As manufacturers begin to introduce Level 3 automated driving systems to the market, the potential impact of such systems on overall traffic flow and capacity needs to be investigated. A key challenge arises from the fact that Level 3 systems require human drivers to take over control when reaching system limits, leading to so-called transitions of control (ToC), which may disrupt traffic flow and reduce road capacity. Despite regulatory advancements concerning Level 3 systems (R157 by UNECE (2023)), the macroscopic impact of such procedural ToC effects on traffic conditions remains insufficiently explored. This raises the general question of how Level 3 control transitions in conditionally automated vehicles (AVs) affect traffic capacity and, more specifically, what characteristics of procedural ToC-induced time headway increments in vehicle strings contribute to this effect. To investigate this, we conduct a large-scale simulation-based analysis and complement it with simplified numerical experiments to estimate macroscopic capacity impacts. Our study also explores the underlying mechanisms of the transition phase in greater detail. Existing research on potential capacity gains from AVs, as exemplified by Friedrich (2016) and Park et al. (2021), has primarily focused on higher automation levels (4–5) under optimistic assumptions of short time headways, e.g., τ A V = 0.5 s, in contrast to observed headways in manually driven vehicles of at least 1   s in freeway traffic, depending on vehicle speed, as shown by Wagner (2012). Our previous work in Alms et al. (2022) and Alms and Wagner (2024) touched on ToC-related capacity effects but lacked a comprehensive quantification of resulting capacity losses. This study addresses these gaps by (i) adopting a macroscopic perspective using realistic, R157-compliant time headways of τ A V = 1.6 s, and (ii) introducing an exploratory estimation approach that explicitly accounts for Level 3 disengagements in road capacity assessment.

The rest of the paper is organized as follows: Section 2 introduces the conceptual aspects of ToCs in Level 3 automated systems. In Section 3, we present a highway scenario calibration based on real-world detector data. Section 4 details our methodology for investigating ToC-related capacity effects in a simulation study, while Section 5 presents and discusses our results, comparing simulated and estimated capacity reductions. Lastly, Section 6 offers our perspective on the interpretation and limitations of this study.

The six levels of driving automation, defined by SAE International (2021), not only classify automated driving functions and capabilities but also specify the human driver’s role in terms of engagement and responsibility, as illustrated in Figure 1. Conditional automated driving (Level 3, highlighted with a purple frame in Figure 1) represents a fundamental shift toward automated vehicle operation within defined Operational Design Domains (ODD), specified in British Standards Institution (2020), allowing human drivers to disengage from the primary driving task. However, if the Level 3 system requires the driver to resume control, a takeover request (ToR) is issued, initiating a critical transfer of authority: these procedures are referred to as transitions of control (ToC, plural: ToCs). Detailed insights into various aspects of ToCs are available through a comprehensive literature review on takeovers in automated driving (McDonald et al., 2019). Further studies examine the intricacies of modeling human factors, such as situational awareness and task demand (Van Lint and Calvert, 2018; Calvert and van Arem, 2020), or reduced driver performance (Wang et al., 2025b), during ToCs.

The current regulations R157 from UNECE (2023) specify technical requirements for the certification of Level 3 Automated Lane Keeping Systems (ALKS) and set the time range T lead to 10   s before a failed transition escalates to a minimum risk maneuver (MRM), which is critical for the process of control transitions. Within the context of the EC project (TransAID, 2021; Lücken et al., 2019; Mintsis et al., 2019) introduced a novel ToC model, which is fully parametrizable to align with these later-established UNECE specifications and for which a detailed description of the model’s implementation is provided. The operationalization of ToCs is further specified in the ongoing EC project Hi-Drive (Bolovinou et al., 2023; Sauvaget et al., 2023) and demonstrated in Schulte-Tigges et al. (2023).

Figure 2 illustrates the basic mechanisms of the ToC model for successful and failed control transitions implemented in the microscopic traffic simulation SUMO (Alvarez Lopez et al., 2018). After a ToR, the AV enters a preparatory phase characterized by headway enlargement and disabled lane changing. Automated driving continues for the limited lead time, after which either the driver resumes control in time (successful transition), or, if not, the AV initiates an MRM (failed transition). For failed transitions, the AV initiates a phase of constant deceleration and may come to a full stop if the human driver does not respond. Although such events are rare, they can have a high impact and are the subject of extensive safety investigations based on disengagement reports (e.g., (Ward, 2024; Kohanpour et al., 2025)). However, this aspect is not the focus of the present work. In the case of a successful transition, the driver state model accounts for a phase of reduced human driving performance, with recent studies gaining further insights into both post-ToC durations (Wang et al., 2025a) and potential negative impacts on traffic stability (Wang et al., 2024).

In Maerivoet et al. (2019) and Lücken et al. (2019) principal transition phase effects of consecutive, quasi-synchronous ToCs in a platoon of Level 3 automated vehicles were previously demonstrated. Figure 3a, which depicts speed and time headways for a string of five AVs disengaging at the same location, illustrates this effect in a simplified simulation experiment with identical vehicle parametrization. The increased time headways, and consequently the cumulative speed reduction, are caused by the preparatory headway increment of the vehicle automation to facilitate a safe takeover (cf. Figure 2, Prep ToC Phase). Figure 3b extends this analysis by showing acceleration profiles for a larger platoon of up to 32 vehicles—the maximum size at which the last AV still manages to prevent a complete stop—using SUMO’s ACC model for AVs, based on Xiao et al. (2017). The main observed effects in the vehicle decelerations include:

• With a moderate default deceleration of 1   m/s 2 during the 10   s transition phase specified by R157, maintaining safe gaps in AV platoons is not feasible without initiating deceleration earlier. Panel (b-2) in Figure 3b illustrates that, starting with the first vehicle behind AV.1 (dark blue line), SUMO’s gap controller begins to decelerate even before the respective vehicles receive ToRs to initiate their ToC.

These numerical experiments are highly simplified due to identical vehicle parametrizations, yet they effectively illustrate the isolated ToC effects discussed. Given the cumulative deceleration patterns observed, we expect that ToC-induced disturbances may lead to noticeable reductions in traffic capacity. To examine whether these effects also manifest under more realistic traffic flow conditions, we calibrate a SUMO simulation scenario to detector data in Section 3.

To analyze the impact of ToCs on traffic capacity, we use real-world detector data from a German highway west of Berlin as a reference for SUMO calibration. The following sections detail the dataset and simulation setup.

To determine ToC-related capacity impacts, we conduct a simulation study with an increasing AV penetration rate and measure the corresponding maximum flows q . Additionally, we estimate the anticipated ToC-induced capacity reduction and later compare these estimates with the measured results from the simulation study.

Figure 12 presents the overall results obtained from the simulation study outlined in Section 4.1. First, we find that all maximum flows in the AV00 share, ranging between 3405 − 3611   veh/h , fall within the expected capacity range from the calibration, i.e., 3281 − 3647   veh/h . We further analyze these results in detail for the four scenarios, following the order in which they were previously defined:

1. No ToCs: The results (green line) show that up to share AV40, maximum flows remain relatively stable, with a reduction of approximately 100   veh/h compared to AV00. Beyond AV50, flow values begin to increase again. This trend can be linked to a homogenization effect in traffic flow as AV shares grow, which is influenced by the AV parametrization—-specifically, the absence of sigma and a very small speedDev value of 0.01.

Overall, the results in Figure 12 show that in the unmanaged scenario, maximum flow declines significantly with increasing AV share. At AV85, the max flow is approximately 500   veh/h lower than in the managed case—indicating that ToC management measures could help alleviate, but not fully prevent, capacity losses.

Furthermore, to compare these simulation results with the theoretical lane capacities estimated in Section 4.2 and Figure 10c, we derive the relative percentage reductions in capacity across the increasing AV share. Figure 13 summarizes these reductions for the unmanaged scenario, differentiating between the estimators max and mean, while constant is included as a reference that ignores ToC effects. For each estimator, we report the root mean squared error (RMSE) and the coefficient of determination (R ²) to quantify the goodness of fit to the simulated capacity reductions—where lower RMSE and R ² values closer to one indicate better agreement with the simulation data. The simulated results reveal a capacity loss of up to nearly 60% at AV85. We also make the following observations:

• While the τ M V dependency is relatively small in the simulation results, it becomes increasingly important in the estimator outcomes.

In summation, the capacity reductions observed in the simulation might align only unsatisfactorily with theoretical estimates, as deviations occur in the mid-range AV shares due to the simplified assumptions of the estimators. This partial mismatch is also reflected in the RMSE and R ² values, for which no established benchmarks exist in this context. Therefore, our assessment of estimator performance focuses on relative differences and qualitative trends within the observed results. However, the estimator max performs best in comparison to the simulation results, substantiating our suspicion that the maxima in time headway increments dominate ToC-related capacity effects. Nevertheless, the overall findings from Figures 12, 13 highlight the potential ToC effects on capacity reductions across various scenarios and parameter dependencies, in line with the stated expectations.

To investigate ToC-related capacity reductions, we conducted comprehensive simulation experiments with a calibrated two-lane highway scenario, as well as numerical experiments to estimate the large-scale impact of time headway increments during consecutive control transitions. Our main findings can be summarized as follows: (i) capacity reductions of up to 2000   veh/h , corresponding to approximately 60 % loss, were observed in shares with near-full Level 3 automation but no traffic management coordination; (ii) ToC effects became notably impactful starting from a Level 3 share of 10 % to 20 % ; (iii) a coordination of ToCs could mitigate losses by roughly 1000   veh/h or 30 % ; (iv) binding Level 3 operation to the rightmost lane resulted in the most severe reduction, with up to 2660   veh/h or 75 % loss; and (v) maxima in time headway increments during ToCs emerge as the dominant factor contributing to these capacity effects.

Several relevant limitations should be acknowledged, as they may affect the applicability and interpretation of our findings. Recent research on data from Level 4 AVs reveals reduced time headways when MVs follow AVs (Jiao et al., 2024). Such effects, which might also apply to Level 3 systems, are not considered in this study. An additional aspect that has not yet been discussed is the impact of human response times for non-emergency ToCs. Throughout this investigation, the response time distribution was kept the same, at μ = 7   s in all simulations, in line with our previous studies. More recent data from real-world tests presented by Pipkorn et al. (2023) indicate response times closer to 5 s, from which the authors infer that a lead time of 10   s , as specified in the R157, should be feasible for human drivers to take over in time. In our simulations, a few random sample reruns with these lower response times indicated approximately 10– 20 % higher capacities compared to the results presented here. A further limiting factor might be SUMO’s ACC model, which is parametrized for full string stability and deployed here as a proxy for Level 3 automated vehicles, although experimental studies and theoretical work have demonstrated string instabilities in ACC-equipped platoons, as we also discussed in Alms and Wagner (2024). All these limitations could potentially affect traffic capacity, though their precise contribution cannot be reliably quantified at this stage.

Future work should therefore include the development of a more accurate Level 3 model in SUMO, for example, an ACC-based ALKS system, as well as a systematic investigation of how model assumptions and human response variability together affect traffic capacity. Another important direction is to further examine the effects of MRMs, which are relevant for failed Level 3 transitions and Level 4 automation, on overall traffic, especially if they are not managed properly.

Lastly, we would like to reflect on the broader capacity implications of AVs. Our overall vehicle parametrization inherently results in slightly reduced theoretical capacities—even without ToCs—due to the implementation of lower time headways for MVs and higher ones for AVs, which contrasts with assumptions commonly made in other studies. While experimental research has demonstrated counterbalancing effects at high AV shares, which our own simulations also imply, this effect is diminished in the context of Level 3 systems. Unlike Level 4 or CACC-equipped vehicles, Level 3 automation, in its current form, does not typically support the low time headways often assumed to contribute to capacity gains. However, practical capacity impacts at relevant market penetration rates between 10% and 20% are likely still many years away, leaving room for further technical and regulatory development of Level 3 systems. Yet, in combination with the ToC-related capacity constraints demonstrated in this study, we take a more cautious view and do not share the seemingly widespread optimism regarding beneficial capacity effects of AVs in the mid-term.