Adaptive Cruise Control (ACC) Overview

• Adaptive Cruise Control (ACC) is a longitudinal driver-assistance system that automatically adjusts speed and headway using sensor data.
• ACC integrates radar, lidar, and camera systems to switch seamlessly between cruise and following modes, enhancing safety and comfort.
• Advanced ACC strategies employ optimization, learning, and communication to improve traffic flow, reduce collision risk, and maintain string stability.

Adaptive Cruise Control (ACC) is a longitudinal driver-assistance or in-vehicle control system that automatically regulates vehicle speed and maintains a desired headway distance to a preceding vehicle by adjusting acceleration and deceleration using onboard sensing such as radar, lasers, or cameras. In its standard operating logic, ACC behaves like cruise control when no lead vehicle is detected and switches to following control when a lead vehicle is present. In contemporary research, ACC is treated not only as a comfort function but also as a safety-critical and traffic-aware control problem in which headway selection, perception, communication, prediction, and control synthesis jointly affect safety, comfort, throughput, energy use, and resilience (Kural et al., 2020, Das et al., 2020).

1. Operational scope and system objectives

ACC is commonly described as a longitudinal control function that maintains a safe following distance while also tracking a driver-selected cruising speed when the road ahead is clear. The literature distinguishes classic highway ACC from Stop-and-Go systems: classic ACC is intended primarily for highway driving and operates at about 40–160 km/h, whereas Stop-and-Go extends the same principle to low-speed congested traffic, operates roughly from 0–40 km/h, and includes full stop and restart behavior (Kural et al., 2020).

The principal design objectives recur across the literature. Safety is defined through safe following distance and rear-end collision avoidance; comfort is associated with smooth acceleration and deceleration; driver workload reduction follows from automating routine longitudinal control; and string stability concerns whether disturbances amplify through a platoon. Efficient traffic flow is an additional objective, especially in mixed traffic and platooning settings. Cooperative Adaptive Cruise Control (CACC) extends ACC by using inter-vehicular wireless communication to share states such as acceleration, enabling shorter gap following and faster reactions than sensor-only ACC (Kural et al., 2020, Lin et al., 2019).

A central correction to a common simplification is that ACC is not merely a fixed-gap convenience feature. On homogeneous highway segments, smaller gaps can increase throughput, but on highways with ramps the best headway depends on the joint state of the mainline, the ramp, and the local geometry; making the headway too small can impede merges, induce braking, and worsen delay. This directly reframes ACC as a traffic-flow control lever rather than a static comfort setting (Das et al., 2020).

2. Headway policies and longitudinal dynamics

The dominant upper-level formulation is constant time-gap following. If $x$ is the follower position and $x_1$ is the predecessor position, the spacing is $\Delta x = x_1 - x$, and the constant time-gap policy is

$$\Delta x(t)-\ell \approx T_s \dot x(t), \qquad T(t)=\frac{\Delta x(t)-\ell}{\dot x(t)}.$$

In CACC notation, the desired spacing may be written as $h \dot l_i + l_0$, with spacing error

$$e_i = l_{i-1} - l_i - b_{i-1} - (h \dot{l}_i + l_0).$$

This time-gap policy is preferred to constant spacing because the physical spacing increases with speed and supports string-stable operation (Tordeux et al., 2019, Lin et al., 2019).

A standard car-following abstraction writes the follower acceleration as

$\ddot x_n = F(\Delta x_n,\dot x_n,\dot x_{n+1}).$

Within this class, the literature compares the Optimal Velocity (OV), Full Velocity Difference (FVD), constant time-gap (CTG), truncated Intelligent Driver (ID), and Adaptive Time Gap (ATG) models. The robustness analysis for full speed range ACC shows that CTG and ATG are systematically stable for positive parameters in the idealized model, whereas OV, FVD, and ID require parameter restrictions. The same analysis also shows that intrinsic stability is not sufficient: latency, stochastic noise, heterogeneity, and actuator saturation can all induce stop-and-go waves even when the nominal car-following law is stable (Tordeux et al., 2019).

Recent work replaces the fixed time gap with a controlled variable. In variable time-gap ACC, the commanded headway is

$\tau_i = \tau_i^\star + u_i(t),$

where the constant component $\tau_i^\star$ dominates during equilibrium car-following and the feedback component $u_i(t)$ is activated to damp disturbances. The proposed nonlinear $x_1$0 design keeps a minimum or desired constant time gap during steady traffic but temporarily relaxes it when perturbations appear, thereby separating equilibrium efficiency from disturbance attenuation (El-Baklish et al., 2024).

3. Sensing, estimation, and cooperative information

ACC is typically organized as a hierarchical system. A two-level architecture is standard: the upper-level controller uses sensor inputs to compute desired acceleration or deceleration, and the lower-level controller converts that command into throttle and brake actuation. Forward-looking radar and lidar dominate the classical sensor stack, with radar described as the most mature practical sensor technology for ACC and Stop-and-Go; camera-based perception has become increasingly prominent in more recent DNN-based systems (Kural et al., 2020).

Purely range-sensor-based following is vulnerable to target detection loss. On curvy roads, steep hills, in adverse weather, or under target-selection failure, the range sensor may lose the preceding vehicle, and standard ACC typically falls back to constant-speed cruise control. An alternative is to approximate inter-vehicular distance using inter-vehicular communication, accurate vehicle localization, and a digital map. In that formulation, vehicle positions are projected onto a lane-center curve $x_1$1, and the distance is approximated by the arc length

$x_1$2

between the projection points, allowing vehicle following to continue during target detection loss rather than reverting to cruise mode (Lin et al., 2019).

Preview information generalizes the sensing problem beyond the immediate lead vehicle. Safe ACC with road grade preview uses a model predictive control formulation and a robust control invariant terminal set to guarantee safe inter-vehicle spacing despite changes in road slope and uncertainty in predicted lead-vehicle motion. Ecological ACC extends the preview horizon further by incorporating Signal Phase and Timing data from traffic lights and probabilistic delay models into an optimization-based reference velocity generator, with a lower-level robust MPC enforcing front-collision avoidance and traffic-light compliance (Firoozi et al., 2018, Bae et al., 2018).

4. Optimization, learning, and correct-by-construction ACC

Model-based advanced ACC has moved from deterministic tracking to uncertainty-aware safety constraints. A learning-based risk-averse MPC formulation models the preceding vehicle as a Markov jump linear system, estimates transition probabilities empirically, surrounds those estimates with $x_1$3-ball ambiguity sets, and solves a distributionally robust AVaR-constrained optimal control problem. A robust terminal invariant set and a nested ambiguity-set condition establish recursive feasibility, yielding an ACC law that is less conservative than fully robust control but safer than nominal stochastic MPC under distribution mismatch (Schuurmans et al., 2020).

State estimation has become an explicit control component when V2V communication is absent. In an adaptive estimation-based safety-critical ACC design, the safety function is

$x_1$4

and Lyapunov functions are combined with Control Barrier Functions (CBFs) to guarantee safety despite estimation errors in predecessor velocity and acceleration. The framework is implemented both with and without V2V communication and reports a four-vehicle string stability gain of

$x_1$5

indicating disturbance attenuation through the platoon (Bohara et al., 2023).

Correct-by-construction ACC uses CBFs to encode hard constraints directly in the control synthesis. Under safety and regulatory constraints such as traffic lights, time-varying CBFs must accommodate finite jump discontinuities in the signal sequence. A piecewise $x_1$6 TV-CBF construction enables forward invariance of safe sets across switching times, allowing the controller to maintain safe headway, respect speed limits, and stop before red lights while minimally modifying a nominal PID-style tracking controller through a quadratic program (Waqas et al., 2022).

Learning-based ACC has expanded along several axes. For highway ramps, D-ACC formulates headway adaptation as a deep Q-learning problem over a state consisting of main-road density, ramp density, main-road average speed, ramp average speed, and ramp length; with SUMO, Veins, and V2X information, it improves traffic flow by up to 70% compared with a state-of-the-art intelligent ACC system in a highway segment with a ramp (Das et al., 2020). SAINT-ACC uses a dual deep RL architecture in which one agent adapts a dynamic $x_1$7 threshold and a second agent selects the inter-vehicle gap, reporting simultaneous gains in safety, efficiency, and comfort across on-ramp and off-ramp scenarios (Das et al., 2021). Under aggressive cut-in maneuvers, an LSTM-based predictor trained on the highD dataset estimates future subject-vehicle acceleration and is 19.25% more accurate than the ANN model and 5.9% more accurate than the MPC model (Singh et al., 2023). At the opposite extreme of architectural abstraction, an end-to-end vision-based ACC trained with Double Deep Q-Networks maps image and speed sequences directly to 21 discretized throttle/brake actions and can generate either a better gap regulated trajectory or a smoother speed trajectory depending on the preset reward function (Wei et al., 2020).

5. Safety, uncertainty, and adversarial robustness

A major contemporary theme is that ACC safety depends not only on control law design but also on the epistemic status of perception outputs. In camera-only ACC, a Deep Ensemble of six heterogeneous CNN regressors—ResNet50, GoogleNet, AlexNet, MobileNetV2, EfficientNet, and VGG16—produces a Gaussian headway estimate $x_1$8, and the downstream controller enforces the chance constraint

$x_1$9

Under out-of-distribution conditions, including a lead-vehicle change from a Lincoln MKZ to a firetruck and a weather change from ClearNoon to HardRainSunset, the uncertainty grows substantially and the MPC decelerates below the nominal set speed, prioritizing safety over speed tracking (Li et al., 2024).

Cyber-physical resilience has become equally central. PID-based ACC that depends on ego-vehicle speed can be destabilized by CAN-bus spoofing: if the speed communicated through CAN is falsified, the controller may accelerate when it should brake. A CARLA-based study shows that spoofing the ego speed to 10 km/h while actual driving speed is 60 km/h or 90 km/h can lead to collisions, and that extending ACC with an ML-based real-time IDS that triggers a “cold brake” can mitigate the attack. The reported IDS performance includes a detection rate of 0.97 and a response time of 1026 ms, with additional scenario analyses reporting 149–152 ms (Jedh et al., 2023).

The same vulnerability persists even when a Kalman filter is used to smooth noisy speed measurements. Analytical results show that the filter can tolerate injected speed values only up to a bounded threshold; beyond that threshold, the filtered speed estimate can exceed a safety-critical value and the next-step gap becomes unsafe. This motivates an ACC-IDS controller in which a binary intrusion flag $\Delta x = x_1 - x$0 switches the control law to emergency braking $\Delta x = x_1 - x$1. Under the paper’s detection-performance and latency assumptions, ACC-IDS preserves collision avoidance by ensuring that the remaining distance is at least the braking distance once intrusion is detected (Othmane et al., 1 Mar 2026).

Perception attacks extend the threat model to DNN-based ACC. Runtime stealthy perturbations applied to live camera frames can alter estimated relative distance and relative speed in the lead-vehicle detector and thereby induce hazardous longitudinal plans. In an enhanced OpenPilot + CARLA platform with AEBS and a driver reaction simulator, the context-aware optimization-based attack CA-Opt achieves 142.9 times higher success rate in causing hazards and 82.6% higher evasion rate than baselines while remaining visually subtle. The same study also shows that alert human drivers and independent AEBS reduce, but do not eliminate, attack success (Zhou et al., 2023).

6. Traffic-flow effects, specialized applications, and adoption

ACC design choices feed back into network-level traffic behavior. On highway segments with ramps, the headway setting is a traffic-flow control variable, and D-ACC reports up to 70% higher average speed than the Manolis baseline at 60% penetration (Das et al., 2020). In mixed traffic, variable time-gap ACC improves traffic outflow by about 1.05% on average versus calibrated CTG ACC at 10% ACC penetration in a highway merging scenario, while also improving minimum TTC and tractive energy consumption relative to constant-gap control (El-Baklish et al., 2024). In signalized arterial driving, ECO-ACC reports a 41.0% reduction in equivalent fuel consumption and 32.91% less wheel energy than an ACC-Only baseline, at the cost of longer travel time (Bae et al., 2018).

Specialized ACC variants further broaden the operating domain. Anti-bullying ACC formulates mandatory cut-in interaction as a POMDP with online inverse optimal control and Stackelberg game-based MPC, allowing the ego vehicle either to protect right-of-way proactively or to yield strategically depending on the competing vehicle’s inferred driving style. It reports travel-efficiency improvement up to 29.55% under different cut-in gaps, mobility improvement up to 11.93%, robustness improvement by 8.74%, and computation time below 50 milliseconds (Hu et al., 2024). For teleoperated driving, steering action-aware ACC defines a safe state as one from which the vehicle can be stopped safely, no matter which steering actions are applied by the operator; by sampling possible future steering trajectories and then optimizing a safe velocity profile, it keeps a 1:10-scale teleoperated vehicle safe even when the operator’s commands would have caused a collision (Schimpe et al., 2022).

Public acceptance and willingness to pay remain distinct from technical feasibility. A survey of 453 respondents in Iran, analyzed with the Technology Acceptance Model and an ordered logit model, shows that perceived ease of use and perceived usefulness affect attitude toward using ACC, and that attitude influences behavioral intention. The same study finds that drivers who find ACC easy and useful, who have higher vehicle prices, and who are women with cruise-control experience are more likely to pay for ACC. The reported pattern is that acceptance is promising, but willingness to pay is weaker, with affordability constraints shaping near-term diffusion (Sahebi et al., 2024).