Six wheeled Agibot G2 robots just worked a full day inside a real Shanghai convenience store, greeting customers, brewing coffee, fetching products and wiping tables while more than two million people watched live. When the 24 hours ended, four earned a pass and two did not. That split result is more useful than the headline task count because it shows where humanoid retail work actually breaks down.
The trial, called “Robots Never Clock Out,” ran from 10 a.m. on November 2 to 10 a.m. on November 3 in an operating store in Shanghai’s Xuhui District. Organized by Chinese robotics company Agibot as the finale of its AI WAVE technology showcase, the live test showed what the Agibot G2 robots handled well, why two failed, and what a 24-hour shift actually proves about robots in stores.
What Actually Happened in Shanghai
Six wheeled Genie G2 humanoids, Agibot’s rolling-base service model, were assigned fixed store roles covering greeting, picking and shelf patrol, coffee making, checkout and hot food, and cleaning. The store stayed open to paying customers through four distinct periods: the lunchtime rush, the evening peak, overnight operations and the morning breakfast rush.
According to Agibot’s official trial summary, the team completed a combined 1,730 tasks. The same company-run assessment that counted those tasks also judged whether each robot met the store’s own standards for time on duty, task completion, customer feedback, response time and handling surprises.
That framing matters. Agibot both ran the trial and set the grading rubric, so the pass and fail labels describe performance against Agibot’s internal probation criteria, not an independent certification. The company has not published minute-by-minute human takeover logs, charging breaks or remote assistance records, so the event should be read as an endurance shakedown in a live store, not proof of unsupervised autonomy.
1,730 Tasks by the Numbers
The total sounds dramatic until you spread it across a full day. Across 24 hours, 1,730 completed tasks averages about 72 tasks per hour for the six-robot crew, or roughly 1.2 tasks per minute combined.
| Task type | Reported count | Share of total | Average pace |
|---|---|---|---|
| Customer-service interactions | 683 | 39.5% | 28.5 per hour |
| Product pickups and deliveries | 369 | 21.3% | 15.4 per hour |
| Shelf inspections and restocking | 277 | 16.0% | 11.5 per hour |
| Coffee orders | 147 | 8.5% | 6.1 per hour |
| Table-cleaning assignments | 128 | 7.4% | 5.3 per hour |
| Breakfast servings | 126 | 7.3% | 5.3 per hour |
That pace describes a quiet neighborhood store rather than a nonstop rush, which is an honest setting for a first 24-hour shift. The more telling detail is the mix. Nearly 40% of the work was talking to people, not moving objects, while food and drink preparation combined for less than 16%. For embodied AI, the hard part was not making one coffee correctly. It was switching between people, products, spills and shelf gaps without losing rhythm.
Meet the Six: Four Passes and Two Failures
Chinese automotive outlet Gasgoo, reporting from Agibot briefings and Chinese press interviews, gave the robots names and individual report cards that go beyond the English summary. Those details help explain the 4-2 split:
- Ahao, the greeter: passed with the top customer rating for steady, proactive welcomes.
- Huobao, the clerk: passed after handling 369 pickup and delivery orders through order surges and cancellations.
- Kasha, the barista: passed after making 147 coffees with consistent quality.
- Dingjiang, hot food manager: passed after serving 126 hot meals, though Agibot noted an occasional dropped corn cob in practice.
- Juanwang, shelf patrol and restocking: did not pass. Agibot said the robot showed little initiative and seldom acted without prompting.
- Lengjie, cleaning: did not pass after drifting from its planned route and positional stability checks.
The names come from Gasgoo’s detailed account of the trial and interviews with Agibot’s embodied intelligence team. While the underlying task counts match Agibot’s official breakdown, the assessments of initiative and route stability remain Agibot’s own internal judgments.
Why the Two Failures Matter More Than the Total
Cleaning and restocking sound simple compared with pulling espresso shots, yet they exposed two classic robot weaknesses: knowing when to act and knowing where to stay.
For the cleaning role, reporting citing Chinese tech outlet ITHome said the failure began when a full trash bin blocked the robot’s planned path. Instead of resolving the blockage or calling for help from its post, the robot deviated from its route, left the store area, and failed to return. For shelf patrol, the problem was the opposite of wandering: the robot waited too long, missed replenishment cues as products sold, and fell behind the pace of the store.
Agibot vice president Ren Guanghui framed both breakdowns as initiative and boundary problems rather than isolated grasp errors. Stores compress product variation, delicate handling, foot traffic, customer queries, and multi-robot coordination into tight aisles where tasks constantly collide. Restocking can block a pickup spot, cleaning can block an aisle, and a single delay quickly cascades into late deliveries.
That is a useful reality check for anyone following retail humanoid launches. Performing a task under supervision is not the same as sustaining it through interruptions, and the trial suggests the threshold for deployment sits closer to recovery behavior and work rhythm than to any single skill like making coffee.
The Secret Shoppers With Sugary Traps
Agibot also invited evaluators to deliberately make life difficult, a stronger test design than a scripted demo.
- Evaluators ordered regular and sugar-free Coca-Cola to check whether robots could distinguish nearly identical packaging.
- They rearranged packs of chewing gum and asked robots to return them to correct shelf positions.
- They blocked an aisle with a person and a box to test obstacle detection and rerouting.
Other examples shared with Chinese media show why small retail actions are technically dense. Picking up a golden sponge cake requires gripping tongs firmly, controlling opening angle and force so the cake neither slips nor crushes, aligning with a bag, withdrawing cleanly, then switching grips to lift the bag. Removing a hooked bag of candy requires sliding the packaging hole off a hook without tearing the wrapper or toppling neighbors, and once a soft package deforms, every subsequent action changes.

Those moments demand vision, motion policy, dexterous gripping and low-level force control working together. If any link slips, the object slips. That full-stack view helps explain why viral robot clips can mislead, because a perfect single grasp on camera does not guarantee the next grasp under slightly different light, placement or crowding.
The System Behind the Shift: Genie Master, GO-3 and Worken
Agibot used the store as an integration test for its software stack, not just its wheeled G2 hardware.
The crew shared an embodied agent framework called Genie Master. Through what Agibot calls the Embodied Execution Protocol, Genie Master assigns work by order priority, robot capability and workspace conflicts, runs tasks in parallel when they do not clash and staggers them when they need the same shelf space. Execution results feed back for troubleshooting, successful adjustments are kept in long-term memory, and effective fixes can become training data.

Three other pieces matter for context:
- GO-3 Preview: a task foundation model built on Agibot’s WRAM architecture for jointly modeling understanding, prediction, and action, designed to transfer experience across picking tasks rather than starting from scratch for each object.
- Genie Evolver: a hybrid post-training setup that combines real hardware feedback, physical simulation, and world-model prediction to refine manipulation skills.
- Worken: Agibot’s proposed labor metric for Effective Intelligent Work Hours, calculated as active robot count multiplied by autonomous work duration and output efficiency ratio. In plain terms, it measures how many qualified deliveries a robot completes per hour of supervised autonomy.
Agibot says livestream data, after filtering, de-identification, and quality checks, typically enters the training queue within one to two weeks. Same-day harness-level fixes can adjust execution quickly, while deeper policy updates iterate over several days. Above all of that sits a strict safety rule paraphrased from the engineering team: better slow than sorry, and never collide. When aisles get crowded, the robots slow down, pause, yield, or replan their path.
The company’s longer endurance work puts this in perspective. The same manufacturer recently completed a 66-mile supervised urban walk with its larger bipedal Agibot A2, a test of stamina outdoors rather than fine manipulation indoors.
What This Trial Does Not Prove
Three cautions should stay attached to any retelling:
- Company-run grading: Four passes and two failures reflect Agibot’s internal probation standards, including customer feedback and response metrics Agibot chose. Independent auditors did not publish separate scores.
- Unclear assistance level: Agibot tracked human takeover frequency, anomaly recovery time and system availability, but has not released those numbers. Without them, viewers cannot tell how often engineers stepped in, how charging was handled or how much remote operation supported the livestream.
- Task counts are not success rates: The 1,730 figure counts completed tasks. It does not reveal first-attempt success rate, retry count or how many requests were declined or rerouted to staff.
Those gaps do not make the trial meaningless. A full day in a working store across lunch, evening, overnight and breakfast shifts exposes robots to changing light, crowds, spills and stockouts that a lab never reproduces. But they do mean the event validates persistence and data collection more than independent commercial readiness.
What Comes Next for Store Robots
Agibot says it will expand testing to more roles and environments, using operational feedback to improve reliability and transfer skills across stores. Ren’s own deployment filter is refreshingly cautious: clear task boundaries, easily verifiable results and an anomaly rate low enough for the system to self-recover.
By that standard, structured and verifiable jobs like sorting, restocking assistance, and basic cleaning are likely to stabilize first, while open-ended customer interaction, identity checks, and delicate food handling will take longer. Continuous-operation designs are already emerging elsewhere, including a 24/7 humanoid worker that swaps its own batteries. Industry shipment numbers explain why robotics companies are pushing into real-world trials anyway. A Counterpoint breakdown cited by Gasgoo put global humanoid shipments above 22,000 units in the first half of 2026, up nearly 300% year on year, with Chinese firms taking the top five spots and most units still going to entertainment, education, data collection, and greetings rather than factory or warehouse floors.
Six Genie G2s finishing with four passes and two failures is not a flawless scorecard. Still, it is far more informative than a polished highlight reel, offering a clear look at the gap between a robot that can perform a demo and one that can hold down a shift.
