AGIBOT Deploys Six Genie G2 Robots for 24-Hour Trial at Shanghai Convenience Store

AGIBOT Innovation (Shanghai) Technology Co., Ltd. deployed six Genie G2 humanoid robots at 10 a.m. on Nov. 2 at an operating convenience store in Shanghai’s Xuhui district and began a 24-hour operational trial. The six robots divided tasks including greeting customers, restocking products, serving coffee, handling checkout and snacks, and cleaning while the store remained open.

AGIBOT described 24-hour store operations by six Genie G2 robots and a system in which Genie Master records task results and incorporates them into training data. In operations that assign multiple tasks to robots in a real store, challenges include not only completing individual tasks but also responding to anomalies and reusing experience. Final results and the extent of human intervention will therefore be important in future evaluations.

The focus of this trial is not the Genie G2’s individual movements, but the deployment of six robots in an operating store as a work system that includes task allocation and continuous operation. Running multiple tasks, such as greeting customers, restocking and cleaning, in the same store puts more than object-grasping ability to the test; it also tests detection of out-of-stock items, task initiation, and responses to pedestrians and obstacles. AGIBOT describes Genie Master’s recording and learning functions as a system that uses on-site failures in subsequent tasks rather than limiting them to retries in the moment. This is a move to connect the technologies the company presented at AI WAVE with work in a real store, advancing from announcements of individual functions to testing that includes operational conditions. In an Oct. 28 article, AGIBOT Unveils Embodied Agent Framework “Genie Master,” Tests Autonomous Improvement in Real Stores, the company outlined a framework connecting execution with feedback on results. The current store trial puts that framework in an environment with actual business hours, changing product arrangements and customer traffic. Compared with the simulation and evaluation process covered in another Oct. 28 article, AGIBOT Updates “Genie Sim” Simulation Platform to Support Deformable-Object Simulation, responding to unpredictable layouts and obstacles in a real store is a separate point of evaluation. The operational structure describes a cycle connecting on-site observation, task decisions, robot actions, recording of results and policy improvement using those records. The design of running all six robots with a shared Agent also appears to be an attempt to handle multiple assigned tasks within the same control and learning framework. However, the article mainly describes specific work scenarios and does not provide quantitative figures for success rates over 24 hours, number of interventions, operating time for each robot or workload for store staff. An article dated Nov. 2, AGIBOT, Tsinghua University and Others Announce CEL, Test Collective Learning With 96 Robots, covers sharing experience across a group, while the current article describes Genie Master’s execution and memory system within a store. Both involve reusing experience, but the articles do not establish that they are connected as the same system. The next questions are what metrics will be used to evaluate store tasks and how failures and human interventions were recorded. AGIBOT said it would release final results later. Whether it reports figures not only for continuous operation but also for task completion and improvements after retraining will help determine whether the system can be deployed at other stores and for other tasks.

According to the company, the trial included visually checking products and returning them to shelves, detecting out-of-stock items and beginning to restock, and recognizing obstacles before stopping or avoiding them. The article describes work in a real store, where lighting and the positions and orientations of products change, as a setting for evaluating the robots’ capabilities. The shared Agent, “Genie Master,” used by all six robots connects task decisions, actions and feedback on results through the Embodied Execution Protocol (EEP), the company said. According to AGIBOT, it compiles successful and failed task trajectories daily and stores them in long-term memory, then uses the adjustment process as training data to improve action policies. At the time the article was published, final results were to be released later.