Supply Chain Action Points
Read this first — the conclusion, and the moves to make:
- Within 14 days, log detect-to-stop for every line stop over two weeks, and publish the median as your baseline before any network or vision spend is approved; if the median is above 60 seconds, fix the escalation rule first.
- Ask for the cluster efficiency figure restated with the utilisation level and the baseline it was measured at, plus the same test run at your average month rather than your peak week; reject a single-point number.
- Write the headway policy and intersection right-of-way rules into the specification, and require simulation at 120 per cent of forecast fleet size with a documented deadlock recovery time before go-live.
- Name the person or the automatic rule authorised to stop the line, and set a 60-second default action (hold, divert or stop) for every anomaly raised by the vision system.
- Re-measure at two load levels, average month and peak week, 90 days after go-live; if the gain at average load is less than half the quoted figure, renegotiate the maintenance contract against the measured number.
JD Logistics' Beijing 5G factory demonstrates more than 100 AGVs and autonomous forklifts coordinated over a private 5G network, lifting cluster efficiency by over 20%. A '5G + machine vision' line monitoring cuts anomaly detection from 20 minutes to 1 second, showcasing the next wave of warehouse automation.
The Analysis
Two numbers came out of JD Logistics' Beijing 5G factory this week, and they are not the same kind of number. The site runs more than a hundred AGVs and autonomous forklifts over a private 5G network and reports cluster efficiency up by more than 20 per cent. On the conveyor lines, 5G plus machine vision cut anomaly detection from twenty minutes to one second.
The first number is a ratio with an unstated baseline and an unstated load. The second is a change in observation interval by a factor of twelve hundred. Put them in the same sentence and most readers will treat them as two proofs of the same claim. They are not. One tells you how much throughput a dense fleet can squeeze out of a constrained floor. The other tells you how fast you can now see a fault, which is only worth something if you can act on it.
Start with what a private 5G network actually does inside a warehouse, because the answer is not bandwidth. Wi-Fi in a racked building is a losing argument with steel: reflections, dead spots, and a handover between access points every time a vehicle crosses an aisle. So every autonomous vehicle has to carry a safety margin in its motion profile, an assumption that the network will sometimes be late. Private 5G removes most of that uncertainty by giving the fleet controller a bounded latency instead of a hopeful one. Now the vehicle can run a shorter headway, because headway is precisely the buffer you hold against message delay. Fleet throughput on a fixed floor is speed multiplied by density, and density is set by the minimum headway you dare to run. That is where the twenty per cent comes from. It is a dividend on shrinking the safety margin, not a dividend on making robots faster.
Which means the twenty per cent is load-dependent, and this is the thing to hold on to when a vendor quotes it at you. A headway dividend only pays where vehicles are actually competing for the same space. At sixty per cent utilisation, a hundred vehicles rarely queue at the same intersection, so cutting headway buys you very little. At eighty-five per cent, the same cut is worth a great deal, because the queue at the bottleneck intersection is what the whole floor is waiting on. A single number measured at one utilisation level tells you almost nothing about your own site. Ask for the curve, or at minimum for the utilisation at which the measurement was taken.
The detection number has a different structure and a much cleaner piece of arithmetic behind it. Twenty minutes at a line speed of one unit every three seconds exposes four hundred units before anyone knows there is a fault. One second exposes a third of a unit. That looks like a four-hundred-fold improvement in exposure, and it is, on paper. Now put a human in the loop. Suppose the decision to stop the line takes four minutes from the moment the alarm fires, counting the phone call, the confirmation and the supervisor's walk.
Four minutes at the same line speed is eighty units. So the real exposure went from four hundred units to eighty, and everything the sensor bought after the fourth minute is exposure you cannot remove because the constraint moved to the response. Cutting detection from twenty minutes to four removes three hundred and twenty units. Cutting it from four minutes to one second removes eighty. Roughly eighty per cent of the value sat in the first step, which was the cheap one.
The impact lands by order profile, not by company size. A high-volume e-commerce fulfilment site pushing tens of thousands of small parcels a day with short cut-offs gains the most, because its floor is congested most of the day and its cost per order is dominated by travel time. A business-to-business distributor shipping two hundred orders a day gets a network bill, a maintenance contract and a skill requirement, and no throughput gain, because its floor never queues. An exporter loading full containers on a weekly rhythm gets nothing at all; its warehouse is a staging area with a forklift and a loading bay, and cluster coordination solves a problem it does not have. An importer replenishing daily against short customer order cycles sits in the middle, and the part it feels is not the picks per hour but the cut-off moving later, which is worth more than the labour saving because it lets the same inventory serve a longer order window.
When does it reach your desk? Not as a technology story. A site that has genuinely cut its internal travel time will first re-price its cut-off, then its service commitments. Assume two to three quarters from commissioning to a commercial offer that reflects the gain. If your third-party provider starts promising a later same-day cut-off at the same rate, that is this, arriving. The inventory effect comes last, because it only appears once you trust the cut-off enough to lower the buffer.
Here is the part the announcement cannot contain, and it is what I would put in front of any buyer. The one-second detection is a monitoring loop. It is not a control loop. Between seeing a fault and stopping the line sits a decision about who is allowed to stop it, and in most factories that authority is deliberately diffuse, because stopping costs money and someone has to answer for it. So the binding constraint after the upgrade is not the camera, not the network, and not the model. It is the escalation rule. A site with one-second detection and a four-minute escalation rule has bought itself a very expensive way to watch defects at high resolution. The metric that decides whether this project paid is detect-to-stop, measured as a median and not as a best case, and it should be measured for a fortnight before any network money is committed.
There is a second order effect worth pricing too. AGV fleets fail differently from people. A human driver who meets an obstruction improvises. A fleet of a hundred vehicles that meets an unplanned obstruction will deadlock in a pattern nobody designed, and the recovery is manual, one vehicle at a time, which means the loss is not proportional to the fault but to the fleet size. That is why the headway policy and the right-of-way rules at intersections have to be written down and simulated at a fleet size above your forecast, not discovered on the floor. The counterfactual here is cheap and unglamorous: keep the network, keep the cameras, but leave headway policy untouched, and the twenty per cent stays in the slide deck while the floor behaves exactly as it did before.
↑ Back to the key points