EN
Search
Photo: The Robot Report

AGIBOT's G2 robots complete 1,730 tasks in 24-hour Shanghai convenience store trial

A 24-hour trial in a working store

AGIBOT ran a 24-hour trial of six of its G2 wheeled humanoid robots at an operating convenience store in Xuhui District in Shanghai, according to The Robot Report. The robots completed a combined 1,730 tasks, from customer help and coffee preparation to restocking shelves and cleaning. Four of the six passed a final performance assessment.

The test started at 10 a.m. on November 2 and finished at 10 a.m. on November 3. The report said it was livestreamed to more than two million viewers on online platforms.

The trial closed AGIBOT's AI WAVE technology showcase, which began on October 22. That showcase released datasets for robotics, simulation tools, AI models, hardware and software for operations. The store test was meant to bring these technologies together in a real setting and show how robots handle everyday work over a longer period. Separately, the company said in September 2026 that its 20,000th robot had come off the production line.

Assigned roles and test results

Each robot had a particular role. Some handled customer requests, some checked and refilled shelves, and others prepared coffee, handled checkout and hot food service, or cleaned the store.

Over the 24-hour period, the robots logged 683 customer-service interactions and 369 product pickups and deliveries. They also prepared 147 coffee orders and 126 breakfast servings, handled 277 shelf inspections and restocking tasks, and cleaned tables 128 times.

The trial covered four periods: the lunchtime rush, evening peak hours, overnight operations and the next morning's breakfast rush. The report said each period posed different challenges, from answering requests at busy times to keeping the store running overnight.

Evaluators scored each robot on how long it stayed on duty, whether it finished tasks, customer feedback, speed of response and how it dealt with surprises. Four robots cleared the final review, while the other two fell short of some requirements, according to the report.

Two evaluators also ran practical tests based on situations robots might face when serving customers. In one, they ordered regular and sugar-free Coca-Cola to see whether the robots could tell similar-looking products apart. In another, they moved packs of chewing gum out of place and asked the robots to put them back where they belonged.

To test navigation, the evaluators put a person and a box in one aisle, checking whether the machines could spot the obstacles and route around them.

Why a store is a demanding test

The report notes that a convenience store can look simple for robots, with a compact layout and many tasks that follow standard procedures. Even so, grabbing goods, refilling shelves and making food still force machines to cope with small, frequent shifts in conditions.

The report gives examples. Candy bags that hang at a slightly different angle are hard to grip without moving neighboring products. A delicate item such as a slice of cake needs careful control of grip force so it is not dropped or damaged.

According to the report, a live store differs sharply from a lab. Customers walk the aisles, goods are picked up and put back, and items vary in position, angle and packaging, so even a repeated task may not go the same way twice.

The report says the G2 robots showed several abilities meant for these problems. They used live visual input to spot items out of place and return them to their designated slots, instead of depending only on preset coordinates.

One robot on restocking duty noticed gaps on the shelves and filled them without being told to. In the cleaning area, another separated spills from patterns printed on the countertop and picked suitable tools.

The report says these tasks went beyond single movements. They called for perception of the surroundings, task-specific decisions and changes to actions as conditions shifted.

Genie Master and learning from operations

The trial also looked at AGIBOT's Genie Master, an embodied AI agent framework. According to the report, it was tested for coordinating tasks, fixing errors and learning from operations.

Genie Master provides a common framework that coordinates all six robots and links their actions with later analysis and learning. The report says the link between planning, execution and feedback runs through the Embodied Execution Protocol (EEP).

The system sorts work by each robot's abilities and the store's conditions. The robots then carry out their jobs based on action policies and what they see in real time. Results from each attempt flow back into the system for review and fault diagnosis.

The framework also learns from earlier attempts. By reviewing successful and failed executions, it can keep useful experience in long-term memory for similar cases. Good adjustments can be turned into training data that refines the action policies.

Next steps

AGIBOT has also been applying what it learns from operations to sharpen G2 results on delicate handling jobs, with added practice during calmer nights. The report says these efforts were aimed at tasks that need precise handling.

The company intends to extend trials to further jobs and real-world sites, using feedback from operations to make the machines more dependable and flexible. According to the report, AGIBOT will also examine which abilities and working methods can carry over between jobs and be copied to other sites. The report suggests these results will shape the next phase, aimed at steadier and more scalable use in the field.

Source