The Pedal Override Is Feedback, Not a Command

TL;DR for operators

An accelerator override during adaptive cruise control (ACC) operation can contain more information than a simple indication that the driver wanted to accelerate. It records how the driver’s preferred acceleration differed from the automated response under particular driving conditions.

Xu and colleagues propose Context-driven Personalized Adaptive Cruise Control (CoP-ACC), which learns recurring acceleration patterns from manual overtaking and shared-control overrides, predicts an appropriate pattern from the approaching maneuver’s context, and refines it into a continuous trajectory.¹

Using 140 overtaking events from one driver, the system achieved 89.29% accuracy in classifying acceleration-profile patterns and 0.0306 m/s² mean absolute error when reconstructing withheld override profiles. Generated profiles also approached observed human velocity-gain behavior while showing lower maximum jerk in offline comparisons.

For ADAS developers, the opportunity is to turn interventions into structured personalization data. The boundary is equally specific: predicting what a driver prefers does not establish that the resulting trajectory is safe to execute. CoP-ACC demonstrates a preference-prediction component, not a validated personalized ACC controller.

An accelerator override reveals a recurring control mismatch

Consider a vehicle approaching a slower car on a highway. Adaptive cruise control maintains its ordinary longitudinal-control behavior, but the driver expects a more decisive increase in speed during overtaking. The driver presses the accelerator.

The immediate intervention resolves the driver’s dissatisfaction with the current response. For the automation system, however, the event could be recorded merely as an exception. That treatment discards information about the conditions preceding the intervention and the acceleration the driver subsequently demanded.

The study examines whether these interventions exhibit enough regularity to support personalized prediction.

Its dataset comes from an instrumented Renault Austral driven by one experienced commuter on a three-lane French highway during low-traffic periods. The experiment distinguishes three driving modes: generic ACC without pedal overrides, naturalistic manual driving, and shared-control driving in which accelerator overrides are permitted.

Of 223 initially identified overtaking events, 140 were retained for personalized-profile training: 88 naturalistic manual maneuvers and 52 shared-control overrides.

The authors identify three recurring acceleration patterns: aggressive, moderate, and nearly passive. Their distribution provides an informative first observation.

Source: Table I, Section V-B. These counts describe training events, not distinct drivers.

Among the 76 maneuvers assigned to the aggressive and moderate patterns, 49 were shared-control overrides, representing 64.4%. Contextual distributions also associate overrides with larger speed deficits, lower vehicle speeds, and shorter distances to the target vehicle.

These descriptive findings suggest that the driver’s dissatisfaction with generic ACC was related to identifiable conditions rather than occurring independently of the maneuver context.

They do not establish that stronger acceleration was objectively preferable. An intervention supplies evidence about desired behavior, not an independent assessment of driving safety.

Predict a behavioral pattern before refining its acceleration

The central methodological choice in CoP-ACC is to avoid treating every overtaking maneuver as an unrelated continuous acceleration-regression problem.

A direct predictor can produce an average trajectory when training examples contain substantially different acceleration patterns. Such averaging risks weakening maneuver-specific peaks. Conversely, selecting only a representative pattern cannot capture the variation between maneuvers assigned to the same group.

CoP-ACC separates these tasks into three stages.

First, discover recurring acceleration patterns. The authors normalize each observed maneuver’s acceleration trajectory to 100 temporal samples and apply hierarchical time-series clustering using Euclidean distance and Ward linkage. Silhouette-based analysis supports three representative patterns, described as aggressive, moderate, and passive.

Second, predict which pattern fits the upcoming maneuver. A balanced Random Forest classifier receives 11 pre-maneuver context features, including vehicle speed, relative speed, distance, time headway, inverse time-to-collision, and short-window steering, throttle, and jerk statistics. These features summarize conditions during the five seconds preceding the annotated maneuver.

Third, refine the selected pattern. A context-conditioned one-dimensional convolutional neural network predicts a continuous residual trajectory. Adding that residual to the selected cluster centroid produces the personalized acceleration profile:

Here, $a^{\ast}_{\hat{y}_i}$ is the representative acceleration trajectory selected by the classifier, and $\hat{r}_i$ is the learned context-specific correction.

The resulting architecture preserves a recognizable behavioral pattern while allowing continuous adjustments. It does not eliminate regression; it assigns regression the narrower task of correcting an already selected trajectory.

This modularity also makes model behavior easier to inspect. Developers can examine whether the classifier chose an inappropriate behavioral pattern or whether the residual decoder produced an inadequate correction.

The experiments separate pattern recognition from trajectory fidelity

CoP-ACC’s results answer several different predictive questions. Their measurements should not be combined into a single measure of controller performance.

Context predicts the acceleration-pattern class

Using an 80/20 stratified train-test split and five-fold stratified cross-validation in the training procedure, the Random Forest achieved 89.29% test accuracy, with macro F1 of 85.65%.

This supports the feasibility of classifying acceleration patterns from pre-maneuver conditions within the studied driver’s data. It is not an accuracy measure for detecting future overrides or safely completing overtakes.

SHAP analysis, reported in Figure 5, identifies ego speed and speed deficit as the most influential classifier features. The extracted short-window behavioral-style features contribute relatively little for this driver.

That result suggests immediate kinematic conditions carry substantial predictive information in this particular setting. SHAP describes the model’s feature attributions, however, not causal effects of speed or distance on driver preferences.

Residual learning improves on fixed templates

The authors compare the residual decoder against a centroid-only reconstruction baseline. This is an ablation of the continuous refinement component, testing whether the additional decoder improves representation beyond selecting a discrete template.

Reported mean absolute error declines from 0.02522 to 0.02398 m/s², while mean squared error declines from 0.001308 to 0.001088.

The improvement is measurable but modest. It supports adding context-dependent corrections rather than relying exclusively on representative trajectories.

In the separate shared-control override test subset, the integrated classifier-and-decoder pipeline achieves an MAE of 0.0306 m/s², average Dynamic Time Warping distance of 2.00, and Pearson correlation of 0.804.

These figures establish trajectory-reconstruction fidelity on withheld override examples. The decoder-versus-centroid comparison and the integrated override-subset evaluation concern different evaluation populations and should not be interpreted as interchangeable error measurements.

Generated trajectories resemble human demand more closely

The authors then apply the model to withheld generic-ACC contexts resembling situations in which overrides occurred. This serves as an offline behavioral-alignment evaluation, rather than another reconstruction ablation or a deployed driving trial.

The generated profiles move toward observed override behavior in total velocity gain, measured through the area under the acceleration curve. They also exhibit lower peak acceleration than human overrides and lower maximum jerk than both human overrides and generic ACC in the reported distributional comparison (Figures 6–7).

One plausible explanation is that anticipating desired acceleration could avoid some of the abrupt late corrections associated with manual intervention.

But the evidence concerns predicted acceleration trajectories. It does not show that drivers experienced greater comfort, intervened less often, or responded differently when the generated trajectories were executed.

Turn intervention data into personalization without delegating safety

For automotive ADAS developers, CoP-ACC suggests a practical way to use information already exposed by human-machine interactions.

A research pipeline could preserve accelerator overrides alongside the preceding vehicle-state signals, identify recurring behavioral patterns, and train context-conditioned models to predict preferred acceleration. The same pipeline could evaluate model outputs against observed driver behavior using trajectory error, velocity gain, peak acceleration, and jerk.

This is a more specific personalization objective than adjusting a fixed preferred following gap. It concerns the acceleration trajectory of an overtaking maneuver and its dependence on the immediate driving situation.

Cognaptus inference: the architecture is particularly relevant where developers want to improve perceived responsiveness while retaining existing controller safety functions. Preference estimation can be developed and evaluated separately from the mechanism that authorizes longitudinal acceleration.

Such separation also provides a clearer governance boundary. A learned trajectory may indicate what a driver is likely to want, but a controller must still determine whether that request satisfies acceleration limits, spacing requirements, road conditions, and other safety constraints.

The paper does not implement that authorization layer or demonstrate integration of its predictions into executable ACC controller parameters.

A single driver’s predicted preferences are not a validated control policy

The most consequential evidence boundary is the scale and design of the experiment.

All 140 training events come from one driver, on predefined highway routes during low-traffic periods. Residual learning uses Gaussian-noise and temporal-warping augmentation to help stabilize training on this small dataset. Neither augmentation nor stratified testing establishes transfer to new drivers.

The feature set also excludes road geometry, including curvature. That omission matters because acceleration appropriate on one road segment may be inappropriate on another, even with similar vehicle speeds and target distances.

Overtaking phases were manually annotated, leaving automatic event detection unresolved. More fundamentally, the predicted acceleration profiles were never executed as closed-loop vehicle commands in the reported evaluation.

The comparison between forced generic ACC and generated personalized trajectories therefore does not identify the causal effect of deploying CoP-ACC on override frequency, actual comfort, or safety.

The authors identify multi-driver validation, automatic detection, adaptive controller-parameter integration, and closed-loop simulation as future work. Those investigations would test capabilities that the current offline results cannot establish.

The useful signal is preference; execution still requires authorization

CoP-ACC provides a concrete demonstration that human interventions can contain learnable, context-dependent information about desired automated behavior.

Its contribution is not simply a higher-fidelity acceleration predictor. By separating behavioral-pattern selection from continuous refinement, it gives developers a structured way to represent individual acceleration preferences and test whether those representations recover observed human actions.

For ADAS product development, that supports research into more responsive personalization. The next evidence requirement is to establish whether those predictions remain appropriate across drivers and operating conditions—and whether applying them through a safety-constrained controller improves actual driving outcomes.

The accelerator override can inform what a driver wanted. It cannot, by itself, authorize what the vehicle should do.

Cognaptus: Automate the Present, Incubate the Future.