During a 24-hour shift in a Shanghai convenience store, two evaluators ordered a regular Coca-Cola and a sugar-free one to see whether a robot could tell the cans apart. They also shuffled packs of chewing gum on a shelf and parked a person and a box in an aisle. AGIBOT’s account of its six-robot trial describes those tests in detail and doesn’t say how often the robots got them wrong. Its headline figure has the same gap: 1,730 tasks completed, with no count of the misses, retries or human help behind them.
What AGIBOT ran and what it published
Six AGIBOT G2 wheeled humanoids worked an operating store in Shanghai’s Xuhui District from 10 a.m. on November 2 to 10 a.m. the next day, livestreamed to more than two million viewers. Each robot had one job: customer help, shelf inspection, restocking, coffee, checkout and hot food, or cleaning. The write-up ran on The Robot Report on November 13, labeled as content sponsored by AGIBOT. It breaks the total down into 683 customer-service interactions, 369 product pickups and deliveries, 277 shelf inspections and restocks, 147 coffees, 128 table cleanings and 126 breakfast servings.
The robots were graded on time on duty, task completion, customer feedback, response time and how they handled the unexpected. Four passed. Two missed at least one criterion. AGIBOT says it used the quieter overnight hours to give the robots extra practice on precise manipulation.
The setting earns AGIBOT some credit. A store with customers moving through the aisles is a harder test than the spotless lab most robot videos come from, and a company that publishes a four-of-six pass rate has told you more than a highlight reel would.
Four numbers a robot field trial report should carry
A buyer pricing a robot needs more than a completion count. Four other numbers decide what those 1,730 tasks are worth to a buyer.
The first is attempts. The account describes the hard cases well: a bag of candy hanging at an odd angle, a slice of cake that needs a gentle grip. It says a robot may fail to grasp an item or need to adjust its movements. It doesn’t count those misses, or report how the robots did on the Coke and gum tests. A coffee order that took three tries counts once.
Then there are interventions. The published account doesn’t say whether store staff stepped in, how often, or whether anyone operated a robot remotely. In a real deployment, the hours a person spends rescuing a robot are part of its labor cost, and a trial that doesn’t log them can’t price the robot against the worker it would replace.
The pass bar comes third. AGIBOT lists what the robots were graded on but not the score each had to hit, or which two fell short. Two misses at the coffee and cleaning stations would be a different finding from two in customer service, the busiest station.
Last is the day itself. The trial ran during a livestream with an audience above two million, and the report doesn’t say how many customers came through compared with an ordinary day at that store. Without that, nobody can tell whether 683 customer-service interactions was a heavy load or a light one.
The same four numbers belong in a founder’s pilot report
When I read a trial, I’m looking for whether it was instrumented for the press release or for the purchase order. One built for the purchase order logs every attempt and every human touch, and it fixes the pass bar before the shift starts. That logging costs more to set up, and a buyer can check the result.
Those logs matter most to whoever pays for the misses, and that depends on who owns the robot. Tutor Intelligence, which sells robot hours, keeps the robot on its own books. A buyer who owns a robot outright pays for every retry and every rescue, and should see the attempt and intervention logs before signing.
Dave’s take
Robot demos have gotten more polished faster than the robots have matured, and a live store with a published pass rate is better evidence than a studio clip. A company that can run six robots in public for 24 hours can count their misses too. Publishing that count beside the 1,730 would make the trial useful to the people who buy robots.
From Dave’s video library
Dave on testing your riskiest assumption before building the product, and asking customers for a costly yes instead of a compliment.
I’m here to help you scale.
Work With DavePrefer a smaller first step? Book a $500 one-hour working session →
Dave Saunders is the founder of Base Reality Group and a Fractional CPO for product companies. He was a founder and operator at Galen Robotics, where the surgical-robotics platform earned FDA De Novo authorization in 2023, and he managed a 35-patent portfolio licensed from Johns Hopkins. He wrote Founders Who Finish and publishes The Build. More about Dave →