 1,730 tasks, including 683 customer receptions, 369 product pickups and deliveries, 147 coffee preparations, 126 breakfast services, 277 restocking inspections, and 128 table cleanings; four successfully "became regular employees," and the livestream attracted more than 2 million netizens to watch online cumulatively. The convenience store is not the endpoint. Next, AgiBot will continue to break down the barriers between technology and real scenarios, validate and accumulate capabilities in more positions, replicate mature solutions outward, and drive embodied intelligence from the laboratory into more offline industry scenarios.