AgiBot G2 robots run a Shanghai store for 24 hours, 1,730 tasks

What did the AgiBot G2 robots actually do?

The numbers are specific. There were 683 customer service interactions and 369 product pickup and handover jobs. The robots made 147 coffees and 126 breakfast orders. They also carried out 277 shelf checking and restocking tasks and 128 table-cleaning jobs.

The day was split into four shifts: the lunch rush, the evening peak, the overnight stretch and the next morning's breakfast rush. Each asked for something different, from busy customer service to simply keeping the shop running while it was quiet.

How were the robots judged?

Evaluators scored them on working time, task completion, customer feedback, response speed and how they handled unexpected situations. Four of the six passed the final assessment. The other two missed some of the required criteria.

The tests were designed to trip them up. Evaluators ordered a regular soda and a sugar-free one to see whether the robots could tell similar products apart. They shuffled packs of chewing gum on a shelf and asked for them to be put back in the right places. They also blocked an aisle with a person and a box to see whether the robots would spot the obstacle and get around it.

A one-day trial is not a long-term deployment, and the report gives no figures on error rates or costs. Still, a 24-hour shift with a live audience is a harder test than a polished demo video.