Pony.ai Verne has begun carrying invited passengers without an onboard safety operator on a 22-kilometre public-road route in Zagreb. The test is a concrete step beyond the supervised service launched earlier, but it is not yet evidence of citywide, unsupervised commercial operation.
- Invited riders travel between Verne’s headquarters, a business district and Zagreb airport.
- The companies say the rides have the approvals required for passenger testing.
- No disengagement rate, incident record, fleet size or launch date was disclosed.
What the Pony.ai Verne test actually changes
The joint announcement says no autonomous-vehicle operator or other staff member sits inside the car. That distinction matters because the August rollout available through Uber used an onboard licensed operator. Removing that person transfers immediate driving responsibility to the automated system and the operating procedures around remote assistance.
The route links Verne’s headquarters and one of Zagreb’s main business districts with Franjo Tuđman Airport. A fixed 22-kilometre corridor is operationally meaningful: airport routes involve faster roads, lane changes, pickups and repeatable demand. It is also narrower than a citywide service, so the announcement should not be read as blanket driverless availability across Zagreb.
EVwire and two European mobility publications independently reported the operator-free passenger rides. Their accounts match the core facts in the primary release: invited passengers, public roads, no employee aboard and a planned expansion over the coming months. None supplied independent safety-performance data, making the companies’ operational claims the evidence ceiling for now.
Pony.ai Verne divides the operating stack
Pony.ai supplies the seventh-generation autonomous-driving system, while Verne owns the local service layer and regulatory work. Uber provides ride access for the commercial service. This division allows each partner to specialise, but it also creates accountability questions when software, fleet operations, customer support and marketplace access sit with different companies.
The vehicle uses Pony.ai’s L4 domain controller on NVIDIA DRIVE AGX and the Halos operating system, according to the release. The company describes 360-degree sensing, redundant systems and fail-operational capability. Those are design descriptions, not public proof of how frequently remote support is needed or how the system performs in Zagreb’s rain, construction zones and unusual traffic events.
The companies say the existing fleet has completed several thousand customer rides and more than 200,000 autonomous kilometres since April. They report a 4.7 out of five passenger rating. The figures provide scale context but were not accompanied by an audit, crash summary, intervention rate or denominator for safety comparisons, so they should not be treated as a safety score.
Safety evidence is the next reporting test
For automated-driving systems, distance alone can conceal the hard questions. Readers need to know the operating-design domain: permissible weather, road classes, speed limits, construction handling, emergency-vehicle response and fallback behaviour. They also need consistent definitions for disengagement, remote assistance and incidents.
The joint release says the rides comply with applicable Croatian rules and required approvals. It does not name the approving authority or publish the permit. That gap does not invalidate the tests, but it limits independent verification of route boundaries, passenger conditions and reporting obligations.
A useful public evidence package would include kilometres by mode, every collision regardless of fault, safety-critical software disengagements, remote-assistance events and service interruptions. Comparable monthly reporting would show whether performance improves as the route expands. Without it, “fully driverless” describes who is inside the car, not the statistical safety of the service.
Commercial expansion remains conditional
Pony.ai says overseas partnerships represent a capital-efficient route to scale. The release points to a pipeline of more than 4,000 robotaxis and plans with Uber for more than 2,000 vehicles in Europe. A pipeline is not a deployed fleet: orders, approvals, financing, vehicles and local operations all have to align.
Zagreb can test whether a Chinese autonomous-driving supplier, a Croatian fleet operator and a US ride-hailing platform can function as one service. The model may reduce the need for Pony.ai to build every local capability, but each new market introduces mapping, language, road-rule, insurance and regulator requirements.
The strongest near-term milestone would be a defined public launch with disclosed fleet size and service area. The stronger long-term evidence would be sustained utilisation and safety reporting across seasons. Until those appear, the development is best understood as a controlled operator-free passenger test rather than a completed commercial transition.
| Verified fact | What remains unknown |
|---|---|
| 22-km public-road route | Full operating area and fleet size |
| No onboard operator | Remote-assistance frequency |
| Invited passengers | Open public launch date |
| Company says approvals obtained | Permit terms and reporting rules |
In plain terms: Pony.ai Verne has crossed an operational boundary by carrying invited passengers without a human in the vehicle, but public evidence still does not establish citywide availability, audited safety performance or a firm commercial timetable.
Related: NVIDIA and Palantir’s supply-chain AI stack and d-Matrix joining NVIDIA NVLink Fusion.
What changes when the safety driver leaves
An onboard operator can perceive a developing hazard, take the wheel and communicate directly with passengers. An operator-free test must replace those functions with automated fallback, remote support and clear passenger instructions. The release says redundancy exists but does not explain how riders contact support, how a stranded vehicle is secured or how quickly an on-road team responds.
Remote assistance also needs careful definition. A person may provide high-level guidance without steering, or a remote operator may exercise more direct control. Those are materially different arrangements for latency, staffing and responsibility. The partners did not publish their remote-operations design, so the article does not assume that “driverless” also means human-independent.
Why the airport corridor matters
Airport connections concentrate demand and can make a small fleet easier to dispatch, charge and monitor. They also expose vehicles to luggage loading, hurried pickups, unfamiliar riders and traffic conditions that change with flight banks. Repeating one corridor can generate useful operational data before a broader geofence is attempted.
The route’s commercial value will depend on more than driving capability. Verne must manage cleaning, charging, vehicle recovery, passenger verification and accessibility. Uber’s marketplace can supply demand, but the announcement does not say whether operator-free rides will be booked through Uber, offered free to invitees or priced as a public product.
How to read the scale claims
Pony.ai’s worldwide autonomous distance and China driverless distance show accumulated experience, but they are not a substitute for Zagreb data. Road design, signage, local driving behaviour and regulation differ. A transparent rollout would separate kilometres driven in Croatia from global totals and specify which software version produced each result.
The companies’ targets also use several different concepts: an overseas deployment pipeline, a planned European fleet with Uber and a global year-end fleet target. Those numbers can overlap. Investors and riders should avoid adding them together and instead watch actual registered vehicles, operating permits, paid rides and published utilisation in each city.
Passenger experience is part of safety
An operator-free service must work for people who have never used an autonomous vehicle. Clear boarding confirmation, emergency-stop instructions, accessibility support and reliable help channels are safety controls, not cosmetic features. The release reports a passenger rating for the wider service but does not separate invited driverless-test feedback from supervised rides.
Public trust will depend on how the partners communicate disruptions. A transparent incident page, plain-language explanation after material software changes and a way for road users to report unusual behaviour would help outside observers evaluate the rollout. Those practices would also reduce the gap between a tightly managed demonstration and a service used by ordinary travellers.
Insurance and responsibility are another open layer. Operator-free passenger tests create questions about how claims are investigated when the vehicle, software supplier, fleet operator and marketplace are separate entities. The announcement does not describe insurance terms or data retention after an incident. Publishing a responsibility map before broad public service would make it easier for passengers and other road users to understand whom to contact and which records are preserved. It would also show whether regulatory approval covers only testing or the eventual paid, no-operator service.
Seasonality will provide a harder test than a launch-day route. Zagreb experiences rain, fog, low winter light and roadworks that alter familiar lanes. Evidence collected across those conditions, paired with clear pause criteria, would tell readers much more than a single milestone statement.
A further measure is service continuity. A robotaxi operator should publish what happens when positioning, cellular connectivity, maps or a remote-support link degrades during a passenger trip. Safe behaviour may mean slowing, stopping in a suitable place or ending service before conditions worsen. The September announcement does not specify those thresholds. Verne can make this milestone more useful by documenting its fallback hierarchy, rider communication and recovery timing, then reporting whether those procedures worked during real interruptions. That operational record would connect the claimed hardware redundancy to an outcome passengers and regulators can evaluate.
Evidence matters.
FAQs
Is the Zagreb robotaxi service fully driverless?
The new test rides have no operator inside, but they are limited to invited passengers and a defined route. That is different from an unrestricted citywide public service.
Who operates the vehicles?
Pony.ai supplies the autonomous-driving technology and Verne runs the local fleet and service operations. Uber provides access for the related commercial service.
How safe are the tests?
The companies describe redundant systems and prior kilometres, but they did not publish Zagreb-specific incident, disengagement or remote-assistance rates.
When will the route expand?
Verne says routes will expand over coming months, but no public schedule, fleet count or full-service date was announced.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
