Predictor-Based Cooperative Adaptive Cruise Control

• Predictor-based CACC is a connected control strategy that augments traditional feedforward loops with explicit forecasts of vehicle, sensor, and communication states.
• It integrates various predictive models such as MPC, LSTM, and GP to anticipate maneuvers, compensate delays, and mitigate sensor and communication losses.
• This proactive approach improves gap regulation, reduces energy consumption, and enhances overall safety compared to conventional ACC and baseline CACC.

Predictor-based cooperative adaptive cruise control (CACC) denotes a class of connected longitudinal control architectures in which the conventional CACC loop is augmented with an explicit predictor of future vehicle, driver, sensing, or communication state. Relative to adaptive cruise control (ACC), which relies on onboard sensing of a single leader, CACC exposes position, velocity, and often acceleration of upstream, downstream, or adjacent vehicles through V2V links; predictor-based variants then use this additional information to anticipate cut-ins and lane changes, compensate actuation and communication delays, bridge packet losses and target-detection loss, incorporate human takeover and human-lead uncertainty, and harden the control loop against adversarial V2V data (Qasemabadi et al., 2023, Kim et al., 2021, Mosharafian et al., 2021, Wang et al., 2024, Samii et al., 29 Jul 2025, Samii et al., 7 Apr 2026).

1. Core concept and relation to baseline CACC

In its canonical form, CACC extends ACC by adding cooperative feedforward to the usual car-following feedback. A standard spacing policy is the constant-time-headway relation

$d_i(t)=d_0+h\,v_i(t),$

with standstill distance $d_0$ and headway $h$, and a representative control law is

$u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$

where $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$. The term $k_a a_{i-1}(t)$ is the characteristic CACC feedforward channel; it uses V2V acceleration to reduce delay-induced oscillations compared with ACC. In the broader literature, this baseline appears in predecessor-following, predecessor–leader following, and multiple-predecessor-following topologies, depending on how much upstream information is exploited (Qasemabadi et al., 2023, Samii et al., 29 Jul 2025, Hu et al., 2024, Samii et al., 7 Apr 2026).

Predictor-based CACC changes the role of the cooperative channel from pure feedforward to anticipative control. Instead of reacting only to measured predecessor motion, the controller acts on a forecast: a preview of leader acceleration, a delay-compensated future state, a probability of lane change or cut-in, a model of a human driver’s best response, or a reconstructed predecessor trajectory during communication loss. This shifts CACC from reactive spacing regulation toward proactive maneuver coordination, safety buffering, and disturbance attenuation across the platoon (Kim et al., 2021, Qasemabadi et al., 2023, Mosharafian et al., 2021).

ACC’s sensing limitations provide much of the motivation for this transition. One lane-change study explicitly frames ACC as relying on onboard radar/lidar/camera measurements of the leading vehicle, typically with limited range of about $150\,\mathrm{m}$ and field of view of about $20$ degrees, whereas CACC can access position, velocity, and acceleration of multiple surrounding vehicles, including adjacent-lane leaders and followers. A plausible implication is that predictor-based CACC is most valuable precisely when the relevant future constraint is not visible in the ego lane alone (Qasemabadi et al., 2023).

2. Predictive targets and model families

The predictive component in predictor-based CACC is not singular. Different works forecast different latent quantities and inject them into the control layer in different forms.

Predictive target	Model family	Control role
Predecessor acceleration preview	robust MPC with trust horizon $N_T$	energy-aware gap reduction
Ego lane-change intention	two-layer LSTM	pre-maneuver headway and gap creation
Predecessor speed during comm gaps	local GP with model-based communication	stochastic MPC under intermittent V2V
Human takeover response	Stackelberg/game-based MPC	authority allocation and smooth takeover
Adjacent-vehicle cut-in trajectory	NAR, NARX, and RNN	risk-adaptive SMPC
Human leader acceleration	OVM + Langevin/CIR scenario tree	safety-aware HL-CACC
Delay-compensated future platoon state	predictor-feedback / Artstein-type state prediction	actuation-delay compensation
V2V plausibility under attack	FNN normal-behavior model	anomaly detection and mitigation
Target distance during sensor loss	EKF + digital map projection	continued gap control

At the data-driven end of the spectrum, one lane-change framework uses a multi-LSTM with two stacked LSTM layers, a fully connected layer with $32$ neurons, dropout $d_0$0, ReLU intermediate activations, sigmoid output, RMSprop, and binary cross-entropy loss; extensive ablation identifies $d_0$1 cells per layer and an input window of $d_0$2 frames as the best setting on HighD. A cut-in framework uses NAR, NARX, and RNN modules with one hidden layer of $d_0$3 nodes, $d_0$4-step memory, and a $d_0$5-step prediction horizon, producing $d_0$6 confidence rectangles whose overlap with a host “bad-set” defines a cut-in probability. A human-lead formulation models the leader’s acceleration with an optimal-velocity mean term plus a speed-dependent diffusion term from an extended CIR-like process, discretized into a scenario tree. A takeover controller models the human as a rational follower with best response

$d_0$7

which is embedded directly in the machine MPC rather than appended as a separate prediction stage (Qasemabadi et al., 2023, Kazemi et al., 2018, Hu et al., 2024, Wang et al., 2024).

At the model-based end, a robust energy-saving design receives a V2V message

$d_0$8

so that a trusted short preview of leader acceleration is used for $d_0$9 steps and then replaced by worst-case braking beyond the trust horizon. A Gaussian-process approach trains a local GP for each vehicle’s speed using the last $h$0 equally spaced samples over the last $h$1, broadcasts hyperparameters $h$2 under a model-based communication paradigm, and reconstructs future speed and acceleration locally during sparse or failed communication. A delay-compensation line of work predicts the extended platoon state at the instant a delayed input will actually take effect, using closed-form matrix exponentials and input-history integrals, both in sampled-data predecessor-following and in multiple-predecessor-following formulations (Kim et al., 2021, Mosharafian et al., 2021, Samii et al., 29 Jul 2025, Samii et al., 7 Apr 2026).

3. Control synthesis patterns and stability criteria

A defining feature of predictor-based CACC is that prediction is not merely informative; it is structurally embedded into the control law. One common pattern is preview-and-worst-case scheduling inside robust MPC. In the energy-focused predecessor–follower design, the lead acceleration is modeled as known over a short trust horizon and equal to the worst-case braking bound thereafter, so the follower can safely close the gap when preview quality is high while preserving hard safety constraints through a robust control invariant terminal set. The resulting RMPC solves a finite-horizon problem with $h$3, $h$4, direct penalization of distance to $h$5, and jerk moderation through input-difference penalties (Kim et al., 2021).

A second pattern is supervisory adaptation of the spacing policy. In the lane-change setting, the LSTM outputs $h$6 at every control step; if the probability exceeds a threshold $h$7 in the range $h$8–$h$9, the cooperative controller enters a pre-maneuver mode, raises the headway from $u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$0 to $u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$1, requests gap creation from target-lane neighbors, modulates longitudinal acceleration to align with the available gap, and reverts to conservative behavior if V2V latency or loss degrades communication quality. The cut-in SMPC uses a different mechanism: a cut-in probability $u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$2 rescales the spacing-error definition,

$u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$3

so that increasing cut-in risk proactively enlarges the effective spacing target; if the QP becomes infeasible, the controller switches to emergency deceleration up to $u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$4 (Qasemabadi et al., 2023, Kazemi et al., 2018).

A third pattern is endogenous prediction, in which the planner optimizes against a predicted response that depends on its own action. The takeover GMPC formulates the transition from CACC to human control as a Stackelberg game with fused acceleration

$u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$5

Because the human reaction model enters the machine dynamics directly, the machine solves a bilevel control problem in which its own command shapes the predicted human input. The human-lead stochastic MPC follows an analogous philosophy at the disturbance level: it propagates a scenario tree for the human leader’s future acceleration and adds a CVaR-style excess-over-threshold penalty to emphasize low-probability but severe headway-error realizations (Wang et al., 2024, Hu et al., 2024).

A fourth pattern is exact delay compensation. In a sampled-data predecessor-following design with heterogeneous third-order longitudinal dynamics and distinct actuation delays, the controller computes a predicted extended state $u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$6 from current measurements, local input history, and received predecessor inputs, and then applies the nominal delay-free law to $u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$7. In the more general multiple-predecessor-following formulation, the predictor-feedback law

$u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$8

exactly compensates any common actuation delay $u_i(t)=k_p e_i(t)+k_d \dot e_i(t)+k_a a_{i-1}(t),$9 at the nominal level while incorporating heterogeneous communication delays in the string-stability analysis. Across these formulations, string stability is expressed either as $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$0 in sampled-data form, as $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$1 in MPF form, or as a time-domain oscillation criterion on successive gap ranges (Samii et al., 29 Jul 2025, Samii et al., 7 Apr 2026, Wang et al., 2024).

4. Maneuver anticipation and human-centered variants

Lane-change anticipation is one of the clearest demonstrations of the value of prediction-rich CACC. Using HighD, which contains trajectories for more than $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$2 vehicles across six locations and includes $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$3 complete lane changes, the LSTM-based intention predictor compares ACC-like sensing with full CACC information. The implemented features are relative Manhattan differences in position, velocity, and acceleration between the ego vehicle and up to eight surrounding vehicles. With three preceding vehicles and position-plus-velocity signals, the ACC-like scenario reaches $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$4 accuracy, $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$5 precision, and $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$6 recall. With all eight surrounding vehicles and position, velocity, and acceleration, the full CACC scenario reaches $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$7 accuracy, $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$8 precision, and $e_i(t)=[p_{i-1}(t)-p_i(t)-d_0]-h v_i(t)$9 recall. The same study reports that rear neighbors alone already yield $k_a a_{i-1}(t)$0 accuracy, whereas adding alongside vehicles to preceding vehicles gives only $k_a a_{i-1}(t)$1, indicating that following vehicles are disproportionately informative for ego lane-change intention. The ablation on signal richness further reports $k_a a_{i-1}(t)$2 for $k_a a_{i-1}(t)$3 only, $k_a a_{i-1}(t)$4 for $k_a a_{i-1}(t)$5, and $k_a a_{i-1}(t)$6 for $k_a a_{i-1}(t)$7, with execution times increasing from $k_a a_{i-1}(t)$8 to $k_a a_{i-1}(t)$9 (Qasemabadi et al., 2023).

Cut-in anticipation addresses a closely related but distinct disturbance: an interfering vehicle entering the host lane. In the stochastic MPC design built from Safety Pilot Model Deployment trajectories, DSRC Basic Safety Messages arrive at $150\,\mathrm{m}$0, the suspicious vehicle’s lateral and longitudinal positions are predicted $150\,\mathrm{m}$1 steps ahead, and overlap between the predicted $150\,\mathrm{m}$2 confidence rectangles and a host bad-set produces a scalar cut-in probability. Reported warning lead time is about $150\,\mathrm{m}$3 for an average $150\,\mathrm{m}$4 maneuver and about $150\,\mathrm{m}$5 for a harsh $150\,\mathrm{m}$6–$150\,\mathrm{m}$7 maneuver. At about $150\,\mathrm{m}$8 host speed, the stochastic MPC reduces worst-case spacing error from about $150\,\mathrm{m}$9 to about $20$0 in the average case, and at the moment of lane entry the spacing error is about $20$1 versus about $20$2 for a conventional MPC without cut-in probability (Kazemi et al., 2018).

Human-centered predictor-based CACC extends anticipation from surrounding vehicles to the driver inside the loop. In takeover scenarios, close CACC headways of $20$3, compared with $20$4–$20$5 in ACC or human driving, make direct authority transfer hazardous, especially because drivers may need $20$6–$20$7 to regain full manual capability. The Stackelberg takeover controller reports string stability in the tested operational design domain when human authority remains below $20$8, a $20$9 reduction in acceleration range at $N_T$0 relative to full human control, and up to $N_T$1 reduction in upstream shockwave propagation duration. In human-lead platooning, a scenario-tree stochastic MPC that predicts the human leader’s uncertain acceleration reports $N_T$2 improvement in perceived safety in oscillating traffic, $N_T$3 improvement in actual safety against hard brakes, string stability in simulation, and computation time of about $N_T$4 on an Intel i5-13500H laptop (Wang et al., 2024, Hu et al., 2024).

5. Delay compensation, degraded information, and resilience

A major branch of predictor-based CACC is concerned not with maneuvers but with information degradation. Under intermittent or low-rate communication, the GP-based stochastic MPC uses model-based communication to avoid degrading to ACC. Vehicles send predicted acceleration profiles when packets succeed and also broadcast GP hyperparameters so that followers can reconstruct future speed and acceleration locally during gaps. In a platoon of $N_T$5 vehicles with sampling time $N_T$6 and prediction horizon $N_T$7, the combined DH-DHSMPC approach remains safe and efficient at communication period $N_T$8 with success probability $N_T$9, closely matching the perfect-communication baseline; under the same low-rate/loss condition, the model-free DHMPC baseline suffers a rear-end collision (Mosharafian et al., 2021).

Sensor degradation motivates a different predictive substitution. During radar or lidar target-detection loss on curves or hills, one CACC design replaces range sensing with an along-path distance estimate obtained from the leader’s communicated localized position, the follower’s EKF localization, and a digital lane-center map. The local lane center is approximated by a quadratic fit $32$0, and the inter-vehicular gap is estimated by arc length between the vehicles’ projections onto that curve. In robot experiments, localization runs at $32$1, the emulated GPS at $32$2, and V2V at $32$3. Estimator-only operation over a full lap yields mean time gap $32$4 with standard deviation $32$5 for desired headway $32$6; in a switching mode that uses the estimator only when the IR range reading leaves $32$7 around a desired $32$8 distance, the mean time gap is $32$9 with standard deviation $d_0$00 (Lin et al., 2019).

Cyber-resilience introduces yet another predictive layer. RACCON trains a feed-forward neural network on benign data only to learn the normal behavior model of the onboard CACC response from inputs $d_0$01, computes the residual

$d_0$02

and switches to mitigation if the discrepancy exceeds an environment-dependent threshold or if communication is lost. In mitigation mode, the system estimates predecessor acceleration from trusted sensors, recomputes a corrected CACC action, compares it against a second response-estimator output through worst-case headway projections, and falls back to ACC when necessary. In the Highway–Day–Windy environment, threshold $d_0$03 is selected to detect biases as low as $d_0$04 while keeping benign false positives at about $d_0$05. Across representative attacks, RACCON maintains time headway in the $d_0$06 band and avoids collisions where naive CACC becomes unsafe (Boddupalli et al., 2021).

Actuation delay is the most classical predictor problem in CACC. Experimental validation with two full-scale electric vehicles uses a digital predictor-based law at $d_0$07 with $d_0$08, $d_0$09, engine lag $d_0$10, leader delay $d_0$11, follower delay $d_0$12, and modal V2V latency about $d_0$13. The reported responses show close speed and acceleration tracking with essentially no overshoot, despite the larger ego delay. The same line of work presents consistent five-vehicle simulations under heterogeneous lags and delays. The more recent multiple-predecessor-following predictor-feedback design extends this idea to heterogeneous MPF platoons with any actuation delay $d_0$14, derives explicit string-stability conditions, and shows in ten-vehicle simulations that MPF can remain string stable at $d_0$15 under parameters for which a single-predecessor design loses string stability (Samii et al., 29 Jul 2025, Samii et al., 7 Apr 2026).

6. Performance envelope, limitations, and research directions

Across the literature, predictor-based CACC is associated with three principal benefits: earlier intervention, smaller conservatism for the same hard safety envelope, and better disturbance attenuation. In the energy domain, short trusted acceleration preview allows a follower to converge to a smaller gap and reduce drag; experimentally characterized small-gap operation reports fuel savings of $d_0$16–$d_0$17 in FC mode and battery-energy savings of $d_0$18–$d_0$19 in FE mode, while RMPC simulations report steady-state wheel energy falling from $d_0$20 of the lead vehicle at $d_0$21 to $d_0$22 at $d_0$23 steps, with the largest benefits occurring at low latency and mild communicated braking bounds. In maneuver coordination, earlier intent detection enables proactive gap creation, mitigates cut-ins, and reduces abrupt braking. In human-centered variants, predictive handling of human behavior reduces perceived risk, actual collision exposure, and upstream shockwave duration (Kim et al., 2021, Qasemabadi et al., 2023, Wang et al., 2024, Hu et al., 2024).

The limitations are equally clear. Many results are domain-specific: HighD is highway-only, robotic target-loss experiments are planar and low-speed, and several human-centered formulations do not explicitly model packet loss or communication delay. Learned predictors may be calibration-sensitive: the lane-change study does not report class balancing, the GP approach notes extrapolation risk under abrupt maneuvers, and the takeover controller states that mismatch in the generic MPC/LQR-like human model reduces optimality. Computational cost can also be nontrivial; in the lane-change ablation, adding velocity and acceleration roughly triples execution time from $d_0$24 to $d_0$25 in training/evaluation, even though inference is expected to be much smaller. A plausible implication is that predictor-based CACC gains are strongest when prediction uncertainty itself is explicitly carried into the controller, rather than treated as an external estimate of fixed quality (Qasemabadi et al., 2023, Mosharafian et al., 2021, Kazemi et al., 2018, Wang et al., 2024, Hu et al., 2024).

The identified research directions are correspondingly structured. The lane-change work points to multimodal sensor fusion, transformer-based sequence models, Bayesian deep learning, uncertainty calibration, and explicit risk-aware control integration. The takeover literature points to driver-specific online identification of the human reaction model and broader mixed-traffic validation. The GP and human-lead works suggest richer uncertainty handling under delay, packet loss, and model mismatch. Delay-compensation studies indicate that exact predictor-feedback remains analytically powerful, but practical deployment increasingly requires coexistence with heterogeneous communication quality, sensing failure modes, and human interaction. Taken together, the field treats prediction not as an auxiliary estimator but as a first-class control primitive: it determines when CACC should tighten a gap, enlarge it, hold it, or relinquish it altogether (Qasemabadi et al., 2023, Wang et al., 2024, Mosharafian et al., 2021, Samii et al., 29 Jul 2025, Samii et al., 7 Apr 2026).