Six AGIBOT G2 robots conduct 24-hour trial at Shanghai convenience store, completing 1,730 tasks

AGIBOT Innovation (Shanghai) Technology Co., Ltd. conducted a 24-hour trial with six wheeled humanoid AGIBOT G2 robots at an operating convenience store in Shanghai’s Xuhui district. The trial ran from 10 a.m. Nov. 2 to 10 a.m. Nov. 3. The six robots completed a total of 1,730 tasks, and four passed the final evaluation.

AGIBOT’s trial combined customer service, product restocking, food and drink service, and cleaning at one store, and evaluated 24-hour operation and responses to unexpected situations. A key question is whether the feedback mechanism in Genie Master, as described by AGIBOT, can use task results to improve action policies, helping extend individual task demonstrations to continuous store operations.

The total of 1,730 tasks completed by six robots alone does not measure the trial’s success. The breakdown includes 683 customer interactions, as well as product handoffs, food and drink preparation, shelf restocking and cleaning. Combining customer service and object handling in a single store operation tests the ability to switch among different types of tasks and sustain them, rather than perform a single-action demonstration. However, the trial was limited to 24 hours at one store, and four of the six robots passed. The results do not show that all robots performed at the same level or that the same results would be achieved at other stores. Including operating time, task completion, customer responses, response time and handling of unexpected situations in the evaluation shows that continuity of store operations, not just movement accuracy, was assessed. However, the passing criteria for each item and the performance of individual robots were not disclosed. The figure of four passing robots should not be taken as a direct measure of readiness for practical deployment; it is necessary to distinguish which tasks or time periods accounted for differences in the evaluation. Technically, Genie Master and EEP are described as a design that connects planning, execution, results analysis and the reuse of experience. AGIBOT says execution results, including failures, can be used to improve subsequent action policies. The trial’s significance therefore lies not only in the number of tasks but also in testing at a real store a process for feeding operational feedback back into learning. However, AGIBOT’s account does not show which results were adopted as training data or how success rates on the same tasks changed after improvements. In a report dated Oct. 29, 2026, AGIBOT introduces ‘Worken,’ a system for measuring robot tasks, we could not confirm details of the measurement method or the specifications of the target robots. The current trial provides task counts and evaluation items, but does not disclose the passing criteria. To connect a framework for measuring work with evaluation in real-world settings, the next points to check are the reproducibility of the criteria, performance when moving to multiple stores and different tasks, and the improvements that experience from the trial produces after retraining. AGIBOT also plans further trials in different roles and real-world settings, as well as evaluations of whether capabilities and operating methods can be transferred to other tasks or environments.

Tasks included customer service, shelf inspection and restocking, coffee preparation, checkout and hot-food service, and cleaning. The breakdown was 683 customer interactions, 369 product pickups and deliveries, 147 coffee orders, 126 breakfast servings, 277 shelf inspections and restocking tasks, and 128 table-cleaning tasks. The trial covered lunchtime, evening, nighttime and the following morning’s breakfast period. The evaluation examined operating time, task completion, customer responses, response time and handling of unexpected situations. According to AGIBOT, the six robots’ work was coordinated using Genie Master, an embodied AI agent platform, while the Embodied Execution Protocol (EEP) connected task planning, execution and feedback. The trial also tested a system that routes execution results to analysis and troubleshooting, stores experiences from successful and failed tasks in long-term memory, and converts effective coordination into training data to improve action policies.