The lane-changing tAV consistently aimed for roughly three-quarters of a second behind its new leader—and three-quarters of a second in front of its follower. Not because it started there. Not because the traffic demanded it. But because somewhere in its code, something in its algorithms, something in its understanding of a safe merge apparently said: this is where you need to be.
That convergence is the most striking finding from a new dataset out of North Carolina, and it's one that should make anyone who drives near autonomous vehicles pay attention. Researchers at North Carolina State University put four instrumented vehicles on a public road in Apex, programmed three of them to make their own lane-changing decisions, and watched what happened 78 times. What they found challenges a common assumption: that autonomous vehicles, left to their own devices, would behave erratically, unpredictably, or differently from humans in ways that make roads less safe.
The opposite may be true. The tAVs—the term stands for transitional autonomous vehicles, meaning systems that can make and execute lane-change decisions without driver input, though a human still supervises—didn't scatter across a wide range of behaviors. They converged. And understanding why they converged, and what that convergence means for the cars that share your highway, is the story this paper tells.
The NC-tALC dataset, as the researchers call it (North Carolina Transitional Autonomous Vehicle Lane-Changing), is not just a collection of GPS traces. It's a precise, millimeter-accurate record of how autonomous vehicles navigate one of the most dangerous maneuvers in driving. And it arrives at a moment when more and more vehicles on American roads—Teslas with Full Self-Driving, Cadillacs with Super Cruise, Fords with BlueCruise—are making those decisions without you.
The Science
The question the researchers wanted to answer was deceptively simple: what does an autonomous vehicle actually do when it needs to change lanes? Not what the marketing says. Not what the simulations predict. But what actually happens on a real road, with real traffic, when a machine decides it's time to move.
Previous research had looked at this question, but with important limitations. The Waymo Open Motion Dataset captures Level 4 autonomous vehicles—true robots, no human backup—operating in Phoenix and San Francisco. Other work has examined driver-initiated assisted lane changes, where a human decides to merge and the car simply executes the maneuver. What was missing was the middle ground: vehicles operating at SAE Level 2 and above that can make the decision and execute it, while a human watches. These are the tAVs.
To study them properly, the researchers needed control. Real-world traffic is chaotic. You can't isolate the effect of an autonomous vehicle's behavior if other drivers are behaving unpredictably around it. So the team went to Sunset Lake Road in Apex, North Carolina—a public roadway with a posted speed limit of 45 mph and a distinctive feature: an exclusive right-turn lane roughly 170 meters long, which creates a mandatory merge scenario. The tAV had to get out of that turn lane and into the through lane. It wasn't a choice. It was a requirement.
Four vehicles ran each trial. One—designated X—was the lane changer, operating in tAV mode. Three others formed a platoon in the target lane: one leader (A) and two followers (B and C). Vehicles B and C also operated in tAV mode, creating a traffic environment where the autonomous system had to interact with other autonomous systems, not just human drivers. The researchers controlled the initial spacing between X and the target lane leader, positioning X in one of three starting locations: near the leader, near the first follower, or roughly centered between them. They then activated the tAV's automation, sat back, and let the machine decide how to merge.
The vehicles were instrumented with RTK-GNSS/INS units—real-time kinematic Global Navigation Satellite Systems integrated with inertial navigation. These aren't your phone's GPS. They're the kind of equipment surveyors use to map construction sites, capable of pinpointing a vehicle's position to within a few centimeters and measuring its speed to within 0.01 meters per second. Each vehicle had one of these units mounted at its center, feeding data at 20 Hz—twenty position updates every second. The result was a dataset of extraordinary precision: every wobble of the steering wheel, every surge of acceleration, every negotiation of gaps captured in enough detail to reconstruct the maneuver frame by frame.
They ran 78 trials. The tAV selected Gap 1—the space between the first follower and the second follower in the target lane—in 52 of them. Gap 0 (ahead of the target lane leader) was chosen 16 times. Gap 2 (between the second follower and third) was chosen 10 times. Gap 3 (behind everyone) was never selected. These weren't random. The tAV had preferences, and those preferences were remarkably consistent.
Target Gap Selection by tAV
| Label | Value |
|---|---|
| Gap 0 (Ahead of A) | 16 |
| Gap 1 (Between A and B) | 52 |
| Gap 2 (Between B and C) | 10 |
| Gap 3 (Behind C) | 0 |
The researchers then tagged each trial with nine key timestamps: activation (when the tAV took over), lane-change start (first lateral movement), left-edge touching (when the vehicle's front corner first crossed into the new lane), lane-change crossing (when the vehicle center entered the new lane), lane-change end (when the vehicle had fully settled into the new lane), and four post-completion timestamps at 2.5, 5, 7.5, and 10 seconds afterward. This granular timeline let them watch not just whether the merge succeeded, but how the vehicle's relationships with others in the traffic stream evolved throughout the process.
The study was conducted in summer 2025, under clear skies, on a public road with live traffic. This wasn't a closed test track. The vehicles had to contend with other drivers, with the variability of the real world, with all the messiness that makes field experiments difficult but also meaningful. The roadway had a grade of less than 2%, was predominantly straight, and had lane widths of 3.7 to 4.0 meters—standard American highway dimensions. The conditions were within the operational design domain of the tested tAVs, meaning the automation considered this kind of driving routine, not edge case.
What They Found
The headline finding is this: despite starting from very different positions within their target gaps, the tAVs arrived at lane crossing with remarkably similar lead and lag gaps. At the moment when the vehicle center crossed into the target lane—the lane-change crossing timestamp, or LCC—the lead gap (the time headway between the tAV and the vehicle ahead in the new lane) consistently exceeded 0.54 seconds, and the lag gap (the time headway to the vehicle behind) consistently exceeded 0.65 seconds. The median lead gap at LCC was 0.75 seconds. The median lag gap was similar.
To understand what this means, consider the alternative. If the tAV had simply maintained its initial position, the lead and lag gaps at LCC would have varied wildly, reflecting the different starting conditions. Some trials would have had the tAV merging with almost no gap ahead and a large gap behind. Others would have had the opposite. The data would have been scattered. But that's not what happened. Instead, all three groups—those that started near the leader, those that started near the follower, those that started in the middle—converged toward the same narrow range.
The near-leader group showed the most dramatic adjustment. At activation, the median lead gap in this group was negative—meaning the tAV was ahead of the vehicle it would need to get behind. A negative lead gap isn't a mergeable situation; you can't slot in front of someone if you're already in front of them. Yet between activation and lane crossing, the tAV adjusted. The median lead gap jumped from -0.24 seconds to 0.75 seconds. The tAV found a way to get behind its target leader.
The near-follower group showed the mirror image. At activation, these trials had small lag gaps—meaning the tAV was close to the vehicle it would need to stay ahead of. By lane crossing, that lag gap had widened to a similar 0.75-second median. The adjustment went the other direction, but the destination was the same.
Even the near-center group, which started with a relatively balanced lead-lag split, showed the pattern. Their gaps didn't need to change as dramatically, but they still moved toward that same target range.
This convergence behavior—different inputs, similar outputs—has implications. It suggests the tAV isn't simply executing a pre-programmed motion based on where it starts. It's actively managing the relative positions of vehicles around it, adjusting in real time to achieve something close to a preferred state. The researchers describe this as the tAV adjusting