Overtake Score: Definitions and Applications

• Overtake Score is defined differently across studies: as a sliding-window posterior probability from CAN data and as a benchmark metric for route-level instruction compliance.
• Truck-based methods employ classifiers like ANN, RF, and SVM to produce calibrated probabilities, with performance evaluated via metrics such as TPR, TNR, and AUC-PR.
• In autonomous driving, Overtake Score aggregates binary success indicators to measure interaction compliance, highlighting challenges in early prediction and scenario variability.

Searching arXiv for the cited papers and closely related work on Overtake Score. “Overtake Score” denotes different but related quantities in recent driving research. In truck overtake prediction from Controller Area Network (CAN) data, it is the calibrated posterior probability that a sliding-window feature vector corresponds to an overtaking event, or an aggregate of such probabilities over a file (Butt et al., 2024, Alonso-Fernandez et al., 1 Jul 2025). In desired-speed conditioned autonomous driving, the same term refers instead to a route-level command-compliance metric: the percentage of overtake/follow scenarios in which the policy executes the commanded interaction outcome (Shao et al., 26 Mar 2026). A plausible implication is that the term is not yet standardized across the literature; its meaning depends on whether the task is maneuver detection from logged signals or instruction-following evaluation in closed-loop driving.

1. Terminological scope

The recent literature uses “Overtake Score” in at least three task-specific senses.

Context	Object being scored	Definition
Truck CAN-window classification	Window $X_t$	$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$
Truck file-level classification	File with $K$ windows	$S=\frac{1}{K}\sum_{k=1}^K \hat p_k$
Desired-speed conditioned autonomous driving	Set of routes $R$	$100\times \frac{1}{|R|}\sum_{r\in R}s_r$

In “Predicting Overtakes in Trucks Using CAN Data,” each classifier $$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$ outputs a score

$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$

described as the “Overtake Score” for window $t$ (Butt et al., 2024). In the 2025 comparative study, the same basic quantity is produced per sample, but the score reported for a file is the arithmetic mean over all sliding-window samples in that file (Alonso-Fernandez et al., 1 Jul 2025). In Bench2Drive-Speed, by contrast, Overtake-Score is a benchmark metric computed from binary route-level success indicators rather than from classifier posteriors (Shao et al., 26 Mar 2026).

This distinction is consequential. In the truck papers, the score is a calibrated confidence signal that can be thresholded for detection. In Bench2Drive-Speed, it is already an aggregated evaluation statistic in percent. Confusing these two usages would conflate posterior estimation with benchmark-level task success.

2. CAN-based overtake score as a posterior probability

In the 2024 truck study, the score is defined over feature vectors extracted from 1 s windows of CAN data ending at time $t$. Let $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$0 be the $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$1-dimensional feature vector extracted from the CAN bus over the 1 s window ending at time $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$2. The logger records 10 Hz signals, so a 1 s window has $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$3 samples, and windows advance every 0.5 s, yielding 21 windows per file (Butt et al., 2024).

For each continuous signal $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$4 within window $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$5, the paper computes

$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$6

For each categorical signal $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$7, the modal value is taken: $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$8 This yields $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$9 features per window (Butt et al., 2024).

The paper configures all four classifiers to output calibrated probabilities. For the ANN, the output layer is a 2-neuron softmax, so

$K$0

where

$K$1

The architecture is input $K$2, hidden $K$3, output $K$4 (Butt et al., 2024).

For Random Forest,

$K$5

where $K$6 is the vote of tree $K$7 for class $K$8, with $K$9 trees, bootstrap samples, and default splits (Butt et al., 2024).

For SVM, both a linear kernel and a Gaussian RBF variant are used. Their decision function $S=\frac{1}{K}\sum_{k=1}^K \hat p_k$0 is mapped to a probability by Platt scaling or built-in MATLAB probability estimates: $S=\frac{1}{K}\sum_{k=1}^K \hat p_k$1 For the linear case, $S=\frac{1}{K}\sum_{k=1}^K \hat p_k$2; for the RBF case, it is the kernel decision function (Butt et al., 2024).

Preprocessing is classifier-dependent. For ANN and both SVMs, the per-feature normalization is

$S=\frac{1}{K}\sum_{k=1}^K \hat p_k$3

with per-feature mean and standard deviation computed on the training set. RF operates on the unscaled features (Butt et al., 2024).

3. Temporal evolution, thresholding, and fusion in early overtake prediction

The 2024 study investigates the score up to 10 s before the overtaking event. Figure 1 shows that for class $S=\frac{1}{K}\sum_{k=1}^K \hat p_k$4 samples, the median $S=\frac{1}{K}\sum_{k=1}^K \hat p_k$5 rises from approximately $S=\frac{1}{K}\sum_{k=1}^K \hat p_k$6 at $S=\frac{1}{K}\sum_{k=1}^K \hat p_k$7 s to approximately $S=\frac{1}{K}\sum_{k=1}^K \hat p_k$8 at $S=\frac{1}{K}\sum_{k=1}^K \hat p_k$9 s for ANN, RF, and SVML. For RF specifically, the median $R$0 for class $R$1 is reported as $R$2 at $R$3 s, $R$4 at $R$5 s, $R$6 at $R$7 s, and $R$8 at $R$9 s. By contrast, the no-overtake scores remain near $100\times \frac{1}{|R|}\sum_{r\in R}s_r$0–$100\times \frac{1}{|R|}\sum_{r\in R}s_r$1 for RF and ANN and oscillate for $100\times \frac{1}{|R|}\sum_{r\in R}s_r$2 (Butt et al., 2024).

This temporal behavior underlies the paper’s main finding: the prediction scores of the overtake class tend to increase as the trigger approaches, while the no-overtake class remains stable or oscillatory depending on the classifier. The best accuracy is therefore achieved when approaching the trigger, and early overtaking prediction is characterized as challenging (Butt et al., 2024).

The paper also defines a score-level fusion of RF and linear SVM: $100\times \frac{1}{|R|}\sum_{r\in R}s_r$3 A single threshold $100\times \frac{1}{|R|}\sum_{r\in R}s_r$4 yields the final decision,

$100\times \frac{1}{|R|}\sum_{r\in R}s_r$5

In practice, $100\times \frac{1}{|R|}\sum_{r\in R}s_r$6 is chosen to maximize $100\times \frac{1}{|R|}\sum_{r\in R}s_r$7, with typical values in $100\times \frac{1}{|R|}\sum_{r\in R}s_r$8; early ($100\times \frac{1}{|R|}\sum_{r\in R}s_r$9 s) the optimum $$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$0 is approximately $$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$1–$$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$2, and closer to the trigger it shifts upward slightly (Butt et al., 2024).

The paper defines

$$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$3

and

$$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$4

At $$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$5 s, RF alone at $$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$6 achieves $$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$7 and $$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$8, whereas the fusion at $$C\in\{\text{ANN},\text{RF},\text{SVML},\text{SVM}^{\text{rbf}}\}$$9 gives $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$0 and $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$1 (Butt et al., 2024). The classifiers show good accuracy in classifying overtakes, with Recall/TPR $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$2, but no-overtake classification is weaker, with TNR typically $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$3–$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$4 and below $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$5 for one SVM variant; fusion improves no-overtake classification at the expense of reducing overtake accuracy, while keeping the latter above $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$6 near the trigger (Butt et al., 2024).

AUC-PR results at $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$7 s further quantify this trade-off:

Classifier	$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$8	$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1],$9	$t$0
ANN	0.931	0.914	0.907
RF	0.896	0.885	0.890
SVML	0.952	0.950	0.946
RF+SVML	0.981	0.981	0.975

The full table also reports $t$1 values of $t$2 for ANN, $t$3 for RF, $t$4 for SVML, $t$5 for $t$6, and $t$7 for RF+SVML (Butt et al., 2024).

4. File-level aggregation and comparative methodology across trucks

The 2025 comparative study reformulates the score at file level. The classifier produces per-sample scores $t$8, $t$9, for the $t$0 sliding-window feature vectors extracted from one file, and the file-level score is

$t$1

A final binary decision is then obtained by thresholding the file score, for example $t$2 overtake and $t$3 no-overtake (Alonso-Fernandez et al., 1 Jul 2025).

The data are cropped around a trigger. Let $t$4 be the moment the trigger fires; each file is cropped from $t$5, and since samples are recorded at 10 Hz, each file yields 60 raw timesteps (Alonso-Fernandez et al., 1 Jul 2025). The trigger is defined by

$t$6

Preprocessing differs by classifier. A sliding window of length $t$7 seconds with 50% overlap is run over the crop, with $t$8 s for ANN and SVM, and $t$9 s for RF. For the seven continuous channels—accelerator pedal position, distance to lead vehicle, lead-vehicle speed, relative wheel speed, ego-speed, lateral acceleration, and longitudinal acceleration—the mean and standard deviation are computed within each window; for the three discrete channels—lane-change flag, left indicator, right indicator—the majority value is taken. This yields $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$00 features for $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$01, or $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$02 features if no windowing is used. For ANN and both SVM variants, the continuous signals or window-features are standardized using training-set mean and standard deviation, whereas RF sees the unscaled window statistics (Alonso-Fernandez et al., 1 Jul 2025).

The score-level fusion strategy is again an unweighted average of RF and linear SVM. If $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$03 and $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$04 are the per-sample scores, then

$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$05

and the fused file-level score is

$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$06

The authors also tried 3-way and 4-way averages, but RF+SVM-lin produced the most consistent per-truck results (Alonso-Fernandez et al., 1 Jul 2025).

After training on balanced data from five trucks and testing on held-out files, RF+SVM-lin yielded the best overall trade-off, with $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$07, $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$08, and $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$09 (Alonso-Fernandez et al., 1 Jul 2025). The per-truck breakdown reported for the fused method is:

Truck	TNR (%)	TPR (%)
t1	70.3	93.7
t2	99.2	80.0
t3	100.0	28.6
t4	95.5	81.3
t5	100.0	62.5

The paper notes that some trucks, specifically t3 and t5, have small numbers of training files, hence lower TPR for those drivers (Alonso-Fernandez et al., 1 Jul 2025). It also reports that variability in traffic conditions strongly influences the signal patterns, particularly in the no-overtake class, and that training with data from multiple vehicles improves generalization and reduces condition-specific bias (Alonso-Fernandez et al., 1 Jul 2025).

5. Overtake-Score as a route-level command-compliance metric

Bench2Drive-Speed defines Overtake-Score in a substantially different way. Let $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$10 be the set of all evaluated routes that include an explicit overtake or follow command; let $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$11 be the user-specified command for route $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$12; let $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$13 indicate whether the scenario actually triggered; and let $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$14 indicate whether the ego-vehicle successfully passed the lead vehicle during the scenario window (Shao et al., 26 Mar 2026).

If the scenario triggered ($S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$15), then for an Overtake command, $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$16 if the ego ever passes the lead ($S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$17), else $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$18; for a Follow command, $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$19 if the ego never passes the lead ($S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$20) for the entire scenario, else $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$21. If the scenario never triggered ($S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$22), the benchmark sets $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$23 to prevent score inflation. The Overtake-Score is then

$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$24

The pass event is defined by

$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$25

The route-level success is written as

$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$26

This construction makes the metric a direct measure of instruction compliance rather than maneuver probability (Shao et al., 26 Mar 2026).

The computational pipeline uses a CARLA ScenarioRunner extension. XML configuration specifies route waypoints, trigger region for overtaking, spawn distance $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$27, front-vehicle speed $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$28, and behavior type. During closed-loop execution, sensor inputs and user commands are collected, the policy outputs control or a planned trajectory, and ego and lead positions are logged along the route centerline. Post-processing then checks whether the ego enters the trigger region, determines whether a pass event occurred between $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$29 and $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$30, sets $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$31, and averages over all routes (Shao et al., 26 Mar 2026).

The benchmark also specifies representative hyperparameters: $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$32 is often zero or a small number such as $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$33 m; the trigger region geometry is chosen to be $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$34 m laterally and $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$35 m longitudinally around the lead-spawn waypoint; timeout is set to the remaining route length or a fixed time such as $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$36 s; initial spawn distance is $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$37–$S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$38 m; and the lead-vehicle speed is sampled as $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$39, with $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$40 m/s and $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$41 m/s (Shao et al., 26 Mar 2026).

Experimental values from the benchmark make clear that commanded overtaking remains difficult. The reported Overtake-Score values are $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$42 for TCP w/o Speed Cmd, $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$43 for TCP-Speed (Expert2.1k), and $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$44 for TCP-Speed (Virtual2.1k) on the “All (48)” split; on “Medium (16),” the corresponding values are $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$45, $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$46, and $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$47; on “Hard (16),” they are $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$48, $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$49, and $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$50 (Shao et al., 26 Mar 2026).

6. Interpretation, limitations, and common points of confusion

A recurring source of confusion is that “Overtake Score” does not refer to a single invariant quantity. In the truck CAN papers, it is a posterior probability estimated by a classifier and then optionally fused and thresholded (Butt et al., 2024, Alonso-Fernandez et al., 1 Jul 2025). In Bench2Drive-Speed, it is an average of binary successes over routes and is already expressed as a percentage (Shao et al., 26 Mar 2026). These are mathematically distinct objects.

Another common misconception is to treat the score as a direct proxy for safe overtaking. The truck studies evaluate overtakes versus no-overtakes using $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$51, $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$52, $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$53, ROC, and PR behavior, and explicitly report trade-offs between detecting overtakes and suppressing false alarms (Butt et al., 2024, Alonso-Fernandez et al., 1 Jul 2025). Bench2Drive-Speed states more directly that its Overtake-Score is a coarse $S_C(t)\equiv P_C(\text{class}=1\mid X_t)\in[0,1]$54 measure: it does not capture how smoothly or safely the overtake was executed, does not penalize a successful overtake that results in a collision, and should be reported alongside standard safety metrics such as collision rate (Shao et al., 26 Mar 2026).

The literature also identifies important failure modes. In the 2024 study, early overtaking prediction is challenging because the overtake-class scores rise as the trigger approaches; performance improves near the trigger, which implies that informative discriminative structure is concentrated in the last few seconds before the event (Butt et al., 2024). In the 2025 study, no-overtake variability under unconstrained real-world conditions affects classification performance if the training data lack adequate diversity, and per-truck accuracy depends on the amount of training data per vehicle (Alonso-Fernandez et al., 1 Jul 2025). In Bench2Drive-Speed, activation failures are counted as failures to prevent score inflation, and a policy could “game” the metric by never attempting to overtake, thereby scoring follow commands perfectly but failing all overtake commands; the benchmark therefore suggests balanced aggregates or separate overtake and follow scores (Shao et al., 26 Mar 2026).

Taken together, these studies situate Overtake Score as a useful but task-dependent abstraction. In signal-based maneuver detection, it is a calibrated probabilistic indicator whose temporal evolution can support early-warning ADAS logic. In instruction-conditioned autonomous driving, it is a compliance metric for interactive behavior. A plausible implication is that future work may need more explicit nomenclature—distinguishing posterior overtake probability, aggregated file-level overtake score, and route-level overtake compliance—to avoid ambiguity across subfields.