---
title: Constant Time Headway Policy Overview
url: https://www.emergentmind.com/topics/constant-time-headway-policy-cthp
type: topic
---

# Constant Time Headway Policy Overview

Searching arXiv for recent and foundational papers on Constant Time Headway Policy and related ACC/CACC string-stability work.
Constant Time Headway Policy (CTHP), also called constant time gap (CTG), is a longitudinal spacing policy for car-following and platooning in which the desired inter-vehicle gap grows linearly with the follower’s speed. In its canonical form, the commanded gap is written as \(d_0+h\,v_i\), where \(d_0\) is a standstill distance and \(h\) is the time-headway; in simplified ACC identification models the standstill term may be absorbed, yielding \(p_{\mathrm{des}}=\tau v\). CTHP is described as the de-facto spacing law in most Level-1 commercial ACC systems, the industry standard for Cooperative Adaptive Cruise Control, and a prevailing policy in CACC research because it enables tractable spacing-error dynamics and string-stability analysis through low-order transfer functions [2306.03458], [1909.10913], [2606.16374].

## 1. Canonical formulations and notation

In the classical platooning formulation, vehicle \(i\) maintains the longitudinal gap
\[
\tilde p_i \;=\; d_0 + h\,v_i,
\]
with spacing error
\[
e_i \;=\; \tilde p_i - (d_0+h\,v_i).
\]
Equivalent notation appears across the literature. In CACC work with parasitic lag, the standstill-gap error is often
\[
e_i(t)=x_i(t)-x_{i-1}(t)+d,\qquad \delta_i(t)=e_i(t)+h_w\,v_i(t),
\]
where \(h_w\) is the time headway and \(\delta_i\) is the velocity-dependent spacing error. In ACC identification work, the desired spacing is written through the equilibrium relation \(p-\tau v=0\), with \(\tau\) playing the role of the time-headway [2306.03458], [2404.08879].

A widely used commercial-ACC surrogate is the second-order car-following law
\[
\dot p(t)=\Delta v(t), \qquad
\dot v(t)=\alpha\,[p(t)-\tau v(t)] + \beta\,\Delta v(t) + d(t),
\]
where \(p\) is space-gap, \(v\) is follower speed, \(u\) is leader speed, \(\Delta v=u-v\), \(\alpha\) is the gap-error gain, \(\beta\) is the speed-error gain, and \(d(t)\) is a lumped disturbance or modeling error. Its forward-Euler discretization is
\[
p_{k+1}=p_k+T(u_k-v_k),\qquad
v_{k+1}=v_k+T\bigl[\alpha_k(p_k-\tau_k v_k)+\beta_k(u_k-v_k)\bigr],
\]
with \(\alpha,\beta,\tau\) carried as constant states during identification [2306.03458].

In connected-platoon models, each follower is often represented by
\[
\ddot x_i(t)=a_i(t), \qquad \tau\,\dot a_i(t)+a_i(t)=u_i(t),
\]
where \(\tau\in(0,\tau_0]\) is an uncertain actuation lag. A single-predecessor CACC law with communication delay \(\ell\) is
\[
u_i(t)
=k_a\,a_{i-1}(t-\ell)
-k_v\bigl[v_i(t)-v_{i-1}(t)\bigr]
-k_p\,\delta_i(t),
\]
while the CACC\(^+\) generalization sums analogous terms over \(r\) predecessors [2404.08879].

## 2. String stability and analytic conditions

The central systems-theoretic role of CTHP is its connection to string stability: disturbances should attenuate, rather than amplify, as they propagate along a vehicle string. In the frequency domain, one common requirement is
\[
\sup_{\omega\ge 0}|G_i(j\omega)|\le 1,
\]
or, in platoon formulations with parasitic lag, \(\|H(j\omega;\tau)\|_\infty\le 1\) uniformly for all \(\tau\le \tau_0\) [1909.10913], [2509.20722].

For the simplified ACC dynamics linearized around equilibrium, the transfer function from leader-speed perturbations \(\Delta u\) to follower-speed perturbations \(\Delta v\) is
\[
H(s)=\frac{\beta s+\alpha}{s^2+(\alpha\tau+\beta)s+\alpha}.
\]
Two classical strict string-stability conditions follow. \(L_2\) strict string stability requires
\[
|H(j\omega)|\le 1 \quad \forall \omega\ge 0,
\]
which is equivalent to
\[
\alpha^2\tau^2+2\alpha\beta\tau-2\alpha \ge 0.
\]
A sufficient \(L_\infty\) strict string-stability condition is
\[
(\alpha\tau+\beta)^2-4\alpha \ge 0.
\]
The identification literature further notes that if \(\beta^2\ge 2\alpha\), then the two conditions coincide, and that the \(L_2\) requirement is the more conservative one [2306.03458], [2309.01211].

For third-order longitudinal models with parasitic lag, robust string stability is coupled to internal stability. In one standard formulation, the characteristic quasi-polynomial is
\[
D(s)=s^2e^{\tau s}+\gamma s+k_p,
\]
with \(\gamma=k_v+h_wk_p\). The delay-robust design results yield closed-form minimum employable time headways for classical ACC, CACC, and CACC\(^+\). These bounds quantify a recurring theme in the CTHP literature: headway is not only a safety parameter, but also a delay-limited stability parameter [2509.20722].

A notable counterpoint is the cooperative extended state observer (ESO) design for third-order longitudinal dynamics. That work shows that, for its decentralized controller using only onboard sensing, the control parameters can be designed to ensure closed-loop and \(L_2\) string stability for any given positive time headway, even arbitrarily small. This suggests that claims about the “minimum employable headway” are controller-class dependent rather than purely properties of the spacing law itself [2004.09778].

## 3. Identification of hidden CTHP parameters in commercial ACC

Because stock ACC logic and parameters are proprietary, a substantial line of work treats CTHP as an inverse-modeling surrogate and estimates its hidden parameters from onboard measurements of spacing and relative speed. The dual unscented Kalman filter (DUKF) formulation performs joint nonlinear state and parameter identification for \(\alpha\), \(\beta\), and \(\tau\), using empirical observations of space-gap and relative velocity, and explicitly adopts the observability rank condition for nonlinear systems to assess whether the parameters are estimable from data [2306.03458].

In that framework, \(\tau\) is the desired time-headway, \(\alpha\) is the gap-error gain, and \(\beta\) is the speed-error gain. Typical production values are described as \(1.0\,\mathrm s \lesssim \tau \lesssim 1.5\,\mathrm s\) in stock ACC, with initial filter guesses \(\alpha_0=0.1\), \(\beta_0=0.2\), and \(\tau_0=1.2\,\mathrm s\). For the Hyundai Nexo, DUKF estimates from three random initializations were
\[
\hat\alpha=[0.146,\,0.199,\,0.213]\;\mathrm{s}^{-2},\quad
\hat\beta=[0.181,\,0.129,\,0.185]\;\mathrm{s}^{-1},\quad
\hat\tau=[1.122,\,1.164,\,1.131]\;\mathrm s,
\]
with spacing and speed tracking MAEs on the order of \(0.1\,\mathrm m\) in gap and \(0.04\,\mathrm{m/s}\) in speed. Those estimates violate both the \(L_2\) and \(L_\infty\) inequalities, leading to the conclusion that the considered Hyundai Nexo stock ACC is neither \(L_2\) nor \(L_\infty\) strict string stable. A parallel least-squares calibration also failed both criteria [2306.03458].

Physics-inspired neural networks (PiNNs) provide a second inverse-learning route. In that approach, a fully connected network predicts spacing and speed trajectories while a physics loss enforces the CTHP ODE residual. The total loss combines the residual term and data misfit, and optimization is carried out by Adam followed by full-batch L-BFGS. Synthetic validation recovered \(\alpha^*=0.08\,\mathrm s^{-2}\), \(\beta^*=0.12\,\mathrm s^{-1}\), and \(\tau^*=1.5\,\mathrm s\) with \(\hat\alpha=0.0784\), \(\hat\beta=0.1200\), and \(\hat\tau=1.500\). Empirical campaigns at Ispra-Casale, Ispra-Vicolungo, and AstaZero yielded headways around \(1.0\)–\(1.52\,\mathrm s\), gains spanning approximately \(0.0104\)–\(0.1760\,\mathrm s^{-2}\) for \(\alpha\) and \(0.0718\)–\(0.4450\,\mathrm s^{-1}\) for \(\beta\), and MAEs below \(0.6\,\mathrm m\) in spacing and \(0.3\,\mathrm{m/s}\) in speed. The learned parameter sets again showed that the considered stock ACC systems were neither \(\mathcal L_2\) nor \(\mathcal L_\infty\) string stable, with only one AstaZero follower narrowly satisfying the \(\mathcal L_\infty\) condition while remaining \(\mathcal L_2\)-unstable [2309.01211].

A recurrent empirical implication is that commercially implemented CTHP appears behaviorally transparent but stability-marginal: it reliably enforces a nearly constant time gap, yet often places the closed-loop system outside the strict disturbance-attenuation region.

## 4. Delay, communication noise, and multi-predecessor information

Connected-platoon analyses extend CTHP by adding parasitic actuation lag, explicit V2V communication delay, and communicated acceleration feedforward. In the delay-free baseline with onboard sensing only, the minimum employable headway is \(2\tau_0\), where \(\tau_0\) is the maximum parasitic lag. Acceleration feedforward and multi-predecessor look-ahead reduce this bound, but do not eliminate it under the classical controller families studied in the delay literature [1803.02900], [2509.20722].

| Architecture | Lower bound on time headway | Conditions |
|---|---:|---|
| ACC | \(h_{\min}=2\tau_0\) | \(k_a=0\) |
| CACC | \(h_{\min}=2\tau_0/(1+k_a)\) | \(0\le k_a<1\) |
| CACC\(^+\) | \(h_{\min}=4\tau_0/[(1+r)(1+r\,k_a)]\) | \(0\le k_a<1/r\) |
| \(r\)-predecessor pos./vel. | \(h_{\min}\ge 4\tau_0/(1+r)\) | V2V, no acceleration |
| \(r\)-predecessor acceleration | \(h_{\min}\approx 2\tau_0/(1+r)\) | as \(r\,k_a\to 1\) |

When explicit communication delay \(\ell\) is present, the CACC transfer function takes the form
\[
H_1(s)=\frac{\mathcal N_1(s)}{\mathcal D_1(s)},\qquad
\mathcal N_1(s)=k_a s^2 e^{-\ell s}+k_v s+k_p,\qquad
\mathcal D_1(s)=\tau s^3+s^2+(k_v+h_wk_p)s+k_p.
\]
Robust string stability requires \(\mathcal D_1(s)\) Hurwitz and \(\sup_{\omega\ge 0}|H_1(j\omega)|\le 1\). A sufficient headway requirement is
\[
h_w\ge \max\!\Bigl\{\frac{2(\tau_0+k_a\ell)}{1+k_a},\frac{\ell}{2}\Bigr\}(1+\eta),
\]
and for CACC\(^+\),
\[
h_w\ge \frac{4(\tau_0+r\,k_a\ell)}{(r+1)(1+r\,k_a)}(1+\eta_r).
\]
The delay enters the bound as the combination \(\tau_0+k_a\ell\), so more delay implies larger headway [2404.08879].

Noisy V2V communication changes the feedforward gain through an averaged effective gain
\[
\widetilde k_a = k_a \bar w,
\]
where the multiplicative channel noise is parameterized by an SNR variable \(\rho>1\). In that setting, the lower bound becomes
\[
h_w
\;\ge\;
2\tau_0\,
\frac{1-(1-1/\rho)\,k_a}
{1-(1+1/\rho)^2\,k_a^2},
\]
with admissible \(k_a\in(0,1/(1+1/\rho))\). Minimizing the right-hand side yields
\[
k_a^*
=
\Bigl(\frac{1 - 1/\sqrt\rho}{1 + 1/\sqrt\rho}\Bigr)\frac1{1+1/\rho},
\qquad
h_{w,\min}
=
\tau_0\,\frac{(1+1/\sqrt\rho)^2}{1+1/\rho}.
\]
For \(\rho\to\infty\), one recovers the perfect-communication limit \(h_{w,\min}\to\tau_0\) [2404.08889].

The communication literature therefore treats CTHP as a co-design problem: headway, gains, and information architecture are selected jointly, not sequentially.

## 5. Safety-oriented and learning-augmented extensions

A major limitation of classical CTHP is that steady-state spacing rules do not automatically provide transient safety during large relative-velocity excursions. The dynamic-headway framework addresses this by augmenting the nominal CTH spacing with a quadratic safety margin
\[
q_i=\frac{v_i^2-v_{i-1}^2}{2a_{\max}},
\]
and, for the first follower,
\[
\bar e_1
=
\tilde p_1-d_0-h\,v_1-\max(0,q_1).
\]
The imposed policy becomes
\[
\tilde p_1(t)
=
d_0+h\,v_1(t)+\max(0,q_1(t))+\bar e_1(t),
\]
which converges to classical CTHP at steady state because \(q_1\to 0\) and \(\bar e_1\to 0\) when velocities synchronize. The associated nonlinear saturated controller renders the unsaturated error dynamics strictly dissipative,
\[
\dot{\bar e}_1
=
-\lambda\Bigl(1+\frac{v_1}{h}\Bigr)\bar e_1,
\]
and the paper proves forward invariance of the first follower’s safe set via Nagumo’s theorem when \(\lambda\ge h/(d_0-d_s)\). By induction, safety of the first pair and non-negativity of \(v_1\) suffice to keep the entire string safe [2606.16374].

This result does not replace CTHP; rather, it wraps CTHP in a transient-safety envelope. The paper’s formulation is explicit that the downstream followers may use any standard CTHP-based CACC law satisfying asymptotic stability of the classical spacing error. A plausible implication is that CTHP remains the steady-state operating law, while safety certification during aggressive transients may require an additional state-dependent margin.

Learning-augmented controllers preserve the same spacing template while modifying the control law. In PERPL, the physics-based baseline is the PD-type CTG controller
\[
u_{n,\mathrm{phy}}(t)=K_a e_{s,n}(t)+K_v e_{v,n}(t),
\]
with delayed predecessor measurements and first-order actuator lag. A PPO-based residual then adjusts the physics action, followed by a quadratic-program safety barrier that constrains the resulting headway to \([1\,\mathrm s,3\,\mathrm s]\). In the reported single-vehicle extreme test, headway RMSE was \(1.726\) for the linear controller, \(2.429\) for PPO alone, and \(0.419\) for PERPL; in a ten-vehicle mixed platoon, PERPL reduced headway RMSE from \(0.439\) to \(0.204\), improved damping ratio from \(0.616\) to \(0.558\), and reduced comfort cost from \(0.301\) to \(0.249\) relative to the pure linear baseline [2409.15595].

## 6. Traffic-flow, energy, and network-level consequences

The traffic-engineering significance of CTHP is its direct coupling between speed and space consumption. Because the desired gap grows with speed, CTHP supports string stability with simple one-vehicle look-ahead structures, but it reduces capacity at higher velocities relative to constant-distance policies. This trade-off is central in comparisons between CTG and constant distance headway (CDH) or constant distance gap (CDG) policies [1909.10913].

In signalized-intersection studies, a calibrated reference CTG uses \(h=0.87\,\mathrm s\) and \(d_0=2.95\,\mathrm m\), derived from \(3\,500\) real-world start-up traces. In a single-traffic-light start-up with \(T_g=15\,\mathrm s\), CTG-Ref yields about \(8\) vehicles, or approximately \(32\) vehicles/min, whereas CDG at \(8.1\,\mathrm m\) yields \(17\) vehicles, or approximately \(68\) vehicles/min, a \(+140\%\) increase. In full intersections, CTG-Ref throughput is about \(5\,250\) veh/h at \(T_g=15\,\mathrm s\), while CDG reaches approximately \(11\,550\) veh/h when no turnings are present. Large-scale simulations further show that CDG can outperform CTG by \(50\)–\(100\%\) when signal timing and intersection spacing avoid blocking disturbances, but long green cycles can cause CDG platoons to block adjacent junctions. The paper therefore advocates situational use of CDH/CDG and notes that CTG remains the safer, more predictable choice in mixed traffic and longer corridors [1909.10913].

Empirical platoon measurements also connect CTHP to energy propagation. In two five-vehicle campaigns, ACC followers clustered tightly around headways of roughly \(1.3\)–\(1.4\,\mathrm s\), whereas human drivers adopted larger and more variable headways of about \(2.1\)–\(2.4\,\mathrm s\). The reported findings are that ACC systems propagate an increasing energy consumption upstream while human drivers do not; ACC succeeds in maintaining a constant time-headway policy, operating very reliably; and ACC operates in different phase-space areas with room for improvement. Quantitatively, tractive energy grew by \(10\)–\(25\%\) per vehicle upstream in ACC runs, while human runs varied within \(\pm 5\%\). Speed variance likewise increased upstream for ACC, indicating string instability in the empirical sense used in that study [2306.04476].

At the network level, headway is treated not only as a local controller parameter but also as a capacity-control variable. In mixed-autonomy routing, the headway of autonomous cars \(\beta^a_e(t)\) is the control input, constrained by \(\beta_{\min}\le \beta^a_e(t)\le \beta_{\max}\), with \(\beta_{\min}=1\,\mathrm m\) and \(\beta_{\max}=10\,\mathrm m\) in the reported experiments. Link capacity depends on headway through
\[
C_e(\alpha_e,\beta^a_e)
=
\frac{v_e^{\mathrm f}b_e}
{\alpha_e\,\beta^a_e(t)+(1-\alpha_e)\beta^h_e}.
\]
A PPO policy that dynamically regulates autonomous-car headway reduced total travel time by \(4.0\%\) relative to a uniform constant headway and by \(11.1\%\) relative to minimum feasible headway on a five-link Braess network; on an eight-link extension, the improvements were \(2.8\%\) and \(0.2\%\), respectively. The reported mechanism is that naïvely minimized headway can induce Braess-type congestion, whereas dynamic headway control can keep at least one corridor in free flow [2303.04266].

Taken together, these results locate CTHP at the intersection of control, safety, and traffic-flow theory. It remains the dominant baseline because of its analytical tractability and implementation simplicity, yet the literature consistently shows that practical performance depends on controller architecture, delays, communicated information, transient-safety augmentation, and the traffic regime in which the policy is deployed.

Source: https://www.emergentmind.com/topics/constant-time-headway-policy-cthp