Executive Summary

Key figures

Source: Cellina et al. (2026), arXiv:2608.26669v1, main text with Figures 4 and 5

The campaign followed a full factorial design. The vehicle under test ran at 100, 115 and 130 km/h, and the distance at which the lane change was triggered ranged from 20 to 60 m. The speed limit on this stretch is 130 km/h, which is where the upper bound came from. For each speed combination the team picked three distances, one near the critical distance set by the regulation and one near the minimum activation distance it specifies. The point was to see whether the system actually tells apart the situations where it may change lanes from the ones where it must not, so the boundary was crossed deliberately from both sides. Combinations corresponding to critical situations were repeated several times.

The 27 recorded manoeuvres fall into four groups by outcome. Twelve were completed normally with the distance requirement met, six were completed even though the distance fell short, two were suppressed by the system although the distance was sufficient, and seven were suppressed because the distance was short. That last group of seven is the intended behaviour, the system doing exactly what the regulation asks. For the two suppressed with room to spare, the paper offers driver attention monitoring or excessive precaution by the system as possible causes, and leaves them as conjecture.

There is a blank inside this classification that a reader should know about. Three of the eighteen completed manoeuvres have no value at the moment of lane crossing, because they took place outside the LiDAR measurement range. For the nine suppressed ones the question does not arise, since suppression happens before the lane is crossed and the measured quantities at that instant are undefined. The number 27 is not built from data of uniform density.

Counting the suppressed manoeuvres is where this work parts from what came before. A comparable open-road study published in 2025 evaluated five commercial lane change systems in live traffic, but manoeuvres the system aborted on its own were left out of the analysis. Look only at completed manoeuvres and you can tell whether a system is too permissive; you cannot tell how well it discriminates. The ability to screen out the situations that should be refused only shows up if you count the refusals.

It is not that nobody had measured commercial driver assistance on public roads. Most of that work sat on the longitudinal side. A 2008 field operational test observed how drivers interact with adaptive cruise control, a 2014 study retrofitted cooperative adaptive cruise control onto four production vehicles and ran them in real traffic, and a 2021 study put seven commercial cruise control systems on the road to see whether speed disturbances grow as they pass down a platoon. Measuring the gap to the car ahead has a long record. Holding the move into the adjacent lane against a distance threshold written into a regulation is, to the authors' knowledge, the first public-road campaign of its kind.

Type approval testing happens on a controlled proving ground. The approach is sound in that the same conditions can be repeated and the results reproduced. The cost is that only a limited set of scenarios fits inside it. Nothing in the design of those tests guarantees that behaviour confirmed on the proving ground carries across the system's actual operating domain.

Lane change assist adds one more condition. The function is usually geo-fenced to selected highways, so switching it on inside a proving ground requires the manufacturer's collaboration. For type approval that is no obstacle, since the manufacturer is the party seeking the approval. The problem comes afterwards. In market surveillance, where the point is to check whether cars already sold behave as the regulation requires, the party under inspection holds the key to the conditions of inspection.

Step outside the proving ground and measurement becomes the obstacle instead. Conventional ADAS testing pins vehicle position to the centimetre with RTK satellite navigation, but that precision does not survive tunnels, canyons and areas with limited satellite visibility. This is why the team chose LiDAR as the main measurement instrument. It is less precise than satellite positioning, but it returns a value almost anywhere.

The acknowledgments credit the technical and operational support of the Netherlands Vehicle Authority (RDW) and of IVEX, the company behind the LiDAR system carried by the approaching vehicle. RDW is one of the type approval authorities in Europe. The side that grants approvals lent a hand to a procedure that measures post-approval behaviour again out on the road. Independence here means independence from the manufacturer, not a position outside the regulator.

Reproducing a regulatory test on a public road requires a car approaching in the adjacent lane. The team assembled three type-approved passenger cars. The first is the vehicle under test. It carries the R79 ACSF of category C lane change capability, which was checked before the campaign and confirmed to work on geo-fenced highway stretches. The second plays the approaching vehicle defined by the regulation, following in the adjacent lane at a target speed higher than the vehicle under test. The third is a support car.

The support car is the most practical piece of this design. It never appears in the experiment itself. It sits behind the vehicle under test in the same lane to stop another car from drifting in and invalidating the run, and at the same time it holds a fixed gap to the vehicle under test with its own adaptive cruise control. That fixed gap becomes the ruler. As soon as the approaching car draws level with the support car, the vehicle under test flicks on its indicator and starts moving toward the lane marking. Instead of a person eyeballing the distance, the gap held by a cruise control serves as the reference.

Using LiDAR for traffic safety assessment is not new in itself. Earlier work mostly parked the sensor at the roadside to detect and track passing vehicles and vulnerable road users, and that arrangement only produces data at the spot where the equipment stands. Studies that put LiDAR on a moving vehicle exist too, but their goal was broader safety assessment rather than type approval or market surveillance. What is different here is that the LiDAR rides on a car taking part in the experiment, measuring the distance between the two vehicles at exactly the instant the regulation nominates for the verdict.

The vehicle under test carried a commercial vehicle detection and tracking system built from a Septentrio dual-antenna GNSS receiver and an Ouster OS1 128-layer LiDAR, while the roof of the approaching vehicle held a LiDAR and camera system developed by IVEX of Belgium. The approaching and support cars each carried an OxTS RT1003 GNSS/INS receiver for ground truth. The LiDAR software estimates the position, velocity, heading, width and length of surrounding vehicles at 10 Hz. Satellite positioning was not the primary instrument for the verdict; it served only as the yardstick for how accurate the LiDAR was.

Marking the timestamps was done by hand. The moment the turn indicator first came on was read from a dashboard camera, the moment of crossing the lane marking from an outside camera used for lane detection, and both were annotated manually for each test. The temporal resolution is one second. The team writes that this procedure proved more accurate and repeatable than manual driving, and states the room for improvement in the same breath. The delay between indicator activation and lane crossing varied from run to run and could not be quantified, and the cruise control of the approaching car was already reacting to the lateral motion of the vehicle under test, effectively reducing its speed.

The yardstick for the verdict is the critical distance defined by R79. At the moment the vehicle under test crosses the lane marking, if the car approaching in the adjacent lane is closer than this distance, the system is supposed to suppress the manoeuvre.

Here $v_{app}$ is the speed of the approaching vehicle and $v_{ALC}$ the speed of the car changing lanes. The rest are values the regulation nails down. The maximum deceleration of the approaching vehicle $a$ is 3 m/s², the reaction time before that deceleration starts $t_B$ is 0.4 s, and the minimum time gap that must remain between the two cars after the manoeuvre ends $t_G$ is 1 s. The first two terms are the distance the trailing car needs to brake down to matching speed, and the last term is the margin it must still be left with.

As the speed difference approaches zero, the first two terms vanish and only the last one remains, because the trailing car has no particular need to slow down. The one-second gap still has to hold. Behind a car travelling at 130 km/h, one second is 36 m. The curve below shows that relationship. At the left edge, where the relative speed is low, the curve is nearly flat, and even there the required distance never drops below 36 m.

All six manoeuvres classified as overshoots came from this low relative speed region. The paper attaches a careful reading to that. Even when the speed difference is small enough that no deceleration is needed, the one-second gap requirement stays alive, and the system may have failed to detect the threshold in this edge case. It is also a stretch where the curve is flat, so a small discrepancy flips the verdict.

Equation (1) is defined on the assumption that the approaching vehicle is faster than the vehicle under test. The six cluster where that assumption barely holds, and the paper's reading is that the one-second gap must be respected there as well. The distance used for the verdict is measured bumper to bumper, and splitting the analysis into three charts by speed rests on the assumption that the speed of the vehicle under test stayed constant through each run. With cruise control holding a target speed, that is not a stretch, but the promises underneath the numbers six and three deserve to be read alongside them.

That is how measurement error becomes part of the verdict. The team kept only the intervals where both vehicles held an RTK fixed integer solution, compared those against the LiDAR, and obtained a position standard deviation of 0.83 m and a velocity standard deviation of 1.40 km/h. Since the lane crossing was annotated by hand at one-second resolution, there is a temporal error as well. Draw those uncertainties as a rectangle around each measurement point, and the margin between that rectangle and the regulatory curve is the strength of the conclusion. The three runs with the vehicle under test at 130 km/h can be called overshoots at 3-sigma, a 99% confidence interval. The other three, at 115 km/h, stay at 2-sigma and do not carry the same certainty.

What RTK was actually for becomes clear here too. Both vehicles held a fixed integer solution simultaneously in only 2.3% of the recorded data. Satellite positioning on its own would not have supported this experiment at all. LiDAR was less precise but delivered a value almost all the time, and that value was accurate enough to pin a regulatory shortfall at 3-sigma. Those two sentences are the whole of what the paper claims on the methodological side.

For a data practitioner, the part of this paper worth taking away is not the number six but the route by which the number was produced. A type approval certificate asserts that the test results are true. Yet the conditions that produced those results, the stretches where the function can be switched on and the venue where the test can be run, sit with the party under inspection. Geo-fencing is a safety design, and to a third party trying to verify compliance it is also an access control. Having a regulation and having compliance independently verified are two different things.

What this campaign changed is not the regulation but the route by which the data is acquired. Put a LiDAR on the roof, line up three cars, and you can measure the distance at the exact instant the regulation names without asking the manufacturer for anything. That is the point at which market surveillance comes to mean measurement rather than document review. If the American highway safety agency's move to define the behavioural competencies of self-driving cars as a test is about deciding what to ask, this is about who is able to mark the answers to questions already written.

Could the records a car keeps of itself answer the same question? A 2022 study analysed driver-initiated Tesla auto lane changes from naturalistic in-vehicle data, without any measurement of the interaction with the surrounding traffic. Data from inside the car tells you when the function was used and how often. It does not tell you the quantity the regulation uses for its verdict, namely how close the car in the adjacent lane was at the moment the lane marking was crossed. If the quantity to be measured is fixed, the instrument has to sit where that quantity can be measured.

Market surveillance is not the only target of this method. The paper states that LiDAR measurement is accurate enough for type approval, and sees it serving future revisions of R79 by showing what those revisions catch in the real operating domain. From the regulator's side, a record made outside the proving ground also indicates which combinations of speed and distance the next revision ought to put into the test programme.

The opposite error sits in the same dataset. Two manoeuvres were suppressed although the distance was sufficient. Nothing about that breaks the regulation, but it is a refusal that did not have to happen, and repeated often enough it teaches the driver not to trust the function. The paper does not settle the cause of those two and leaves it as conjecture. Which is another way of saying there is a region an observer outside the system cannot see into.

The paper is also clear about the lines it draws around itself. One vehicle was tested, and none of the manoeuvres classified as overshoots were perceived as dangerous by the testing personnel. So the result should be read not as a defect list for a particular manufacturer but as evidence that an operating region exists which the type approval test does not represent. The next steps the team names point the same way, improving LiDAR accuracy and temporal annotation and moving toward multi-sensor fusion and automated event detection, which would turn the procedure into a repeatable instrument for market surveillance. As it stands it is twenty-seven lines of record from a single campaign. The moment it becomes repeatable, it becomes a second document to set beside the certificate.

Primary Source

• 1.Cellina, M., Kriston, A., Migneco, A., Maggi, D., Favelli, S., Re, F., Minarini, F., Nuovo, A., Donà, R., & Ciuffo, B. (2026). "Beyond the Proving Ground: Independent Public-Road Testing of Assisted Lane Change Systems using LiDAR." arXiv:2608.26669.

Regulation

• 2.UNECE. (2023). "UN Regulation No. 79: Uniform Provisions Concerning the Approval of Vehicles with Regard to Steering Equipment." United Nations.

Related Work

• 3.Mattas, K., Vass, S., & Zachar, G. (2025). "Safety Perspective on Assisted Lane Changes: Insights from Open-Road, Live-Traffic Experiments." arXiv:2508.09233.
• 4.Viti, F., Hoogendoorn, S. P., Alkim, T. P., & Bootsma, G. (2008). "Driving Behavior Interaction with ACC: Results from a Field Operational Test in the Netherlands." 2008 IEEE Intelligent Vehicles Symposium.
• 5.Milanés, V., Shladover, S. E., Spring, J., Nowakowski, C., Kawazoe, H., & Nakamura, M. (2014). "Cooperative Adaptive Cruise Control in Real Traffic Situations." IEEE Transactions on Intelligent Transportation Systems, 15(1).
• 6.Gunter, G., Gloudemans, D., & Stern, R. E. (2021). "Are Commercially Implemented Adaptive Cruise Control Systems String Stable?" IEEE Transactions on Intelligent Transportation Systems, 22(11).
• 7.Noonan, T. Z., Gershon, P., Mehler, B., & Reimer, B. (2022). "Characterizing the Use of Tesla's Auto Lane Change Feature in Driver-Initiated Maneuvers." Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 66(1).