Supply Chain Action Points
Read this first — the conclusion, and the moves to make:
- Within 30 days, request four figures from any provider: cycle time per pick, picks per human intervention, mean minutes to recover, and station uptime per shift; a provider that cannot supply them is unproven regardless of the video.
- Compute your own breakeven intervention rate before any trial, using the formula above with your labor rate and pick volume, and hold the go/no-go at one intervention per 30 picks until eight weeks of your own data exist.
- Model effective throughput as the minimum across stations, and name the station that becomes the constraint if pick rate rises 30 percent, then put that upgrade in the same budget cycle rather than the next one.
- Run a six-week pilot in one zone against an untouched control zone, tracking interventions per 1,000 picks weekly, and stop if the rate has not fallen 30 percent by week four.
- Put a productivity and data-sharing clause in the next contract: a stated reduction in cost per order within two quarters, evidenced by audited volumes and exception counts.
At WAIC 2026 on July 17,神州控股's supply-chain brand Kejie and Daka Robotics demonstrated multiple embodied-AI robots performing live order picking inside Kejie's Tianjin Wuqing warehouse, completing task receipt, path planning, item recognition and sorting-cart return. Kejie, which runs nearly 200 warehouses with a peak of 5 million daily orders, is opening real scenarios to train and deploy logistics robots.
The Analysis
On July 17 at WAIC, Kejie, the supply-chain arm of Digital China Holdings, and Daka Robotics ran embodied-AI robots through live order picking inside Kejie's Tianjin Wuqing warehouse. The robots took the task, planned a path, recognised the item and returned the sorting cart. Kejie runs close to 200 warehouses with a peak of five million orders a day, and it is opening real sites as training and deployment ground for logistics robots.
Strip the demo language and what remains is a claim about a service rate at one station in one process. That is a testable claim, and the right kind to make. What the announcement does not carry is the four numbers that would let you evaluate it: cycle time per pick, picks between human interventions, minutes to recover from one, and station uptime across a shift.
A warehouse is a queue network in a fixed sequence. A faster server in the middle does not raise output unless the stations on either side can absorb the new rate. That is arithmetic rather than opinion, and it is where most automation cases quietly fail.
Start with the numbers that were given. Roughly 200 warehouses, a peak of five million orders a day. That peak is a portfolio figure, and what a robot business case needs is the peak at one site, which was not given. If volume were spread evenly, the average site would see 25,000 orders a day, and I want to be explicit that this is my assumption, not a disclosed figure. Real networks are not even. A handful of sites carry the peak, so the addressable footprint for a picking robot is much smaller than 200, and the honest question for any provider is how many of those buildings have the volume to justify one station.
Picking is where travel dominates, and travel is set by slotting and order profile: lines per order, cube per line, how stable the SKU set is. Embodied AI's advantage over fixed automation is that it tolerates an unstructured layout, so you do not rebuild the building to accommodate it. Its disadvantage is speed and reliability inside a highly structured aisle, where fixed automation is simply faster. The fit is therefore narrow. It is best where the layout is awkward and the order profile repeats often enough that the robot meets the same geometry.
The second mechanism is cost structure. Robots convert labor hours into capital, maintenance and floor space. That trade rewards utilization and punishes idleness, which makes the design variable the peak-to-average ratio rather than the volume. A site with a threefold peak needs enough stations to cover the peak and eats idle time the rest of the day, or a hybrid crew that keeps robots busy on the base load and humans on the spike. A site with a flat profile can run stations near saturation and the case looks wonderful. Same technology, opposite conclusion, and the difference is a number most proposals never ask for.
Then there is whether the gain lands at all. Effective throughput is the minimum across stations, not the sum. Raise pick rate and replenishment must keep pace, the pick face empties faster, pack and sort receive faster, and the dock sees tighter windows. A picking-only deployment therefore delivers less than the labor arithmetic suggests, and the shortfall shows up as congestion at the next station rather than as a shortfall at the robot.
The gains land unevenly. High-volume sites with stable SKUs gain first, which in practice means a large third-party operator or a brand's own distribution centre. Multi-client operators with volatile mix gain last, because exceptions and changeovers dominate and a robot fleet is least flexible exactly where flexibility is worth the most. Smaller shippers gain only indirectly, through their provider's cost per order, and that gain is captured by the provider until the next contract cycle. If you are the shipper, the productivity improvement is not yours by default. It becomes yours at renewal, and only if you have a baseline.
Timing runs in quarters, not weeks. The demonstration was July 17. Credible third-party numbers, meaning cycle time and intervention rates measured outside a showcase, are a two-to-four-quarter proposition on my reading. A single-zone pilot also tells you nothing about a network, and rollout risk scales with how different your sites are from each other.
Here is the part the coverage leaves out, and it is the only part that decides the investment. A robot that picks faster but needs a human every N picks can be slower in effective terms than a steady human. Write it down. Effective rate equals gross rate minus gross rate squared, times interventions per pick, times minutes per intervention, divided by sixty. With a gross rate of 180 picks an hour and three minutes to recover, both my assumptions rather than the article's, breakeven against a 120-pick human arrives at roughly one intervention per 27 picks.
Nobody published interventions per pick, and that omission alone makes the capability claim unfalsifiable. The other missing piece is where the cost lives. Training on real scenarios is the genuine asset here, because the long-run cost driver in embodied picking is the tail of rare items, not the average one. A demo on popular SKUs proves the head of the distribution. Your cost sits in the tail, and the tail is what you should be pricing.
Run the arithmetic for your own site rather than borrowing mine. Take 25,000 orders a day at three lines each, 75,000 picks, over 20 operating hours, so 3,750 picks an hour. At 120 picks per human hour that is around 31 pickers, and those inputs are my assumptions, not figures from the article. Swap in stations at a gross 180 with the intervention rate you actually measure, and breakeven stops being a matter of opinion. It becomes a line you can draw before you sign.
Both failure modes are expensive. Deploy only where the profile fits and the rest of the portfolio waits for volume to justify a station, which is the correct answer and a slow one. Deploy to prove coverage and the average site carries capital for a demonstration, which is the most expensive way to learn. Do nothing and your provider keeps the productivity gain as margin until you renegotiate, at which point you have no baseline to argue from.
Three traps. Benchmarking a demo against your own operation, where the SKUs are less friendly and the demand is less stable. Counting gross picks instead of effective ones, which is how a pilot looks better than a rollout. And forgetting the physical envelope, because aisle width, charging positions, floor space and the replenishment path all change when a fleet arrives, and those costs sit in the building rather than on the invoice.
Ask for four numbers and a control zone. With both, the decision makes itself. With neither, you are being shown the head of a distribution and asked to price the tail.
↑ Back to the key points