Four of AGIBOT's Six G2 Humanoids Passed a 24-Hour Store Shift

AGIBOT put six G2 humanoid robots to work in an operating Shanghai convenience store for 24 hours from 10 a.m. on Nov. 2, logged 1,730 tasks, and said four of the six passed its assessment. The other two did not, and the company has not published a failure rate or a count of human interventions.

The tally comes from AGIBOT itself. The Robot Report ran it on Nov. 13 labelled "Sponsored content by AGIBOT." Gasgoo, an automotive trade outlet, filled in the detail the sponsored piece leaves out.

What the 1,730 tasks were

The store sits in Shanghai's Xuhui District; neither outlet names it or its operator. The livestream drew more than two million viewers. The breakdown, per The Robot Report:

Task type	Count
Customer-service interactions	683
Product pickups and deliveries	369
Shelf inspection and restocking	277
Coffee orders	147
Table cleaning	128
Breakfast servings	126

1. Customer service683 tasks
2. Pickups and deliveries369 tasks
3. Shelf inspection and restocking277 tasks
4. Coffee orders147 tasks
5. Table cleaning128 tasks
6. Breakfast servings126 tasks

Source: AGIBOT figures as published by The Robot Report (sponsored content), accessed 2026-11-14

The number that matters is the spread across robots. Gasgoo gives per-robot counts for five of them: Huobao, a clerk, handled 369 pickup-and-delivery orders; Kasha, a barista, made 147 coffees; Dingjiang made 126 hot meals; Juanwang patrolled and restocked 277 times; Lengjie cleared tables 128 times. Those five sum to 1,047. The remaining 683 match the customer-service line, and the sixth robot, Ahao, is the greeter. One robot at the door accounts for 39 percent of the headline count (our arithmetic), and it is a task that needs no manipulation.

Why two robots failed

Per Gasgoo, Juanwang, the patrol and restocking robot, "showed little initiative" and did not keep up with the store's pace. Lengjie, the table-clearing robot, drifted off its route and left the store, which Gasgoo describes as weak positional stability. Dingjiang passed but occasionally dropped a corn cob. Early ordering software had failed orders, missed status updates and positional conflicts between robots.

Ren Guanghui, vice president of AGIBOT's Embodied Intelligence Research Center, said, as Gasgoo reports it, that robots still lag a seasoned clerk on adapting to on-site change, fine manipulation rhythm and anomalies, and that deformable objects and continuous multi-step operations are weak spots. He also said that if adaptation and maintenance always depend on engineers, replication costs will not fall.

What neither report says

AGIBOT tracked "human takeover frequency" but published no figure, per Gasgoo. The Robot Report does not say whether the robots were teleoperated or fully autonomous. Price, runtime between charges and the number of staff on site are absent from both. A 24-hour livestream does not establish that each robot worked continuously without charging.

The fleet behind the demo

AGIBOT said in September 2026 that its 20,000th robot had come off the production line. Gasgoo cites a Counterpoint report that global humanoid shipments passed 22,000 units in the first half of 2026, up nearly 300 percent from a year earlier, with AGIBOT at about 9,700 units (over 43 percent) and Unitree above 7,000 (about 31 percent). Entertainment, performance, education and data collection made up over 60 percent of deployments; manufacturing took 13 percent and logistics 5 percent.

That mix is the context for a store trial. Retail does not appear in the breakdown Gasgoo cites. For comparison, see our reports on Agility Robotics' stock debut and order book and OpenAI disbanding its robotics team.

What would change my read

A published takeover count per robot-hour would settle most of this. If AGIBOT releases the task log and the number is low, 72 tasks an hour across six robots (our arithmetic, 12 per robot-hour) becomes a real throughput figure. If it stays unpublished, the trial remains a demonstration.