Six humanoid robots work 24-hour shift in Shanghai store

The 24-hour retail trial

Chinese robotics company AGIBOT recently conducted an experiment titled "Robots Never Clock Out" to test the capabilities of humanoid machines in a real-world setting. The company deployed six G2 wheeled humanoid robots into an operating convenience store in Shanghai's Xuhui district. These machines were tasked with performing a variety of retail duties over a continuous 24-hour period starting on November 2.

The robots were assigned specific responsibilities to cover the full scope of store operations. They were tasked with greeting customers, preparing coffee, serving hot food, and restocking shelves. The trial was designed to move beyond controlled demonstrations by exposing the machines to varying levels of customer activity and changing store conditions.

To track the progress of the experiment, AGIBOT livestreamed the challenge across multiple online platforms. The event attracted more than two million viewers who watched the robots navigate the store. The machines were required to work through four distinct periods, including the lunchtime rush, evening peak hours, overnight operations, and the following morning's breakfast rush.

Task completion and performance

During the 24-hour shift, the six robots collectively completed 1,730 tasks. These activities included 683 customer-service interactions and 369 product pickups and deliveries. The robots also handled 147 coffee orders, 126 breakfast servings, 277 shelf-inspection and restocking tasks, and 128 table-cleaning assignments.

AGIBOT evaluated the robots based on several specific criteria to determine their success. The company looked at the total time on duty, the number of tasks completed, and the feedback received from customers. They also measured response times and the machines' ability to handle unexpected situations that arose during the shift.

At the end of the assessment, four of the robots passed the final evaluation. The remaining two robots were placed in an observation group for further improvement. These two machines were specifically assigned to cleaning and shelf management, which proved to be the most difficult responsibilities for the robots to master consistently.

Failures in unpredictable settings

The two robots that failed the assessment struggled with tasks that required navigating unpredictable obstacles. One cleaning robot encountered a full trash bin that blocked its planned route. Instead of finding a way around the bin or resolving the issue, the robot deviated from its path, left the store, and failed to return to its position.

The robot responsible for shelf inspections and restocking also failed to meet the company's performance criteria. It reportedly lacked the initiative to identify and replenish products as they were sold. This highlighted a major distinction in robotics: performing a task under predictable conditions is much easier than completing it reliably in a changing environment.

The experiment also included deliberately difficult tests to push the robots' limits. Evaluators asked the machines to distinguish between regular and sugar-free Coca-Cola and had them rearrange chewing gum on shelves. They also blocked an aisle with a person and a box to test the robots' obstacle detection and navigation capabilities.

The role of embodied AI

The trial served as a testbed for AGIBOT’s Genie Master embodied AI agent framework. This system connects task planning, execution, and feedback across the robots to coordinate their work. It uses an Embodied Execution Protocol to manage assigned tasks and collect results for analysis by the company's engineers.

A key feature of this framework is its ability to retain both successful and unsuccessful experiences. AGIBOT says the system can use this data to troubleshoot problems and convert effective adjustments into training data. This process is intended to improve future robot behavior by teaching the machines how to handle complex physical manipulations.

For a humanoid robot to operate outside of a laboratory, it must combine perception, decision-making, and physical movement. A machine must be able to recognize an object and then determine how to grasp and move it without damaging nearby items. The Shanghai experiment provided a useful look at how these systems handle the nuances of retail work.

What comes next for AGIBOT

The results of the "Robots Never Clock Out" challenge suggest that humanoid machines are becoming increasingly capable of handling specific retail duties. However, the failure of two robots shows that managing an entire shift reliably remains a significant challenge. The data gathered from the 1,730 tasks will be used to identify specific gaps in autonomy and adaptability.

AGIBOT plans to use the operational feedback from this trial to improve the performance of the G2 models. The company intends to refine the robots' ability to recognize products and navigate obstacles more effectively. They aim to make the machines more capable of handling the