PaperlayerView exact revision on arXiv

Source-linked AI summary

Pre-Lane-change Signal in Transitional Autonomous Vehicles: Results from Controlled Experiments

Zeyu Mu, Danjue Chen, Abhinav Sharma, George F. List

arXiv:2609.02575v1cs.ROeess.SY

TL;DR

Production tAV lane-change decisions are difficult to observe separately from lateral execution. Using 150 controlled mandatory lane changes, this paper defines an observable pre-LCS signal, predicts eventual target-gap choice, and examines longitudinal preparation before lateral movement; it finds substantial gap-choice information at SigT and preliminary evidence of distinct pathways for in-position and repositioning cases.

  • Problem

    Empirical understanding is limited regarding how production tAVs develop lane-change decisions and whether target-gap choice can be distinguished from lateral maneuver execution.

  • Method

    The study analyzes 150 controlled NC-tALC mandatory lane changes, defines SigT as an operational pre-LCS reference point, and examines gap choice and longitudinal progression to LCS.

  • Results

    The traffic state at SigT contains substantial information about eventual target-gap choice, while in-position and repositioning cases show preliminary evidence of different longitudinal pathways to LCS.

  • Takeaways & Limitations

    The findings support modeling longitudinal preparation before lateral execution and applying gap-choice analysis to in-position, repositioning, and longitudinally overlapping cases.

  • Takeaways & Limitations

    The results come from controlled experiments with limited scenarios and should not be generalized directly to other tAV systems, software versions, roadways, or traffic conditions.

Abstract

from arXiv · show

This paper investigates how a production transitional autonomous vehicle (tAV) develops and executes mandatory lane-change decisions. Using 150 controlled mandatory lane changes from the NC-tALC experiments, the study examines whether the eventual target gap is observable before lateral movement begins and how the tAV progresses longitudinally from that pre-lane-change state to lane-change start. Signal time (SigT) is defined as an operational pre-lane-change-start reference point. A Firth logistic regression predicts whether the tAV eventually merges in front of or behind its nearest target-lane vehicle using relative position and relative speed at SigT. Longitudinal progression from SigT to lane-change start is then examined separately for in-position and repositioning cases. The traffic state at SigT contains substantial information about eventual target-gap choice and provides meaningful lead time before lateral movement begins. The proposed formulation predicts whether the tAV remains with its current gap or repositions to a neighboring gap by moving forward or dropping back, including cases with longitudinal overlap and ambiguous current-gap geometry. The model achieves an average five-fold cross-validated accuracy of 0.89. Results also provide preliminary evidence that in-position and repositioning cases follow different longitudinal pathways from SigT to lane-change start. These findings support a two-stage conjecture of the observable lane-change process: longitudinal preparation from SigT to lane-change start, followed by lateral maneuver execution. The formulation applies to in-position, repositioning, and longitudinally overlapping cases, and can support lane-change models that distinguish target-gap choice from lateral-onset timing while representing longitudinal preparation before lateral movement begins.

Zeyu Mu

Production tAV lane-change research has limited empirical grounding, and existing approaches often conflate target-gap choice with lateral-movement onset. This study frames lane changing as an observable process that separates these components.

  • Motivation: Production tAV lane-change behavior has not yet been empirically examined using high-resolution controlled trajectories.The NC-tALC experiments systematically varied initial spacing and relative speed during mandatory lane changes.
  • Related Work: Most lane-change models formulate the decision as accepting or rejecting the current gap, while multi-gap and interactive approaches remain less common.Existing frameworks include critical-gap, utility-based, game-theoretic, and multi-gap selection models.
  • Related Work: Prior studies often operationally equate the lane-change decision with lateral-movement onset because the true decision is unobservable in trajectory data.Some work conceptually separates gap selection from maneuver execution, but does not directly observe or test gap selection.
  • Study Objective: The study analyzes 150 controlled mandatory lane changes to test whether target-gap information appears before lateral movement and how longitudinal preparation proceeds to lane-change start.The analysis uses the NC-tALC datasets and addresses two process-level questions.
  • Study Objective: The proposed two-stage conjecture separates longitudinal preparation before lane-change start from subsequent lateral maneuver execution.The paper also conjectures that target-gap decision making may precede longitudinal preparation, but does not evaluate that possibility because of data limitations.

DATA

The dataset contains controlled mandatory merges by a production tAV under systematically varied initial conditions, with synchronized trajectory measurements and operationally defined lane-change events. These data support analysis of signal time, gap choice, repositioning, and maneuver initiation.

  • Dataset: 150 mandatory lane-change trajectories come from controlled NC-tALC field experiments in which subject vehicle X merged into an adjacent target lane.Vehicles A, B, and C occupied the target lane, with A as leader and B and C as potential followers.
  • Experimental Design: Initial longitudinal position and relative speed of X were varied at automation activation, ACT, after manual positioning established prescribed conditions.The automated driving system then independently conducted the mandatory lane change under continuous driver supervision.
  • Experimental Design: The experiments include 78 cases with tAV followers and 72 cases with ACC followers.Together these conditions yield 150 lane-change cases.
  • Measurements: RTK-GNSS/INS instrumentation recorded synchronized position, speed, heading, and acceleration at 20 Hz.Processed trajectories were transformed into a common road-based coordinate system for subsequent analyses.
  • Key Events: SigT is an operational pre-LCS reference: it is the observed delayed switch time in 23 cases and ACT otherwise.SigT is treated as an observable signal associated with target-gap information and longitudinal preparation, not as the confirmed internal target-gap decision time.
  • Key Events: LCS marks lateral-movement onset, while LET, LCC, and LCE mark first target-lane entry, center crossing, and maneuver completion with stabilization.Figure 2 presents these timestamps alongside ACT and post-LCE observation times.

OBSERVATIONS AND THE TWO-STAGE CONJECTURE

The paper proposes an observable two-stage lane-change process: longitudinal preparation before lateral movement, followed by lane-switching execution. SigT provides a pre-LCS reference for studying whether target-gap choice is already informative before lateral onset.

  • Signal time: 23 of 150 cases exhibit a sustained longitudinal kinematic switch at least 0.3 s after ACT, while the eventual gap generally remains unchanged after the switch.
  • Signal time: SigT is the observed delayed kinematic switch, or ACT when no significant delay occurs, and is treated as a reproducible pre-LCS signal rather than the exact internal gap-decision time.
  • Two-stage conjecture: Before LCS, the tAV primarily adjusts longitudinal position and speed relative to target-lane vehicles; after LCS, lateral movement proceeds with continued longitudinal adjustment until LCE.
  • Two-stage conjecture: A median pre-LC duration of 3.83 s for Gap=2 indicates that longitudinal preparation can be substantial before lateral movement begins.
  • Target-gap information: The study tests whether the SigT traffic state predicts eventual target-gap choice separately from the timing of lateral onset.
  • Target-gap information: In 76% of merges, X is already within its final gap at SigT; the remaining 24% reposition longitudinally to a neighboring gap.
  • Target-gap information: The gap outcome is represented relative to the nearest target-lane vehicle as merging in front of or behind it, using relative position and relative speed at SigT.

SigT-Based Front/Behind Model

At SigT, relative position and relative speed predict whether the tAV eventually merges in front of or behind its nearest target-lane vehicle. Position supplies the dominant signal, while speed resolves near-aligned cases.

  • The Firth logistic regression predicts the eventual front/behind outcome from X–N time gap and relative speed at SigT.The outcome is 1 for merging in front of N and 0 for merging behind N.
  • Relative position provides the dominant predictive information in the standardized model.The larger standardized position coefficient indicates greater predictive contribution than relative speed.
  • 0.97 AUC is achieved by the full model, compared with 0.87 using position alone.Relative speed primarily resolves ambiguous cases when X and N are nearly aligned.
  • 0.89 average held-out accuracy is achieved in five-fold cross-validation.Each fold trains on 120 cases and tests on 30, so all 150 cases are tested once.
  • 0.90 accuracy is achieved by leave-one-out cross-validation, correctly classifying 135 of 150 cases.The leave-one-dataset-out accuracy is 0.87.
  • Relative speed is especially informative when X and N are nearly aligned or when X subsequently crosses N.The sign of relative position generally predicts which side of N X eventually occupies.

In-Position and Repositioning Cases

The analysis distinguishes cases already positioned in the eventual gap from cases that reposition across a target-lane vehicle. Repositioning direction corresponds to relative speed and reveals distinct longitudinal pathways.

  • An in-position case places X between the leader and follower of its eventual target gap at SigT.A repositioning case begins in a different gap region and subsequently crosses a target-lane vehicle into a neighboring gap.
  • 36 of 150 cases are repositioning, while 114 are in-position.Among repositioning cases, 30 are dropbacks and six are move-ups.
  • 30 dropbacks move X backward across a vehicle, whereas six move-ups move X forward across one.Dropbacks include moves into the A–B or B–C gaps; move-ups enter the front-A or A–B gap.
  • Dropbacks have a median relative speed of −0.6 m/s, while move-ups have a median of 2.5 m/s.The direction of repositioning is consistent with X’s relative speed to the crossed vehicle.
  • Relative speed helps explain how X reaches an eventual neighboring gap when its SigT gap occupancy differs from the final target gap.These cases show why the SigT state can predict a final gap that X does not initially occupy.

Lead Time of Gap-Choice Information

Gap-choice information is often observable at SigT before lateral movement begins. The measured lead interval is substantial in a notable share of the controlled mandatory lane changes.

  • The lead interval is defined as LCS − SigT, measuring how early gap-choice information precedes lateral movement.SigT is the pre-lane-change reference state and LCS marks lane-change start.
  • 38.7% of cases have LCS − SigT > 2 s, 24% exceed 3 s, and the median is 1.3 s.These values quantify the observed availability of gap-choice information before LCS.
  • The results indicate that substantial gap-choice information is systematically observable at SigT before LCS.The paper treats this as evidence that target-gap choice is resolved before lateral movement begins.
  • The reported fractions apply to controlled experimental scenarios and should not be interpreted as naturalistic rates.Because SigT equals ACT in most cases, the observed lead interval may be a lower bound on earlier availability if automation had activated sooner.

Implications Relative to Conventional LC Decision Models

Conventional lane-change models typically evaluate acceptance of a current candidate gap, but that representation becomes unclear during longitudinal overlap and cannot capture some neighboring-gap choices. The SigT-based front/behind formulation provides a common outcome representation across these cases.

  • Conventional gap-acceptance and game-theoretic models usually evaluate whether X accepts or rejects the current candidate gap.The conventional candidate gap generally requires positive lead and lag bumper-to-bumper clearances.
  • Longitudinal overlap creates unclear zones because the bumper-to-bumper clearance with an overlapping vehicle is negative.In these zones, conventional models may not clearly define the current candidate gap.
  • The front and behind outcomes remain separated by the decision boundary within the unclear zone.This indicates that eventual gap choice remains predictable even when current-gap geometry is ambiguous.
  • At SigT, X can be ahead of N yet eventually merge behind it, or behind N yet eventually merge ahead.These outcomes are consistent with evaluating a neighboring gap and repositioning longitudinally.
  • The SigT-based formulation covers in-gap, repositioning, and longitudinally overlapping cases with one front/behind target-gap representation.It therefore spans straightforward, surprising, and geometrically unclear cases.

LONGITUDINAL PREPARATION FROM SIGT TO LCS

From SigT to LCS, the tAV’s longitudinal preparation is examined separately for in-position cases using safety margins within the selected gap. Lateral movement often begins before those margins become positive, but no sharp onset threshold is established.

  • In-position cases: 114 in-position cases place X between the leader and follower of its eventual target gap at SigT.
  • Safety-margin measure: After emergency spacing (AES) measures whether longitudinal spacing is sufficient to avoid a rear-end collision under hypothetical emergency braking.Negative AES indicates insufficient projected spacing, while larger AES indicates a greater safety margin.
  • Safety-margin measure: The binding AES is the more restrictive safety margin between the selected-gap leader–X and X–selected-gap follower pairs.
  • Lateral onset: Approximately −3 m: the binding AES rises most sharply and has the same median at LCS, although the distribution remains dispersed.Lateral movement therefore often begins while the binding AES is slightly negative rather than positive.
  • Lateral onset: Approximately 0.71 AUC: binding AES is the strongest individual indicator of LCS for in-position cases; additional variables reach approximately 0.74–0.76 AUC.These values indicate moderate rather than strong discrimination of exact lateral-onset time.

Repositioning Cases

Repositioning cases require X to move longitudinally across a target-lane vehicle before entering its eventual target gap. Signed bumper-to-bumper clearance is the main longitudinal measure examined at LCS.

  • Repositioning cases: 36 repositioning cases require X to cross a target-lane vehicle before entering the eventual target-gap region.
  • Clearance measure: Signed bumper-to-bumper clearance c_XN is positive after X clears the crossed vehicle N and negative while the vehicles still overlap longitudinally.
  • Lateral onset: Approximately −4 m: lateral movement begins most sharply as clearance to N approaches this value, with the same median at LCS.X often still overlaps longitudinally with N when lateral movement begins, and the distribution is not a sharp threshold.
  • Prediction evidence: Approximately 0.84 AUC: signed bumper-to-bumper clearance is the strongest individual LCS indicator for repositioning cases; expanded specifications reach approximately 0.84–0.87 AUC.The repositioning group contains only 36 cases.

Comparison, Summary, and Future Research

In-position and repositioning cases show different longitudinal progressions toward LCS, but neither group yields a single clear lateral-onset condition. Additional state variables, temporal information, and larger repositioning samples are needed.

  • Comparison: In-position cases change safety margins within the selected gap, whereas repositioning cases move longitudinally relative to the crossed vehicle.
  • Limitations: The observed relationships do not establish one clear lateral-onset condition for either group.
  • Future research: Future research should examine additional state variables, pre-LCS temporal evolution, and larger repositioning samples.These directions aim to determine whether a more consistent lateral-onset pattern can be identified.

CONCLUSIONS AND FUTURE WORK

Using 150 controlled mandatory lane changes, the study tests whether target-gap information is available before lateral movement and characterizes longitudinal progression from SigT to LCS. The results support a two-stage observable process, while controlled-scenario limits constrain generalization.

  • Study scope: 150 controlled mandatory lane-change trajectories from NC-tALC experiments provide the basis for examining pre-LCS target-gap information and longitudinal progression.
  • Main findings: SigT contains substantial information about eventual target-gap choice and provides meaningful lead time before lateral movement begins.
  • Main findings: The formulation predicts staying with the SigT gap or repositioning to a neighboring gap, including longitudinal overlap and ambiguous current-gap geometry.
  • Process conjecture: The findings support a two-stage conjecture: longitudinal preparation from SigT to LCS followed by lateral maneuver execution from LCS to LCE.
  • Main findings: In-position and repositioning cases show preliminary evidence of different longitudinal pathways, associated respectively with a surrogate safety measure and bumper clearance.
  • Limitations: Controlled experiments covered limited scenarios rather than naturalistic driving, so findings should not be generalized directly to other tAV systems, software versions, roads, or traffic conditions.Repositioning cases were also a small fraction of the sample.
  • Future work: Further studies should extend observation before LCS, test broader initial conditions, generate more diverse gap choices, and examine additional tAV systems.
Loading 2609.02575v1…