Neolix secures Dapeng routes to complete autonomous delivery coverage of all 10 Shenzhen districts

The Analysis

When I read that one operator now holds normalised operating rights across all ten of Shenzhen's districts, my first reaction is not to think about vehicles. I think about a queue. Route rights are admission control. Somebody at the district level decided that five specific edges in one specific district may carry a new class of vehicle, for twelve months, under a defined set of conditions. Admission is not throughput. Getting into the queue is not the same as moving through it, and this announcement tells us a great deal about the first and almost nothing about the second. That gap is where I would put my attention if I were the network operator or the shipper deciding whether to put real parcels on those five lines.

The shape of it is worth stating precisely, because the numbers here are unusually clean. Five core routes. A twelve month normalised trial, which is a different animal from a demonstration and a different animal again from an open permit. Three subdistricts, Kuichong, Dapeng and Nan'ao, out on the far eastern edge of the city. Three kinds of node stitched together by those routes: parcel depots, industrial parks and coastal communities. Ten districts once this one lands, which the company says makes it the first operator holding citywide route rights. And a qualification process run across five dimensions and eighteen indicators, with the operator describing technical controls, institutional safeguards and emergency response as three parts of one safety design.

Here is my observation, and it is the reason I am writing this as a column about acceptance rather than a column about technology. Five dimensions and eighteen indicators is a serious assessment framework, and I genuinely want to read the indicator list. But the announcement publishes no operating number at all. No parcels per day. No availability. No intervention rate. No cost per parcel. If you are the person who has to decide whether this enters your network, a route licence is an input to that decision, not the output. So the useful question is not whether the technology works. It is what you would need to measure, over what window, against what control group, and at what threshold you stop, before you let it carry anything you have promised a customer.

Start with the geography, because geography is the cost driver and it is the part people skip. Kuichong, Dapeng and Nan'ao are not the dense core where a courier clears sixty stops on an electric trike before lunch. They are the far eastern peninsula: long distances between stops, industrial parks that release bulk outbound traffic at fixed hours, and coastal communities whose population swells on summer weekends. Long lines and low stop density are exactly the profile where a human courier's cost per parcel climbs, because you are paying for windscreen time rather than doorsteps. That is the economic case for putting a machine on the line, and it is a genuine case. It is also the only part of the case that a licence actually proves.

Now find the bottleneck, because this is where most pilots fool themselves. A delivery route is not one process, it is four in series: consolidate and sort at the depot, load, move, then hand off the last hundred metres to a person or a locker bank. A vehicle that drives itself replaces the third stage. It does nothing for the first, second or fourth. If your depot needs ninety minutes to build a route ready cage, or if your coastal stops have nowhere safe to pull over and no locker within a reasonable walk, the machine will wait, and a waiting machine is pure cost. In queue terms you have shortened service time at one station in a four station line. Throughput rises only until the next station saturates, and then the queue simply relocates.

That is why I keep saying find the bottleneck before you sign anything. Little's Law is unforgiving on this point. Work in process equals throughput multiplied by cycle time. Cut cycle time at one stage and, if the stages either side of it cannot absorb the change, you have moved the waiting, not removed it. The honest way to run this trial is to write down the four stages, put a measured service time against each one for the current human operation, and only then predict what happens when stage three gets faster. If stage one or stage four is your constraint today, a self driving vehicle will not save you money this year and you will learn that expensively.

Let me put numbers on it, and I will flag every input as an assumption, because none of these come from the announcement. Assume one vehicle per route at the start of the trial, so five vehicles. Assume two shifts of eight hours per day and six loaded trips per shift. Assume two hundred parcels per loaded trip. That gives five routes times one vehicle times two shifts times six trips times two hundred parcels, which is twelve thousand parcels a day. Assume three hundred operating days in the year. That is three point six million parcels a year on these five lines, and I want to be clear that this is my model, not anybody's published forecast.

Now price it, again with assumptions stated. Assume the human courier cost in this far eastern, low density profile is one point eight yuan per parcel, which is high, but this is exactly the kind of profile where it is high. Assume the vehicle's variable cost, energy, maintenance, connectivity, comes to zero point nine yuan per parcel. Assume the fixed cost per vehicle per year, covering the vehicle itself, the remote safety operator's share, insurance, charging and yard space, is three hundred and sixty thousand yuan, so one point eight million yuan across five vehicles. Variable cost on three point six million parcels is three point twenty four million yuan. Add the one point eight million fixed and the total is five point zero four million yuan, which is one point four yuan per parcel.

Against the human operation, three point six million parcels at one point eight yuan is six point four eight million yuan. The model therefore saves one point four four million yuan a year, a reduction of zero point four yuan per parcel, or about twenty two percent. That is a decent number, and it is also a thin one, because it depends entirely on volume actually showing up. The break even volume is where fixed cost divided by the per parcel saving clears: one point eight million divided by zero point nine yuan is two million parcels a year. Spread across five vehicles and three hundred days, that is about one thousand three hundred and thirty three parcels per vehicle per day. Below that line the machine costs more than the person. I would write that number into the trial agreement as a hard gate.

Run the sensitivity, because this is where the decision really lives. Suppose availability lands at eighty five percent instead of the ninety five percent the model assumed. Volume drops to about three point two two million parcels. Vehicle cost becomes roughly two point nine million variable plus one point eight million fixed, around four point seven million. The human baseline for the same volume would be about five point eight million. The saving falls from one point four four million to about one point one million, a loss of roughly a quarter of the benefit. Now suppose volume is only half the model, one point eight million parcels. Vehicle cost is one point six two million plus one point eight million, three point four two million, which is one point nine yuan per parcel, worse than the human one point eight. At half volume the whole case inverts.

So here is what I would measure, with thresholds, and I would insist on receiving the raw event log rather than a dashboard, because a dashboard is someone else's opinion about which numbers matter. Availability, meaning the share of scheduled shift hours the vehicle is actually in service, at ninety five percent or better, with a hard stop below ninety. Interventions per thousand kilometres of autonomous operation, at one or fewer, with a hard stop above two. Task completion without human fallback, at ninety eight percent or better, counted against trips actually dispatched. On time arrival at the handover point inside the promised window, at ninety five percent or better. Cost per parcel against the control group, down by fifteen percent or more, and volume per vehicle per day above one thousand three hundred and thirty three.

Then the safety and data thresholds, which are the ones everybody forgets to write down until they need them. Attributable incidents per one hundred thousand kilometres, at one or fewer, with one hundred percent of any collision event producing a retrievable log within twenty four hours. Time from a remote takeover command to the vehicle reaching a minimal risk state, thirty seconds or less, and I would test that rather than accept it. Time for a field responder to reach a stopped vehicle, twenty minutes or less, and you should set that number yourself from your own subdistrict geography, not inherit it from anyone. Raw event log exported at least once a day, with timestamp, distance, intervention cause category and a weather flag on every record. If the vendor cannot export that log, you cannot audit anything, and an un-auditable trial is worth nothing.

One threshold that almost never gets written down is unplanned dwell, and on coastal routes it is the one that bites. These are narrow village roads with parked cars, holiday traffic and the occasional delivery truck blocking a lane. A vehicle that stops and waits because it cannot resolve a scene generates no parcels and costs the same fixed money. I would set unplanned dwell at forty five minutes or less per vehicle per day, measured from the event log rather than from a driver's recollection, and I would require a cause code on every instance, because the cause distribution tells you whether you have a technology problem or a road furniture problem. Those two have completely different fixes, and only one of them is the operator's to solve.

The ramp deserves a rule too, because ten districts of rights creates a temptation to grow fast. Add routes one at a time, and only add one when the routes already running have held every threshold for four consecutive weeks. That sounds cautious, and it is, but the failure mode here is not a dramatic one. It is a slow bleed: you add a fourth route, availability drops because the remote team is now watching more vehicles per person, intervention response times stretch, and nobody connects the two because the metrics are reported per vehicle rather than per operator. Report span of control alongside availability every week, and make both of them gates on adding capacity.

The five dimensions and eighteen indicators deserve a specific word. They were built to answer the regulator's question, which is whether this operator is safe enough to hold route rights. They were not built to answer your question, which is whether this vehicle makes your network cheaper or faster. Those two questions overlap but they are not the same, and it is a common error to treat a safety qualification as an efficiency qualification. Ask for the indicator list, map it against your own table, and note where you have no coverage at all. In my experience the gaps sit exactly where you would expect them: cost per parcel, handover dwell time, and the cost of the exceptions nobody counts.

Design the pilot as an experiment or do not bother running it. Five routes is a gift here, because it lets you run a crossover. Put three routes on the vehicle and keep two on the human courier for the first eight weeks, then swap, so that every route spends time in both arms and you control for the fact that routes differ. Pair on the things that actually drive cost: stop density, round trip distance, and whether the route serves an industrial park or a coastal community. Randomise by route and by week, not by individual parcel, because splitting one route across two methods inside a single day destroys the comparison and creates a dispatch nightmare.

Set the sample size and the thresholds before the first vehicle moves, and write them down where everybody can see them. Eight weeks per arm, and I would not draw any conclusion on fewer than forty thousand parcels per arm, because below that the week to week noise from weather, promotions and staff absence will swamp the signal. Pre registering the thresholds matters more than the thresholds themselves. If you decide what counts as success after you have seen six weeks of data, you will rationalise whatever happened. Decide now that ninety five percent availability is the bar, decide now that one thousand three hundred and thirty three parcels per vehicle per day is the gate, and then let the data rule.

Seasonality will wreck a twelve month average if you let it. The trial starts in late July, so it runs straight through the summer coastal peak, then through typhoon season, then into a quiet winter. August traffic in a seaside subdistrict and December traffic are not the same operating condition, and an annual key performance indicator will average them into something meaningless. Report monthly and quarterly, never annually, and separate planned downtime from unplanned. A week of storm closures is not a technology failure, but if you let it sit inside the availability figure without a flag you will draw the wrong conclusion in both directions.

Now the measurement traps, and there are more of them than people expect. Distance: is the reported figure autonomous distance or total distance including any human driven repositioning? Interventions: does a remote operator tapping confirm on a screen count as an intervention, or only a full takeover? Incidents: are you counting only injury accidents, or every contact? Availability: is the denominator scheduled hours or total hours, and does charging time sit inside or outside it? Each of those choices can move a headline number by several points, and every one of them should be defined in the trial agreement before day one, in writing, with an example.

There is a bigger trap sitting behind the phrase citywide coverage. Route rights are not capacity, and capacity is not a service commitment. Ten districts of route rights tells you where a vehicle is legally allowed to operate. It tells you nothing about how many vehicles exist, what they cost, whether they run in the rain, or whether anyone will commit to a delivery window you can sell. If a commercial team reads this announcement as a promise of citywide autonomous coverage and prices a product on it, that is a self inflicted problem. Keep the licence, the fleet and the service level agreement as three separate documents with three separate owners.

Responsibility and insurance need the same treatment. Who pays when a parcel is damaged inside an unattended vehicle, and how is that proven? Who is the responsible party when a remote operator takes control, and what is that person's span of control, meaning how many vehicles they are watching at once? That span is an operating parameter and it belongs in your threshold table, because one operator watching twenty vehicles and one operator watching five are not the same safety case. Ask also who owns the operational data, how long it is retained, and what happens to it if the trial ends. Data questions feel like paperwork until the day they are the whole dispute.

One more, from the handover end, because this is the one that generates customer complaints. If the vehicle drops into a depot or a locker bank, what is the proof of delivery? A scan at the locker, a photograph, a signature on a handheld, or nothing? On a coastal route with a holiday population, a failed collection or a disputed delivery will be argued weeks later, and the only thing that settles it is the record. Decide the proof of delivery format before launch, make sure it lands in your existing system rather than in a separate portal, and test it with a deliberately failed delivery in the first month.

Alternatives deserve honest treatment, because the machine is not the only lever on this cost. A hybrid is often the better first move: let the vehicle take the long transfer leg from depot to village hub, and let a person do the final hundred metres where the doors, the stairs and the conversations are. Densifying pickup points does something different again, it changes the cost structure of the last mile instead of the cost of moving between stops. Concentrating vehicle runs on the industrial park's fixed outbound window and leaving the scattered residential stops to people is a third option. Each has a different break even volume, and the one you pick should follow from where the bottleneck actually is.

Here is the calendar I would put on it. Before the end of August, get the eighteen indicator list and the raw log field definitions into the trial agreement, and name the person on your side who audits them. By the end of October, at roughly the three month mark, run the first gate: if availability is below ninety percent or interventions exceed two per thousand kilometres, pause any plan to add routes and fix the cause first. By the end of January, at roughly six months, read the crossover result: if cost per parcel is not at least ten percent below the control arm, scale back to two routes rather than pushing through. Before the twelve months close, decide on renewal, and note that what happens to route rights after a normalised trial ends is not something this announcement answers, so do not sign anything that assumes it.

I will say the plain version of all this. A route licence is a ticket, not a report card. It tells you the regulator let a vehicle onto five roads for a year. It does not tell you whether your parcels get there cheaper, and nobody in this story has published the numbers that would tell you that. Put the queue on paper, locate the bottleneck, write the thresholds down before the first vehicle moves, and keep two routes on people so that you can actually tell the difference. Do that and the decision gets simple, because the arithmetic will tell you what to do by the end of January.

↑ Back to the key points

— By Ravi Chandran