Six AGIBOT G2 robots run a Shanghai convenience store for 24 hours
AGIBOT says six wheeled G2 humanoids completed 1,730 tasks over 24 hours in a working store, and that four of the six passed its final assessment. The account comes from the company.
AGIBOT ran six of its G2 wheeled humanoid robots in a convenience store that stayed open for business in the Xuhui District of Shanghai from 10:00 on 2 November to 10:00 on 3 November 2026. The account, published by The Robot Report on 13 November, reads as a company release. It does not give the G2’s degrees of freedom, payload or reach, and it does not say whether any task was teleoperated.
Each robot had one role: customer assistance, shelf inspection, restocking, coffee, checkout and hot food, or cleaning. According to AGIBOT, the six completed 1,730 tasks between them. The total breaks down into 683 customer-service exchanges, 277 shelf checks or restocks, 369 items picked up or delivered, 147 coffees, 126 breakfasts and 128 tables cleaned. Each robot was scored on task completion, hours on duty, response time, customer feedback and how it coped with surprises. Four of the six passed and two did not meet every criterion.
Two evaluators set tests during the run. They asked for both sugar-free and regular Coca-Cola to see whether the robots could tell the cans apart, moved chewing gum packs to the wrong place on a shelf, and put a person and a box in an aisle as obstacles. AGIBOT’s Genie Master agent framework assigned work, collected execution results and, the company says, turned useful corrections into training data.
Why it matters
A 24-hour run with a pass mark is a better test than a short demo, and AGIBOT published the failure count as well as the task total. Two of six robots failing their assessment is the number to remember. The trial also shows where the hard parts sit in retail: grasping items that hang at odd angles and handling soft food. What is missing is a breakdown of failures by task, the intervention rate and any teleoperation share, which an integrator would need before pricing a deployment.