China Service Robot Pilots: Measure Human Minutes per Order

A service robot can take longer than a person to finish an order and still reduce some hands-on work. It can also move that work into refilling, cleaning, supervision and recovery. A useful pilot measures the human work around the complete service, alongside the time the customer waits.

For a retail or hospitality operator, the practical question is how much usable staff time changes when the robot joins the workflow. Start with human minutes per correctly completed order, then check whether any released time can actually be reassigned.

What the Sharpa report can tell a buyer

In a 21 October 2026 Jazzyear report, Sharpa describes a robot making ice cream at a DQ shop on Shanghai's Wujiang Road. The account says it uses the shop's existing equipment and operating procedures, and currently makes a cup at roughly half a human worker's speed. That is an attributed comparison in a reported deployment; it does not establish the site's total labour cost or staffing requirement.

The report also describes a plan to deploy 30,000 robots across restaurants, retail and hotels. That planned volume is separate from completed installations. China AI Wire's related coverage preserves that distinction and notes that prices and delivery dates for the new hardware were not announced.

The useful next question is therefore operational: which work does the robot complete, which work stays with people, and what happens when service cannot proceed normally?

Define a completed order before timing the pilot

Choose a small set of products and the conditions under which an order counts as correctly fulfilled. Keep the order mix, demand period and quality requirements comparable between the current service and the robot-assisted trial. Record preparation and close-down work even when it happens outside the customer-facing test period.

Use the customer's completed order as the denominator. A restarted robot cycle is another attempt at the same order, not another customer served. Keep abandoned orders, wrong items, remakes and customer waiting times in separate columns so a lower labour figure cannot conceal a weaker service.

For each human task, record the person, start and end time, purpose and order or batch affected. Include remote assistance. Two people working for five minutes use ten person-minutes. If one person handles overlapping tasks, allocate that person's time once rather than counting the same minute twice.

Follow the work outside the visible robot cycle

Work to observe	What to record	What the record clarifies
Preparation and close-down	Loading, start-up checks, cleaning and shutdown time	Whether the visible cycle omits recurring work
Consumables and stock	Refill visits, stock checks and interruptions	How much human work keeps the station available
Exceptions and recovery	Failed attempts, diagnosis, reset assistance and remakes	Whether labour moves into intermittent interventions
Customer handoff	Collection, questions, wrong orders and refunds	Who completes the service after the robot stops moving
Monitoring and remote help	Active attention, support sessions and required dedicated coverage	Whether active work and reserved staffing tell different stories

Count preparation or cleaning shared by a batch once, then allocate it across that batch's completed orders. Include human work spent on unsuccessful attempts in the numerator, even though those attempts add no completed orders to the denominator.

Keep active work and dedicated coverage as separate measures. A worker required to remain beside a station for an hour is still reserved for that hour, even if only a few interventions occur. Avoid adding active minutes to that same person's coverage minutes as though they were separate staff members.

Worked example: less active work, unchanged staffing

Illustrative trial worksheet. These figures are assumptions, not Sharpa or DQ operating results.

Suppose the baseline and robot-assisted trial each complete 30 comparable orders during a one-hour observation period, to the same agreed quality threshold. The baseline uses 45 active human minutes across those orders, including its allocated preparation and close-down work.

For the robot-assisted trial, assume 15 person-minutes for preparation and cleaning, 12 for refilling, intervention and customer handoff, and six for remote assistance. Those are distinct tasks with overlapping entries removed.

The recorded active work is 33 person-minutes, or 1.10 human minutes per completed order. The baseline is 45 / 30 = 1.50 minutes per order. The difference is 12 active minutes across the trial, or 0.40 minutes per completed order.

That calculation establishes a change in active work under the assumed conditions. If the site still requires a worker dedicated to the station for the full hour, local coverage remains 60 minutes; remote help also requires a resource allocation. The worksheet has not demonstrated that a shift or position can be removed.

Check whether the twelve minutes occur in useful blocks and whether the worker can perform another needed task while meeting the station's supervision requirements. Repeat the comparison during representative busy and quiet periods. A single convenient hour cannot establish the operating pattern of a full week.

Leave the pilot with a staffing question answered

Keep the completed-order count, active person-minutes, dedicated coverage, waiting times and quality outcomes on the same dated sheet. Name the unresolved staffing assumption: for example, whether remote assistance can support several stations, or whether refilling interrupts another essential task.

Use the retail AI field guide for the wider customer-journey and operating questions. Bring the completed worksheet into the brief for an industry-focused China AI program, so the next discussion can focus on the work that still needs evidence.