France has become the first European country to open its roads to autonomous delivery robots. The information is accurate, it is official, and it went round the French press in a week.
It is also almost systematically told wrong.
A decree of 28 July 2026, published in the Official Journal on 7 August, sets out the technical requirements for type-approving these vehicles. Transport minister Philippe Tabarot welcomed a decision that “reflects our ambition to accelerate the deployment of autonomous vehicles in France”. The phrase that circulated in the coverage — “we are moving from experimentation to industrialisation” — sums up the stated ambition well.
Except that three elements of the text contradict the picture being painted. None of them is hidden: they are in black and white in the official statement. You only have to read them.
What the decree actually does
Let us start with the real content, because it is more technical and less spectacular than announced.
The government has not created a new legal category. The decree amends an existing text — the one of 17 August 2016 on the approval of category L vehicles, the regulatory family covering mopeds, motorcycles, tricycles and motorised quadricycles — to bring fully automated vehicles into it.
Two approval routes are opened: national small-series approval, and individual approval, that is to say vehicle by vehicle. These are heavy administrative procedures, designed for low volumes.
The conditions set are as follows:
- travel on the roadway only, pavements remaining prohibited;
- permanent remote human supervision;
- compliance with cybersecurity and data protection standards;
- zones and routes defined by local authorities, which alone have competence;
- maximum dimensions of 4 metres long, 2 metres wide, 2.5 metres high;
- payload up to 1,000 kilos.
The ministry stresses the objective: cutting polluting emissions and noise, smoothing urban logistics. It recalls that more than 200 trials have been run in France since 2015.
So much for the text. Now for what it does not say.
First point: nothing is rolling, and nothing will be any time soon
“Delivery robots are coming to our roads” assumes they are on their way. They are not.
A type-approval decree puts no vehicle on the road. It defines the criteria a manufacturer will have to meet to get its machine approved. In other words, it opens a counter.
To date: no approved vehicle, no government timetable announced, and above all an extra step the coverage almost always forgets — even once a model is approved, each local authority will have to authorise each zone and each route. A municipality can say no. Most, lacking technical expertise and faced with no feedback from experience, will say no out of caution.
Add one element noted by the specialist press: the harmonisation work started at UN and European Union level to define an approval framework for these vehicles is still under way. France has moved ahead of an international framework that does not exist. If that framework diverges from hers, it will have to be revised.
The honest headline is the one at L’Usine Digitale: France opens the way to type approval, but you will have to wait longer to see them on the road.
Second point: the “strict rule” is an admission, not a constraint
This is the most interesting element in the file, and it is presented back to front everywhere.
The press describes a “very strict rule” imposed on the robots, with the implication that France is tightly regulating a mature technology. The reality is the opposite. Permanent human supervision is not an excessive precaution applied to an autonomous machine. It is the acknowledgement that the machine is not autonomous.
A genuinely autonomous vehicle does not need an operator able to take over at any moment. If it needs one, it is because its driving involves situations it cannot handle alone — complex junctions, unpredictable behaviour, degraded conditions, roundabouts. The remote human is not there on a precautionary principle: they are there because they are necessary to make it work.
We met exactly the same mechanism a few days ago when examining what humanoid robots really do when nobody is driving them: a home robot sold as autonomous, but whose complex tasks are performed by a human operator in a virtual reality headset. The technology differs, the pattern is identical — and it is the same pattern we had already described with those platforms where AI hires humans to act in the real world.
Which raises an economic question nobody has put a number on: how many vehicles can one supervisor monitor at once? If the answer is twenty, the model is profitable. If it is two or three, a delivery driver has been replaced by a remote operator, an electric vehicle and a software infrastructure — and the financial equation changes entirely. Neither the decree nor the statement sets a ratio.
That is the constant of this generation of systems: they are presented as autonomous for as long as nobody looks at who is holding the controls. We observed it with software agents starting to act alone, where human supervision remains the implicit safeguard of every serious deployment. The difference here is that an error does not produce a bad file: it produces an impact.
Third point: the detail the illustrations give away
Look at the photos accompanying articles on the subject. They show small six-wheeled machines, the height of a cool box, trundling along among pedestrians. Those are the robots of Starship Technologies or Amazon Scout — pavement machines, present in more than 25 cities in the United States, the United Kingdom, Estonia, Denmark and Australia.
That is not what France has just authorised.
The French text covers vehicles that can reach 4 metres long, 2 metres wide and 2.5 metres high, with 1,000 kilos of payload, travelling on the roadway. These are not delivery robots as the public understands the term: they are driverless electric vans. A Renault Kangoo is about 4.5 metres long.
The confusion is not trivial. It changes the risk assessment entirely. A 40-kilo machine hitting a pedestrian at 6 km/h and a tonne of mass moving in traffic pose neither the same safety problems, nor the same insurance questions, nor the same acceptability issues.
And that explains the choice of the roadway: you do not run a tonne along a pavement.
Two incompatible models: the American pavement, the French roadway
That choice of the roadway is not a detail of layout. It makes the American business model non-transposable, and it is probably the most structuring piece of information in the file.
The model that works in the United States rests on three pillars: light and slow machines, a low unit cost, and the pavement as the space to travel in. Starship Technologies claims deployments in more than twenty-five cities in Europe, the United States, Estonia, Denmark and Australia; the company is now redirecting its business from university campuses towards food delivery and urban grocery. Serve Robotics operates on the same logic with Uber Eats in several American cities.
These machines work because they are harmless. At 6 km/h and 40 kilos, a collision with a pedestrian is an incident, not a tragedy. That is what allows cautious driving, frequent stops, and relative social tolerance.
France has made the opposite bet. By banning the pavement in the name of pedestrian safety — a defensible position — it sends these vehicles onto the roadway, with the cars. And a machine travelling on the road has to keep up with the flow of traffic, handle merges, overtaking, roundabouts, cyclists. The level of reliability required is of another order entirely.
In other words: France has ruled out the easy version of the problem and authorised only the hard one. That is consistent with the size bracket chosen — you do not run a tonne among passers-by — but it means the first French deployments will look like nothing that exists elsewhere. Not the same vehicles, not the same routes, not the same uses. Logistics zones, closed campuses, links between platforms and pick-up points are far more credible candidates than last-mile delivery in a city centre.
The environmental case deserves a look
The ministry puts forward a clear objective: cutting polluting emissions and noise, for urban logistics that is “cleaner and quieter”. The argument is repeated as is everywhere. It deserves a close look.
These vehicles are electric — the text requires it. But the van they would replace in the city increasingly is too, under the effect of low-emission zones and fleet renewal. And taking the driver out of an electric vehicle reduces neither its emissions nor its noise. The environmental gain does not come from automation: it comes from electrification, which is a separate movement and already under way.
What automation removes is not carbon, it is a wage cost. You may judge that desirable or not, but those are the right terms for the debate. The environmental promise only holds in one precise scenario: one where these robots would replace still-existing combustion rounds, or make it possible to pool journeys currently run empty. No public impact study establishes that at this stage — and the question of these machines’ material footprint, sensors and batteries included, is no better documented, at a time when the hidden environmental bill for AI is becoming better known.
The black hole of liability
Which brings us to the most serious point, and the one most absent from the coverage.
The French legal framework is not blank. The mobility orientation law of 24 December 2019 empowered the government to legislate by ordinance. Ordinance no. 2021-443 of 14 April 2021 established the criminal liability regime applying to vehicles with delegated driving, and decree no. 2021-873 of 29 June 2021 set out the detail, with specific provisions for automated road transport systems supervised and controlled remotely.
On paper, then, the ground is prepared. In practice, a fundamental difficulty remains, well identified by lawyers: French criminal law has no liability for things in one’s keeping. That principle exists in civil law, where it allows a victim to be compensated without having to prove fault. It has no criminal equivalent. And a criminal offence presupposes a fault attributable to a person.
Who is that person when a driverless vehicle, supervised from a remote centre, injures somebody? The supervisor, who was perhaps monitoring ten vehicles at once? The manufacturer, whose software misread the scene? The logistics client? The local authority that approved the route?
These questions are not theoretical: they will arise at the first accident. And they will arise before a judge who will have to decide with no settled case law. That is a major difference from the debate about software AI, where the stakes play out in civil and contractual liability — here, there are bodies.
The European AI regulation brings no direct answer on this point: it deals with the obligations borne by systems, not with the criminal attribution of an accident, as we set out in our analysis of what changes with the AI Act.
Who is ready on the French side
The decree was not written in a vacuum, and one national player clearly stands out.
TwinswHeel, a subsidiary of the small firm Soben, created in 2016, designs goods transport “droids” in various sizes. Its machines were tested in Montpellier under the Carreta project, with two uses: carrying parcels and mail for La Poste, and delivering fresh produce from a STEF platform to restaurants and trade customers. The company also claims Enedis, SNCF and Renault among its customers, and has deployed robots for the postal services of Ljubljana and Bratislava.
That is good industrial news: unlike humanoid robotics, where France has no player of any size, here there is a French company with real references and deployments abroad.
It is also the limit of the exercise. A small firm, a few dozen vehicles in trials, against American and Chinese players deploying in the thousands. The decree creates a domestic market protected by a demanding approval procedure — whether that protection will work in favour of French industry, or simply slow everybody down and give foreign manufacturers time to line up, remains to be seen.
What to take away
The fact is real and it counts: France is indeed the first European country to define an approval framework for these vehicles, and the text is serious, technically precise, backed by more than 200 trials.
But three things have to be said clearly.
Nothing is circulating. The decree opens an approval counter. No vehicle has been through it, no timetable exists, and every municipality will keep a right of veto on its own territory.
These vehicles are not autonomous. Permanent human supervision is not a constraint imposed on a mature technology, it is the finding that the technology is not mature. The one figure that will determine the model’s economic viability — how many vehicles per supervisor — is nowhere to be found.
These are not the robots in the photos. The text authorises 4-metre machines carrying a tonne on the roadway, not cool boxes on wheels trundling along. And the question of who will answer in criminal law for the first accident remains, to date, without a solid answer.
For delivery jobs, then, there is no immediate urgency — the subject joins what we found when analysing what studies on jobs at risk actually measure: between a regulatory authorisation and a deployment that changes a sector, there are years, and a great many things that never happen.
The figure to keep
1,000 kilos. That is the maximum payload the French text allows. Set against the images of small pavement robots illustrating this announcement everywhere, the gap says it all: France has not legalised an urban gadget, it has opened the way to driverless vans. And it did so before the UN and the European Union had agreed on the rules.