China Robot Cells: Count Good Units per Scheduled Shift

A robot cell with the shortest demonstration cycle can produce fewer good units during a working shift. Changeovers, recovery, missing inputs and quality losses consume the time and output that a cycle-time comparison leaves unexplained.

When evaluating an AI-enabled robot cell in China, ask for two reconciled records: where the scheduled time went, and what happened to the units processed during it. Compare first-pass good units per scheduled hour before using a faster cycle as evidence of more usable capacity.

Define one cell, product and observation window

Choose the same product specification, cell boundary and acceptance rule for both configurations. State which loading, inspection and handoff steps the boundary includes. A robot-arm movement and a complete cell cycle can cover different work.

Agree the scheduled production window before collecting results. Keep its length fixed when comparing configurations, and show any exclusions explicitly. Removing an inconvenient stoppage from the denominator changes the question the result answers.

MEASURE 2.3 in NIST's AI Risk Management Framework 1.0 (2023) addresses performance evidence under conditions similar to deployment. For this exercise, use a representative product mix, staffing pattern and input supply, then retain the conditions alongside the results.

Reconcile the time before calculating output

For each interval, record the start, finish and operating state. Separate active processing, changeover, fault recovery and waiting for input or downstream space. Together, the states should account for the whole scheduled window without overlapping minutes.

If two problems overlap, record one primary time allocation and retain the other cause as context. An interval waiting for material while a fault remains open cannot be subtracted twice from the same clock. Likewise, a short pause already included in the measured average cycle time needs no second deduction.

Use existing logs, permitted records or an offline exercise. The evaluation does not require visitors to interrupt production or introduce faults into live machinery.

Worked example: the faster cycle loses the shift

Hypothetical comparison. All quantities below are assumptions.

Both configurations have a 480-minute scheduled window. Assume one completed processing attempt on a different workpiece per active cycle, constant active-cycle duration, inspection completed within the window and no reworked unit returning during it.

Record	Faster-cycle configuration	Steadier configuration
Scheduled minutes	480	480
Active processing minutes	240	360
Changeover minutes	90	45
Recovery minutes	60	30
Waiting minutes	90	45
Seconds per active cycle	30	40
Completed processing attempts	480	540
First-pass good units	432	513
Units not accepted first pass	48	27

The first time budget reconciles as 240 + 90 + 60 + 90 = 480 minutes. Its processing count is 240 × 60 / 30 = 480 attempts. With an assumed 90% first-pass yield, it produces 480 × 90% = 432 good units.

The second budget is 360 + 45 + 30 + 45 = 480 minutes. It completes 360 × 60 / 40 = 540 attempts. At an assumed 95% first-pass yield, it produces 540 × 95% = 513 good units.

Across the same eight scheduled hours, the results are 54 good units per hour and 64.125 good units per hour respectively. The steadier configuration produces 81 more first-pass good units despite its longer active cycle.

The comparison identifies the operating assumptions that matter. It does not establish why one configuration has fewer losses. Investigate the separate changeover, recovery, waiting and quality records before attributing an improvement to AI or to a particular hardware change.

Keep attempts, unique units and later recovery separate

In an actual pilot, give each workpiece a stable identifier. A second attempt on the same piece uses cell time but does not create a second unique good unit. Record rework and retest activity against that original identifier.

Distinguish first-pass good output from units accepted after rework. If inspection remains unresolved at the cutoff, keep those units pending. If a later shift releases them, preserve the link between processing time, disposition time and the final acceptance record. State which measure the report uses before comparing totals.

Repeat the observation across representative shifts. A quiet window with one product may omit the changeovers, replenishment and support delays your intended operation must absorb. Keep the distribution of results and the reasons for unusual shifts alongside the average.

Turn the largest loss into the next pilot question

Leave with a reconciled time budget, unique-unit counts, acceptance status and the configuration used. Name the largest unresolved loss and the record needed to explain it. A faster motion setting is useful only if the resulting operating system meets the required output and quality conditions.

Use the robot recovery guide to examine restart responsibilities and unfinished work. Bring the time-and-output record to an industry-focused China AI program so the next discussion can address the cell constraint your team needs to resolve.