CES® 2027 NEWS
ROAD TO CES 2027Live now
ROAD TO CES · AN INTELITOYS EVENTJAN 6–9 · VEGAS
Home Robots

AGIBOT's G2 Humanoids Ran a Shanghai Convenience Store for 24 Hours and Logged 1,730 Tasks

Six wheeled humanoids covered customer service, coffee, restocking and cleaning in a live retail trial that has now ended.. AGIBOT put six G2 wheeled humanoid robots through a 24-hour trial inside an operating convenience store in Shanghai's Xuhui District, completing 1,730 tasks.

InteliToys Editors ? AGIBOT G2
AGIBOT wheeled humanoid robot, studio product shot (image: AGIBOT)

AGIBOT put six G2 wheeled humanoid robots to work inside a convenience store in Shanghai's Xuhui District for a full day, and the fleet finished with 1,730 tasks completed. The run began at 10 a.m. And wrapped at 10 a.m. The following morning, with more than two million people watching a livestream of the shift. When the scores were tallied, four of the six machines met every criterion in the final assessment; the other two fell short.

The trial served as the finale of AGIBOT's AI WAVE showcase, a program the company kicked off. Earlier installments of that program touched on robotics data, simulation, AI models, hardware and operational software one at a time. This event, the company says, was the first time all of those pieces were combined in a single location and left to handle a normal retail day.

Responsibility was divided among the six units. One handled customer assistance, another checked shelves, a third restocked them, and others covered coffee preparation, checkout and hot food service, and cleaning. Across the shift the group recorded 683 customer-service interactions, 369 product pickups and deliveries, 147 coffee orders, 126 breakfast servings, 277 shelf inspections and restocking jobs, and 128 table cleanings. The clock was set up to hit four distinct periods: the lunch rush, the evening peak, the overnight stretch, and the next morning's breakfast crowd.

Ride the Las Vegas Monorail to CES - stops at the Las Vegas Convention Center

Scoring covered time on duty, task completion, customer feedback, response times and reactions to surprises. Evaluators also staged scenarios drawn from real shop work. They asked for regular and sugar-free Coca-Cola to see if the machines could separate look-alike items, moved packs of chewing gum around a shelf and told the robots to restore order, and put a person plus a box in an aisle to test whether the units could spot the blockage and route around it.

AGIBOT's argument is that a corner shop is a tougher proving ground than it appears. The floor plan is small and plenty of jobs follow a routine, yet a candy bag tilted a few degrees can be hard to pick up without knocking neighbors loose, and a slice of cake calls for precise grip control. A store also refuses to hold still: shoppers wander the aisles, goods get moved and returned, and packaging, position and orientation shift constantly. In the trial, the G2s used live visual input to find misplaced items and put them back on their assigned shelf rather than following fixed coordinates, one restocker noticed empty shelf space and filled it without being told, and a cleaning unit told spills apart from countertop patterns before choosing a tool.

Genie Master, AGIBOT's embodied AI agent framework, was the other thing under test. The company describes it as a shared layer that coordinates the fleet and ties each action back to later review. Its Embodied Execution Protocol connects planning, execution and feedback: assignments are matched to each robot's abilities and the state of the shop, the units act on policies and live observations, and results flow back for analysis and troubleshooting. The framework keeps useful experience from both successful and failed attempts in long-term memory for future situations, and AGIBOT says good adjustments can be turned into training data to sharpen the robots' action policies. The company says it has already been using operational feedback to improve G2 handling of delicate manipulation, with added practice during the quieter overnight hours.

What remains unclear is how much of this carries beyond one shop in one district. AGIBOT says it intends to keep testing in more roles and more real environments, using what it learns to raise reliability, task execution and adaptability, and that it will weigh which skills and operating practices can move between tasks and settings. For readers following embodied AI, the more telling detail is the format of the test: a complete working day, a public livestream, and a pass count published rather than a staged demo video.

Source: AGIBOT.

InteliToys is editorial. We do not sell AGIBOT G2. Watch the source tape in the branded lightbox below.