12664 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO.

10, OCTOBER 2023

Centralized Model Predictive Control With

Human-Driver Interaction for Platooning
Justin M. Kennedy , Member, IEEE, Julian Heinovski , Student Member, IEEE,
Daniel E. Quevedo , Fellow, IEEE, and Falko Dressler , Fellow, IEEE

Abstract—Cooperative adaptive cruise control presents an op-

portunity to improve road transportation through increase in road
capacity and reduction in energy use and accidents. Clever design
of control algorithms and communication systems is required to
ensure that the vehicle platoon is stable and meets desired safety
requirements. In this paper, we propose a centralized model pre- Fig. 1. System model of the used CACC concept: a platoon controller regulates
dictive controller for a heterogeneous platoon of vehicles to reach the velocity and inter-vehicle distances utilizing the vehicles’ position p, velocity
a desired platoon velocity and individual inter-vehicle distances v, and acceleration a, exchanged via inter-vehicle communications.
with driver-selected headway time. As a novel concept, we allow
for interruption from a human driver in the platoon that tem-
porarily takes control of their vehicle with the assumption that the
driver will, at minimum, obey legal velocity limits and the physical Adaptive Cruise Control (ACC) systems use on board sensors
performance constraints of their vehicle. The finite horizon cost to measure the distance and velocity of a predecessor vehicle to
function of our proposed platoon controller is inspired from the
infinite horizon design. To the best of our knowledge, this is the operate an autonomous cruise control system. However, these
first platoon controller that integrates human-driven vehicles. We systems are prone to string stability issues resulting in ghost
illustrate the performance of our proposed design with a numerical traffic jams [7]. To ensure stability of the platoon, either a
study, demonstrating that the safety distance, velocity, and actua- large, velocity-based inter-vehicle distance is required using
tion constraints are obeyed. Additionally, in simulation we illustrate a headway time [8] or more than just the preceding vehicle’s
a key property of string stability where the impact of a disturbance
is reduced through the platoon. state is required. By utilizing inter-vehicle communication sys-
tems, Cooperative Adaptive Cruise Control (CACC) systems
Index Terms—Cooperative adaptive cruise control, human- can reduce the inter-vehicle distance and avoid string stability
driver interaction, hybridized cyber-physical systems, model
issues by sharing the desired control actions from other vehicles.
predictive control, platooning.
Fig. 1 illustrates our system model for communication, which
represents a potential CACC design of a coordinated platoon of
I. INTRODUCTION vehicles driving with small inter-vehicle distances with a platoon
communication network.
UTONOMOUS vehicle platooning with inter-vehicle
A communication permits road vehicles to travel close to-
gether increasing road capacity while reducing energy use and
Many platoon control designs have focused on the technical
aspects of algorithm design, communication constraints, or ex-
perimentation isolated from real traffic [9]. However, studies
associated vehicle emissions [1]. This cooperative connected of driver behavior have found that human drivers maintain
cruise control technology can reduce the incidence of so-called small inter-vehicle distances below safety margins [10]. This
ghost traffic jams [2] and highway accidents [3]. The au- effect is exacerbated when humans drive when near autonomous
tonomous cruise control problem was first posed as a centralized platoons [11]. As pointed out in [12], during implementation of
platoon design approach in [4], and has seen recent attention with vehicle platooning systems, it will be important to incorporate
several survey papers [2], [5], [6]. legacy vehicles that are unable to integrate with a platoon com-
munication system and human-driver interaction for passenger
Manuscript received 10 February 2023; revised 18 April 2023 and 15 May comfort and well-being, e.g., some passengers feel uncomfort-
2023; accepted 15 May 2023. Date of publication 18 May 2023; date of current able with too small safety gaps, as well as motion sickness.
version 17 October 2023. This work was supported by the Project NICCI2 funded Additionally, some drivers may wish to switch automation lev-
by the German Research Foundation (DFG) under Grants DR 639/23-2 and QU
437/1-2. The review of this article was coordinated by Dr. Tao Dusit Niyato. els, connecting or disconnecting from the CACC. The inclusion
(Corresponding author: Justin M. Kennedy.) of human factors is in line with research integrating human
Justin M. Kennedy and Daniel E. Quevedo are with the School of Elec- behavior and Cyber-Physical System (CPS), moving from CPS
trical Engineering and Robotics, Queensland University of Technology, Bris-
bane, QLD 4000, Australia (e-mail: [email protected]; daniel.quevedo to Hybridized Cyber-Physical System (H-CPS) which is also
@qut.edu.au). called Cyber-Physical-Social System (CPSS) [13].
Julian Heinovski and Falko Dressler are with the School of Electrical En- While modern communication systems can enable the trans-
gineering and Computer Science, TU Berlin, 10623 Berlin, Germany (e-mail:
[email protected]; [email protected]). mission of large volumes of data very quickly between physi-
Digital Object Identifier 10.1109/TVT.2023.3277451 cally separated vehicles [14], the inherent reliability limitations
0018-9545 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
KENNEDY et al.: CENTRALIZED MODEL PREDICTIVE CONTROL WITH HUMAN-DRIVER INTERACTION FOR PLATOONING 12665

controller. The exogenous inputs to our controller are the desired

platoon velocity and individual inter-vehicle distances, and the
desired safety constraints on the inter-vehicle distance, velocity,
and acceleration. The output from our controller is the desired
control action for every vehicle in the platoon.
In MPC design, the control law is based on a prediction of the
states of the system. To compute a mathematical prediction of
the system states, we utilize an abstract mathematical model of
the system dynamics. This technique is widely used in control
engineering practice [19] and [20].
To ensure convergence of the vehicle positions and velocities
to the desired references over a finite horizon, we design a time-
varying reference for all vehicles in the platoon. We consider
Fig. 2. Platoon control architecture. The exogenous inputs to the controller a combined mathematical model of the platoon of individual
are the desired inter-vehicle distances and platoon velocity to form the platoon
reference. The output is the control action for all vehicles in the platoon. The vehicle dynamics to predict the motion of all of the vehicles
controller reconfigures when a human-driver temporarily takes control. The from the multiple control inputs. Inspired by the infinite horizon
block connected by the dashed lines engages predicting a minimum control optimal control algorithm [21], we propose a finite horizon cost
action of the human-driver.
function. We introduce time-varying references to encode the
desired trajectories of positions, velocities, and accelerations
of wireless communications systems, including from packet for all vehicles in the platoon. We emphasize that in our de-
delays and dropouts, can degrade the information available to a sign, the desired inter-vehicle distance and platoon velocity are
control system. When implementing a CACC system for a practi- allowed to change over time. This feature enables a controlled
cal vehicle platoon, the information available to the local control dynamic behavior of the platoon. Our cost function penalizes the
system at each individual vehicle should be considered. When ig- quadratic error of the predicted vehicle positions, velocities, and
noring communication imperfections such as delays and packet accelerations to the desired time-varying references, as well as
loss, a centralized platoon control design that incorporates con- the predicted inter-vehicle distances to the desired inter-vehicle
straints and guarantees safe inter-vehicle distances could lead distances. The use of a quadratic cost function allows for simple
to optimal decisions. Therefore, dealing with realistic wireless application of hard constraints on the cost function in a quadratic
communications, a number of decentralized and distributed program optimizer to guarantee minimum and maximum inter-
control designs have been proposed using local information vehicle distances, velocities, and accelerations of all vehicles in
in combination with information transmitted from neighboring the platoon. The inclusion of constraints in the controller ensures
cars (usually those in front), cf. [2], [5], [8], [12]. However, the that the control actions do not result in unsafe vehicle behavior,
lack of full information of all platoon members means that the including reversing, extreme accelerations, and most critically,
optimality of the control actions may be degraded. This mo- collisions between multiple vehicles.
tivates the development of decentralized control architectures. In our design, we include a switch to reconfigure our controller
One aim in the design of decentralized controllers is to recover for a vehicle temporarily under human driver control, shown as
as close as possible the performance provided by a centralized the dashed line in Fig. 2. Early implementations of potential
controller that has access to all system and state information [15]. CACC designs will need to operate with non-CACC enabled
While many decentralized CACC designs have been posed, a vehicles and with drivers that choose to switch automation
centralized platoon control design that incorporates constraints levels [12]. Our approach accommodates a human driver to
and guarantees safe inter-vehicle distances has not yet been make an emergency brake, reduce speed, or travel at a different
presented. This is of particular interest, when dealing with het- inter-vehicle distance. Our method could also be utilized to
erogeneity and human-driver interaction, which cannot easily be incorporate legacy vehicles. When a vehicle is human driver
realized with incomplete information in a distributed way. Thus, controlled, we adjust our time-varying reference to return the
to enable comparison to decentralized or distributed control platoon to the desired reference and remove the vehicle from the
designs with a limited or reduced information availability, it platoon control. A key feature of our method is that it incorpor-
is first necessary to understand centralized CACC designs. ates the vehicle states such that the control actions for the platoon
In this paper, we present a novel constrained Model Predictive still ensure the inter-vehicle distance constraints for the other
Control (MPC) approach for the centralized control of a platoon vehicles in the platoon. We predict the control actions of the
of heterogeneous vehicles. In our design, we include reconfig- human driver by assuming that the driver will only change their
uration under temporary human-driver control. The goal of our control action to obey a minimum set of constraints of the legal
controller is to reach a desired platoon velocity and individual road speed limits and their vehicle performance constraints. Our
inter-vehicle distances, while guaranteeing safety constraints. centralized platoon control design ensures safety for the remain-
The desired inter-vehicle distance is based on driver-selected ing platoon-controlled vehicles by producing control actions that
headway times which are variable between individual vehicles maintain safe inter-vehicle distances. More complex predictive
in the platoon, and can change over time [16], [17], [18]. Fig. 2 approaches, such as economic cost functions [22] or inter-
illustrates the control loop of our proposed centralized platoon vehicle interactions, could be utilized to predict a human driver.

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
12666 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 10, OCTOBER 2023

The key contributions of this paper are block the transmission and the number of neighboring vehicles to
r proposal of an optimal algorithm for centralized platoon calculate the best beacon interval. Focusing specifically on pla-
control of heterogeneous vehicles with constraints, and tooning, Segata et al. [28] proposed a dynamic approach called
r reconfiguration under temporary human-driver control by jerk beaconing, which exploits vehicle dynamics to share data
incorporation of a human-driver model. only when needed by the controller. Here, the beacon interval
is computed dynamically based on changes in acceleration over
II. RELATED WORK time, i.e., the jerk. This approach shows huge benefits in terms
of network resource saving and is able to keep inter-vehicle
In this paper, we propose an MPC design to perform the
distance close to the desired gap even in highly demanding
CACC task to safely control a platoon of vehicles to the desired
scenarios.
inter-vehicle distances and velocity. Through this section, we
Going beyond IEEE 802.11p, Segata et al. [29] proposed the
present the literature on communication protocols and control
Distributed EDCA Bursting (DEB) protocol, which extends
law designs for CACC, to argue that MPC is an appropriate
the frame busting mechanism of IEEE 802.11p such that only
technology for CACC with interruption by human drivers. We
the platoon leaders content for the channel. In case of success-
highlight that while many control designs have been posed,
ful channel reservation, all vehicles in the platoon transmit a
a centralized control design that obeys safety constraints and
coordinated burst of frames, thus, sharing the platoon leader’s
incorporates human drivers has yet to be posed.
transmission opportunity. This helps overcoming channel limits
by reducing the number of nodes contending for the channel and
A. CACC Communication improving spatial reuse. Amjad et al. [30] extend IEEE 802.11p
In order to work reliably and with small safety gaps, CACC by adding a full-duplex relaying system, which enables platoon
requires periodic updates of vehicles’ data (e.g., acceleration, members to simultaneously receive and relay the leader beacons.
speed, position). Typically, the data from at least the vehicle in 2) Cellular V2X (C-V2X): Albeit all of the above protocols
front and often also the first vehicle (i.e., the platoon leader) is and modifications, IEEE 802.11p alone seems not to be sufficient
necessary. If the updates arrive with a high enough frequency for for meeting the strict requirements of CACC (i.e., ultra-low
the control system to react properly, string-stability, i.e., keeping reliability and latency) [31]. The most prominent alternative
the desired gaps without accumulating control errors throughout for enabling V2X communication is Cellular V2X (C-V2X),
platoon members, can be achieved [23]. If the updates are which uses 3GPP standardized 5 G cellular networks. Radio
delayed, “string-stability is seriously compromised” [24]. resources are scheduled by either the base station if vehicles are
1) IEEE 802.11p: Up to a few years ago, the main Vehicle- in-coverage (operation mode 3) or by a distributed resource allo-
to-Everything (V2X) technology considered was IEEE 802.11p cation scheme if vehicles are out-of-coverage (operation mode
as a basis for quite advanced protocol families such as European 4). The latter allows vehicles to select resources in a stand-alone
Telecommunications Standards Institute (ETSI) ITS-G5 [1]. fashion with semi-persistent scheduling. While mode 3 in gen-
The most simple approach for exchanging the vehicle up- eral allows for high packet receptions ratios, mode 4 produces
dates is to use static beaconing, where vehicles broadcast their lower beacon update delays [32], which are also required for
information in regular, periodic intervals. Yet, static beaconing platooning.
can lead to a congested channel, especially in highly dense For example, Vukadinovic et al. [33] compare IEEE 802.11p
scenarios, e.g., with long or many platoons, thus reducing the to 3GPP C-V2X based on LTE in both operation modes for truck
stability of a platoon. Thus, Segata et al. [18] proposed to use platooning. Results show that C-V2X in both modes allows for
slotted beaconing, which splits the time for the leader beacon shorter inter-truck distances than IEEE 802.11p due to more reli-
into transmission slots for all platoons members. The authors able communication in a congested wireless channel. However,
show that this can greatly improve the beaconing performance short communication distances and large vehicle densities seem
in crowded scenarios, especially when combined with transmit to be covered better with IEEE 802.11p instead of C-V2X [34].
power control, thus, reducing the load and improving the relia- Therefore, general modifications for improving the scheduling
bility. of sidlelink radio resources in mode 4 have been proposed [35],
In order to reduce the channel load further, dynamic beaconing [36]. In order to reach the performance required for CACC,
schemes have been proposed. Sommer et al. [25] presented Hegde et al. [37] propose to schedule the sidelink radio resources
the Adaptive Traffic Beacon (ATB) protocol, which adaptively for the platoon members by the platoon leader. Similarly, the
adjusts the beaconing period according to the current channel radio resource coordination method by Campolo et al. [38]
quality and the message utility. Following up on this, ETSI fulfills the ultra-low latency requirements of CACC and is able to
standardized Decentralized Congestion Control (DCC) [26]. provide spatial reuse of LTE resources among platoon members.
It uses a simple final state machine to adjust, among others, 3) RADCOM: Complementary to IEEE 802.11p and C-
beacon interval and transmit power based on the current ob- V2X, joint communication and sensing approaches, also known
served channel busy ratio. Sommer et al. [27] proposed DynB as Radar-based Communication (RADCOM), have been pro-
to avoid overloading the wireless channel and allow low-latency posed. Following the trend of using higher communication fre-
communication by using very short beaconing intervals. The quencies for radio communication, Millimeter Wave (mmWave)
protocol continuously observes the channel load and considers technologies have recently become interesting to the V2X re-
detailed radio shadowing effects, even by moving vehicles, that search community. mmWave technology promises high bitrates

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
KENNEDY et al.: CENTRALIZED MODEL PREDICTIVE CONTROL WITH HUMAN-DRIVER INTERACTION FOR PLATOONING 12667

and low delays due to its wide channel bandwidth and dynamic modeling approaches have used the energy based port-
beam-forming [39]. However, using it as a single communication Hamiltonian system model [44], [58]. While use of a homoge-
technology may be difficult due to its highly volatile transmis- neous platoon with identical dynamics makes the control design
sion channel, especially in an automotive environment [39], [40]. and tuning simpler, it is unrealistic to real world heterogeneous
Nevertheless, initial works indicate that mmWave can be very platoons of different vehicles [59]. Certain controller stability
valuable when complementing the other alternatives [14], [41]. properties can change with different types of vehicles such
as platoons of heavy vehicles [60], and environment effects
including changes in road slope [61] and wind [59].
B. CACC Controller A variety of control design approaches have been proposed in
Control design for vehicle platooning has focused on meeting the literature including the classic Linear Quadratic Regulator
string stability conditions with several definitions in the litera- (LQR) [4], [45], [62], Proportional Integral Derivative [63],
ture [42]. In addition to stability requirements, the control design H-infinity [59], sliding mode control [43], and MPC. MPC
also needs to consider the information flow topology arising algorithms optimize a finite horizon cost function at each time
from the available communication links, formation geometry or step, and allow for the inclusion of hard constraints [64], such as
spacing policy, vehicle dynamics, and desired platoon conver- road speed limits and minimum safe inter-vehicle distances. The
gence. desired control actions from a constrained MPC controller will
The information flow topology of how information is shared not exceed a vehicle’s performance limit or control vehicles into
between vehicles influences both the control algorithm design situations that could lead to an accident. Additionally, so-called
and the required communication system. Many control designs economic MPC [65], that assigns real values, such as fuel costs,
utilize a leader-follower approach where a lead vehicle sets to the weights in the cost function, has been utilized [66] to link
the platoon speed and each follower vehicle maintains their vehicle performance to an energy or financial metric.
own spacing to the predecessor, such as the sliding mode con- Distributed MPC algorithms in the leader-follower approach
troller in [43] and employed in [18]. Other designs consider have been applied to platoons with poor communications [46],
bi-directional information sharing from the neighboring vehicles with extension to heterogeneous platoons [67], and string sta-
such that leader information is not required, e.g. [44] and [45]. bility was enforced using constraints [68]. A more complex
These distributed approaches consider that the lead or reference approach was employed in [69] to include network information
vehicle is exogenous to the platoon controller [46], either con- as a delay on the desired control action in the dynamics.
trolled by a human driver or by a separate ACC system [18]. Most control designs for the platooning of vehicles consider
Many designs focus on Vehicle-to-Vehicle (V2V), also of in- a distributed approach with the use of a lead or ego vehicle that
terest are V2X where infrastructure can monitor and coordinate provides an input reference to the platoon. This is a flexible
a platoon [5], as well as interactions with other platoons [47]. approach as it allows for control designs to break apart and
Recent works have included unreliable communication channels reform platoons [5] However, distributed policies have been
in the control design using time delays [48] and packet loss [49]. introduced that slow front vehicles and speed up later vehicles to
A key design factor is the formation geometry of the inter- form a platoon, while observing that this may be in conflict with
vehicle distances [8], which was considered fixed in early the lead driver’s goal of reaching their destination quickly [70].
works [4] but caused string instability for ACC [42]. To achieve Additionally, [53] noted that the use of a variable headway and
string stability for ACC, [50] proposed a velocity based spacing ACC controller introduced a “group conscience” such that the
policy following the concept that a human driver should follow leading vehicles were designed with reduced performance to
a preceding car with a certain headway time, with refinements take into account later vehicles in the platoon. A centralized
in [51] and [52]. To account for the slower braking performance control design would utilize all platoon information and a pla-
of heavy vehicles, a variable spacing policy with the headway toon reference to design the control actions for the vehicles as a
a function of the difference in velocity [53]. The authors note collection.
that if lead vehicle information is shared, then the headway is The original work on the control of vehicular platoons is [4].
able to be reduced to zero [54]. Often a common (non-unique) The authors designed a centralized LQR controller that took a
and constant headway time is utilized [55]. Alternative spacing target reference velocity for the platoon and desired inter-vehicle
policies have also included use of the traffic density [56], as spacing, to generate the control action for all vehicles in the
constant time headway can result in unstable traffic flow [57]. platoon, which was furthered in [71]. However, in [21] it was
The vehicle dynamics used for control design of vehicle pla- shown that the original cost function in [4] is not string stable
toons have included complex models that model torque output as the length of the platoon goes to infinity, such that as more
of the engine with variable gear ratios as well as simplified linear vehicles are added the convergence time expands, and the initial
models. In reference to the nonlinear engine and gearbox models, control action increases. The authors posed an alternative state
it is noted in [43] that a first order lag model is suitable for higher representation and cost function that penalized both the abso-
level control of the vehicle, such as for platooning applications. lute position error to the reference as well as the inter-vehicle
This simplifies the vehicle, engine, and braking systems into distances to achieve finite convergence [21].
a single constant. In [18], [43] the mechanical lag coefficient However, this approach is criticized in [72] which shows that
for a standard passenger vehicle is assumed to be τi = 0.5 [s], an infinite length platoon is not equivalent to a large but finite
with a heavy vehicle having a larger coefficient. Alternative platoon. In [73] it is shown that the optimal control design

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
12668 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 10, OCTOBER 2023

fails for certain initial conditions with large control values such that the platoon dynamics are
resulting from the static gain computed from the LQR such
Xk+1 = AM Xk + BM Uk + Wk (4)
that desired control action could be larger than the maximum
allowable control. The poor performance of large platoons also where AM and BM are block diagonal matrices of the collection
occurs in decentralized designs where the state feedback control of single-vehicle dynamics and control input matrices, Wk is a
gain reduced for vehicles further away [74]. This reduction in vector of the i.i.d. process noise acting on each vehicle which
state feedback gain was used to argue that for an M length can be modeled as Wk ∼ N (0, W). The matrices AM and BM
vehicle platoon, there should M independent controllers with are given in Appendix B.
M separately tuned gains [75]. In MPC design, instead of directly computing the platoon
control action Uk , we optimize for the change in control actions
III. PLATOON ARCHITECTURE ΔUk from the previous control action such that the applied
In this section, we state the single vehicle dynamics and de- control (3) to the platoon dynamics (4) is
velop the centralized platoon model of heterogeneous vehicles. Uk = Uk−1 + ΔUk (5)
We then state our human driver model.
(1) (M ) (i)
where ΔUk = [Δuk , . . . , Δuk ]T and Δuk is the change
A. Vehicle Dynamics in control action for vehicle-i. This optimization is computed
over a finite horizon of N time steps into the future. We use the
We consider a commonly utilized linear dynamics for longi- platoon model (4) with (5) to predict the value of the state over
tudinal motion of a vehicle-i. Let us define our state variables the next N time steps. We introduce the predicted state value of
as p(i) [m] as a point at the front bumper, v (i) [m/s] the velocity, the platoon at time k + j for j ∈ {1, . . . , N } from the measured
a(i) [m/s2 ] acceleration, and u(i) [m/s2 ] control input or desired state value at time k using the model denoted as X̂k+j|k , with
acceleration. Following [43], the time derivative of each state is the prediction window defined as
defined as ṗ(i) = v (i) , v̇ (i) = a(i) , and  T
T T
1 (i) 1 Xk = X̂k+1|k , . . . , X̂k+N |k
ȧ(i) = − a + u(i)
τi τi
for the predicted value of the change in control from the platoon
where τi [s] is the mechanical actuation lag. controller as
We write the state vector of a single vehicle-i as x(i) = T T T
(i) (i) (i) T
[p , v , a ] , which gives the standard state space form ΔÛk = [ΔÛk|k , . . . , ΔÛk+N −1|k ]

ẋ(i) = A(i) (i)

+ Bc(i) u(i) , where from the measurement at time k the predicted applied
c x (1)
control at time k is Ûk|k = Uk−1 + ΔÛk|k and the predicted
(i) (i)
where Ac and Bc are the dynamics and control input matrices control at time k + j is Ûk+j|k = Ûk+j−1|k + ΔÛk+j|k for j =
with the mechanical lag term for vehicle-i and are given in {1, . . . , N − 1}.
Appendix A. We note that ΔUk is the change in control applied at time
Following [76], a continuous-time system (1), can be dis- k, while ΔÛk is the predicted change in control over the finite
cretized in time with constant sampling interval (period) Δt horizon of length N . The actual applied control action is not
[s] to necessarily equal to the prediction.
(i) (i) (i) (i) Using algebraic manipulation as illustrated in MPC texts
xk+1 = A(i) xk + B (i) uk + wk (e.g. [64]) the state prediction of the platoon Xk can be written
(i) as a linear combination of the current state Xk , the previous
where the subscript k is discrete-time, wk is i.i.d. process
applied control Uk−1 and the predicted change in control ΔÛk
noise representing error in the discrete-time prediction model,
modeled as zero mean normally distributed with covariance Xk = ΦXk + λUk−1 + ΓΔÛk (6)
(i)
W (i) > 0, wk ∼ N (0, W (i) ) and the dynamics matrices are
where Φ is the propagation of the state through the dynamics
discretized using
matrix AM , λ and Γ are the propagation of the control inputs
 Δt
(i) (i) (i)
through the dynamics and control matrices, and are given in
A = exp(Ac Δt ) and B = exp(A(i)
c m)dm Bc
(i)
Appendix B.
0
We utilize this state prediction model to design a centralized
and are given in Appendix A. MPC for the coordinated control of a platoon of vehicles.
Following [4], [71] and [21], we consider a combined model of
the platoon. For M vehicles, we define the centralized multiple- B. Human Driver Model
output state and multiple-input control vectors as
 T We consider scenarios where during operation of the platoon
(1) (M ) (1) (M ) (1) (M )
X k = pk , . . . , p k , v k , . . . , v k , a k , . . . , a k (2) a human driver temporarily takes control of their vehicle. This
raises the critical challenge that our platoon controller must
 T
(1) (M ) reconfigure to this human driver to ensure safety of the platoon.
Uk = uk , . . . , u k (3)
A human driver may choose to modify their interaction with

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
KENNEDY et al.: CENTRALIZED MODEL PREDICTIVE CONTROL WITH HUMAN-DRIVER INTERACTION FOR PLATOONING 12669

the CACC automation due to traffic conditions or for safety ()

where Φ̄, λ̄ and Γ̄ are given in Appendix B and ΔÛk =
reasons [12]. After switching to manual operation, we make no () ()
[Δûk+1|k , . . . , Δûk+N −1|k ]T is the prediction of the human
assumption about the human driver’s decision to maintain a new
driver change in control of vehicle-.
velocity, or inter-vehicle distance. While algorithmic methods
We utilize (10) to compute a basic prediction of the human
could be utilized to produce a prediction of the human driver
driver’s control action, which we can utilize in the platoon
actions, such as economic MPC [22] or human behavior [12],
prediction model (9) to design a centralized MPC for the co-
these methods would require an assumption or model of the
ordinated control of a platoon of vehicles.
intended human driver action. For our CACC design, we make
limited assumptions. We assume that the human driver is solely
focused on the state of their own vehicle, does not interact with IV. PLATOONING PROBLEM
any other vehicles in the platoon, and issues control actions that We desire to control the entire platoon to reach a target velocity
are consistent with physical (engine limit) and legal (road speed of vd [m/s] with the desired distance between vehicle-i and its
limit) constraints. immediate predecessor vehicle-(i − 1) as
Consider a vehicle- has temporarily left the platoon and
d¯k  di + hk vk = li−1 + ri + hk vk
() (i) (i) (i) (i) (i)
has the change in control action Δuk from the human driver (11)
()
replacing the platoon control Δuk such that the control action where di = li−1 + ri is the constant inter-vehicle distance, li−1
() () ()
is uk = uk−1 + Δuk . For ease of notation we modify the [m] is the length of vehicle-(i − 1), ri [m] the desired standstill
platoon change in control action (5) with a switch (i)
distance in front of vehicle-i, and hk [s] the desired headway
time. The headway time quantifies the distance to the preceding
Uk = Uk−1 + αk ΔUk + ᾱk ΔUk (7) vehicle at the current velocity. The desired standstill distance ri
(i)
where ΔUk is the control action applied from a human driver, a and headway time hk are vehicle specific and can be chosen
binary switch αk as a diagonal square matrix of size M that takes by the respective driver, whereas the desired velocity is platoon
ones on the diagonal for the vehicles controlled by the platoon specified.
and zero in the i, ith element when vehicle-i is not controlled by We consider a unique headway time for each vehicle, which
the platoon controller, and ᾱk = IM − αk . When the platoon can be modified by the occupants of the vehicle. Commercially
is fully controlled by the centralized platoon controller αk ≡ available ACC systems allow for user selection of headway
IM and ᾱk ≡ 0M , and (7) reduces to (5). The dynamics of the time [8], with increments at 1, 1.5 and 2 seconds [77]. We write
(i)
platoon (4) are now the individual headways, hk , as a function of time k, to indicate
that these can be modified but consider that a reasonable driver
Xk+1 = AM Xk + BM Uk−1 + BM αk ΔUk + BM ᾱk ΔUk . would not be constantly changing their headway.
(8) Additionally, we desire to ensure the following constraints for
Based on the applied control at time k − 1, the platoon all vehicles i ∈ {1, M }
controller is aware if every vehicle has utilized the centralized r C1: p(i−1) − p(i) ≥ dmin , minimum safe distance between
platoon controller or an alternative control value. As such, αk is vehicles to ensure that no vehicle impacts its predecessor,
known to the controller at time k. If a vehicle has temporarily r C2: p(i−1) − p(i) ≤ dmax maximum distance between ve-
left the platoon, we assume that the vehicle will continue to be hicles to ensure (random) communications are maintained,
human controlled until informed otherwise, and αk is constant r C3: vmin ≤ v (i) , minimum velocity set to zero on the as-
for the finite prediction horizon. The finite horizon prediction sumption that no vehicle in the platoon reverses on the
for the state of the platoon is expanded from (6) to road,
r C4: v (i) ≤ vmax , maximum velocity chosen based on the
Xk = ΦXk + λUk−1 + Γ(IN ⊗ αk )ΔÛk + Γ (IN ⊗ ᾱk ) ΔÛk road speed limit, or the performance limitation of a vehicle,
(9) r C5: amin ≤ a(i) , minimum acceleration bounded based on
T T T
where ΔÛk = [ΔÛk|k , . . . , ΔÛk+N −1|k ] are future change in the performance of the braking systems, and
controls from the human driver, ⊗ is the Kronecker product and r C6: a(i) ≤ amax , maximum acceleration chosen based on
IN is the identity matrix of size N . the engine performance of the vehicles.
In the following, we design a finite horizon cost function The acceleration bounds could be further limited for the
to find the optimal change in control ΔUk for the platoon, comfort of the vehicle occupants.
which requires knowledge of any human driver control action Finally, we also consider that our proposed controller can
ΔUk . Ideally for the platoon controller, the future human driver accommodate a human driver taking temporary control of their
change in control actions ΔUk are known exactly, however, this vehicle within the platoon. This could include a driver initiating
is unlikely to be the case. This motivates the use of a predicted an emergency brake, reducing speed, or temporarily maintaining
control action for the human driver control values. For a finite a larger distance from the previous vehicle than specified. This
prediction horizon of length N , the human-driver model can be accommodation allows for a human driver to drive within the
written as bounds of the platoon to their own comfort. Additionally, it may
() () () () allow for the inclusion of legacy vehicles. We make the minimum
Xk = Φ̄xk + λ̄uk−1 + Γ̄ΔÛk , (10) assumption that the vehicle and driver will obey performance

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
12670 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 10, OCTOBER 2023

limits of the vehicle: the minimum and maximum accelerations, A. Reference Design
and legal limits on velocity: non-negative and not exceeding the
Proportional state feedback controllers have used in several
road speed limit. platooning works [43], including LQR [4], [21], [62], [71],
We consider the situation where it is more important for the
[74]. However, for constant gain feedback regulators, the control
platoon to stay together, and an emergency brake for one vehicle
value increases the further the states are from the desired refer-
should be obeyed by the platoon. This is in contrast to control ence [80]. In reasonable platooning scenarios [73], such as zero
policies in [5], [62] where each vehicle has individual goals and
initial velocity, the desired initial control actions could exceed
platoons are allowed to split and reform. We desire to ensure that
maximum allowable control action [81].
our proposed controller yields a stable closed-loop to temporary To avoid this issue with constant static references used in
inputs from a human driver to their individual vehicle within the
regulators, we propose a time-varying reference for the desired
minimum constraints. platoon states. Using a slowly increasing reference, all vehicles
We finish our platooning problem by stating our problem in in the platoon are able to converge to the desired reference before
the form of an optimization problem:
the position and velocity references reach the desired steady-
M 
∞  2  2  2 state. This allows convergence to the reference from any initial

+ pk −pk − d¯k
(i) (i−1) (i) (i)
min vk −vd + ΔÛ (i) condition.
k=0 i=1 We consider a slowly increasing ramp for the velocity refer-
ence with constant acceleration from initial time k0 as
subject to: (8), (10), and constraints C1–6. 
First, the inclusion of the system and safety constraints ne- ak Δt k + v̄, k0 ≤ k < k0 + km
vk =

cessitates a careful reference design. As demonstrated in [78], v d , k ≥ k0 + k m
careful design of reference signals may be required to ensure
(i)
that computed control actions smoothly converge the states to where v̄ = min vk0 is initialized to the minimum velocity of the
the desired values while not violating system dynamics and platoon, the acceleration reference is
constraints. We present a suitable time-varying reference in  v −v̄
d
, k0 ≤ k < k0 + k m
Section V-A. a k = Δt k m

Second, it is not possible to solve the infinite horizon cost 0, k ≥ k0 + km
function due to the required computations to ensure con-
and km is the sampling periods to reach the desired velocity.
straints and potential interruption from a human driver. However,
The time constant km is a tuning parameter of the controller. We
through careful design of the terminal cost in a finite horizon
include acceleration in our reference design as it is necessary to
cost function, the infinite horizon design can be recovered [79].
provide a reference for all states. It is demonstrated in [78] that
We present our platoon cost function and optimization in
it is important to design a reference for all states, such that the
Section V-B.
reference satisfies both desired constraints and system dynamics,
Third, the future states of the human driver vehicle are
while providing a smooth path to the final desired values.
required for the prediction of the platoon state, see (9). To
For the position reference, we take inspiration from [71]
predict the state of the human driver vehicle, we predict the
and [21] to establish the position reference of all the vehicles
potential control actions of the human driver by posing a limited
as the cumulative sum of the desired distances from a virtual
assumption human driver cost function. We present the human
lead vehicle-0. The lead vehicle position reference is
driver control action prediction in Section V-C.

Finally, we summarize our control design with an algorithmic ak (Δt k)2 + v̄Δt k + p̄, k0 ≤ k < k0 + km
1 
implementation in Section V-D. pk = 2

vd Δt k + p̄, k ≥ k0 + km

where p̄ = pk0 + d¯k0 is initialized from the position of vehicle-

(1) (1)
V. CONTROLLER DESIGN
We now design our controller using the models given above to 1. The position reference for each vehicle-i is
⎛ ⎞
achieve the desired platoon velocity and inter-vehicle distances  i i
d¯k = pk − ⎝
(i) (j) (j)
while guaranteeing the constraints and able to reconfigure to pk = pk − dj + hk vk ⎠ ,
a temporary human driver. First, we design a time-varying j=1 j=1
reference for the platoon. Second, we design our finite horizon
where d¯k is defined in (11). By using the desired inter-vehicle
(i)
cost function for the platoon inspired by the infinite horizon
cost function of [21]. We also apply the desired constraints on distanced to form the position referenced for each individual
the cost function to propose our constrained MPC controller vehicle, the headway times are included as part of the state
to centrally control the platoon to the desired platoon velocity reference.
and inter-vehicle distances. Third, we propose a simple finite When a vehicle leaves the platoon under human driver control,
horizon cost function to predict the human driver control actions we desire to drive the platoon forward at the desired velocity but
for use in the platoon controller. Finally, we summarize the within the platoon constraints. We reset the platoon reference
implementation of our platoon control design with human driver based on the human controlled vehicle state. The initial time is
interaction in an algorithm. set as k0 = k, and velocity reference is set to the velocity of

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
KENNEDY et al.: CENTRALIZED MODEL PREDICTIVE CONTROL WITH HUMAN-DRIVER INTERACTION FOR PLATOONING 12671

() and Pk+N is the terminal state cost. To achieve convergence

vehicle-: v̄ = vk0 , and the virtual lead vehicle position as the
independent of platoon length, it is necessary to penalize both
desired distance from vehicle-: p̄ = pk + j=1 d¯k .
() (j)
the relative and the absolute position errors [21].
For convenience we define our desired reference for the
(1) (M ) Using algebraic manipulation and (11), the relative position
platoon at time k as the vector Xk = [pk , . . . , pk ,
 T error (12), can be written as a function of the errors
vk , . . . , vk , ak , . . . , ak ] and over the finite prediction horizon
  
(i) (i) (i−1) (i) (i)
as Xk = [(Xk+1 
)T , . . . , (Xk+N

) T ]T . η k = ξk − ξ k + h k ζk
such that the relative position errors can be incorporated as cross-
B. Cost Function Design
terms of the absolute position errors and velocity errors, with
To design our MPC platoon controller we establish position, the headway times as a weight on the velocity errors. While one
velocity and acceleration error states using our time-varying could think of the headway times as a reference to the problem as
references. We propose a finite horizon cost function of these introduced in the desired inter-vehicle distance (11), it is more
errors and discuss how our cost function can be rearranged to be convenient as a weight on the state deviation. By forcing the
in a quadratic function of the vehicle states. Finally, we apply the headway time to be a state reference, it may lead to a nonlinear
desired inter-vehicle distance, velocity, and acceleration limits control problem.
as state constraints on the cost function. Our final constrained Our cost function can now be efficiently written as a quadratic
cost function is in the form of a quadratic program, which can function
then be solved using standard convex optimization techniques.  T  
The constraints on the states and control are a boundary in the J = X̂k+N |k − Xk+N 
Pk+N X̂k+N |k − Xk+N 

cost function solution space, such that the predicted optimal


N −1  T 
control action is guaranteed to not exceed the desired constraints. T
+ X̃k+j|k Qk+j X̃k+j|k + ΔÛk+j|k RΔ ΔÛk+j|k
There exist several quadratic programming solvers to establish
j=0
the optimal control action within constraints [64] which reduces
(14)
to optimization of a convex function [82].
Consider for each vehicle-i for i ∈ {1, M }, the abso- where X̃k+j|k = X̂k+j|k − Xk+j

, RΔ = rIM , and
lute position, velocity, and acceleration errors as the differ- ⎡ ⎤
(i) q1 T M + q2 I M
ence between the current state and desired reference ξk = q1 T κ 0
⎢ ⎥
(i) (i) (i) (i) (i)
pk − pk , ζk = vk − vk , and ψk = ak − ak . For the
(i) Qκ = ⎣ q1 TκT q 1 Hκ + q 3 I M 0 ⎦ (15)
entire platoon, these errors can be written as Xk − Xk = 0 0 q4 I M
(1) (M ) (1) (M ) (1) (M )
[ξk , . . . , ξk , ζk , . . . , ζk , ψk , . . . , ψk ]. For conve- where 0 is a square matrix of zeros of size M × M , TM is a
nience below, we define η̂k+j|k , ξˆk+j|k , ζ̂k+j|k , and ψ̂k+j|k as the
(i) (i) (i) (i)
symmetric Toeplitz matrix of size M × M with the first row
predicted errors where the subscript indicates the state prediction of the form [2, −1, 0, . . . , 0], Tκ is an M × M matrix with the
at time k + j given the state at time k. (1) (M )
headway times of all vehicles [hκ , . . . , hκ ] on the diagonal
Following [71] and [21], we introduce virtual reference ve- and negative headway times of vehicles-2 to-M on the first upper
hicles on the platoon boundary that perfectly follow the refer- (2) (M )
(0) (M +1) (M +1) (0) (M +1) diagonal [−hκ , . . . , −hκ ], and Hκ is a diagonal M × M
ence pk = pk , pk = pk , vk = vk = vk , and (1) (M )
(0) (M +1) matrix where Hκ = diag[(hκ )2 , . . . , (hκ )2 ]. In the case of
ak = ak = ak . and introduce the relative position error (i)
a common constant time headway across the platoon hk = h,
between vehicle-i and vehicle-(i − 1) for i ∈ {1, M + 1} as
then Tκ reduces to a Toeplitz matrix with h on the diagonal and
(i) (i)
η k = pk − p k
(i−1)
+ d¯k .
(i)
(12) −h on the first upper diagonal, and Hκ reduces to the identity
IM multiplied by h2 . The reduction of (13) to (14) is given in
Inspired by the infinite horizon cost function of [21] we Appendix C.
propose a finite horizon cost function over a prediction horizon The terminal penalty P is the penalty on the final state in the
of N steps with our time-varying references prediction horizon. Choosing P as the solution of the algebraic
 +1 M   Ricatti equation implements the infinite horizon cost on the final
N−1 M   (i) 2  2
q2 ξˆ
(i)
J= q1 η̂ +
k+j|k k+j|k
state such that the final control action is the infinite horizon
j=0 i=1 i=1 optimal control action [79].
 2  2  2  Using algebraic manipulation and (9) the cost function can be
(i) (i) (i)
+q3 ζ̂k+j|k + q4 ψ̂k+j|k + r Δuk+j|k written in the form of a quadratic program

 T   J(Xk , ΔÛk ) = f (Xk , Uk−1 )

+ X̂k+N |k − 
Xk+N Pk+N X̂k+N |k − 
Xk+N  
+ ΔUkT (IN ⊗ αk )T Ψ + ΓT ΩΓ (IN ⊗ αk ) ΔUk
(13)  T
where q1 is the penalty on relative position error, q2 the penalty on + 2 ΦXk + λUk−1 + Γ (IN ⊗ ᾱk ) ΔÛk − Xk
absolute position error, q3 the penalty on velocity error, q4 as the
penalty on the acceleration, r the penalty on the control inputs, × ΩΓ (IN ⊗ αk ) ΔÛk (16)

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
12672 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 10, OCTOBER 2023

where f (Xk , Uk−1 ) is a constant term and Ω = The optimal platoon control action is the change in control
diag{Qk , . . . , Qk+N −1 , Pk+N },Ψ = diag{RΔ , . . . , RΔ } are that minimizes the constrained finite horizon cost function
block diagonal matrices. Thus the optimization problem can be
solved using a standard quadratic program solver. ΔÛk = min J(Xk , ΔÛk ) (18)
Consider the desired constraints on the vehicles’ veloci- ΔÛk
ties and accelerations, and the inter-vehicle distances outlined
(i−1) (i)
in Section IV of inter-vehicle distance dmin ≤ pk − pk ≤ which we then implement the first element ΔÛk|k

to the platoon,
(i)
dmax , velocity vmin ≤ vk ≤ vmax , and acceleration amin ≤ before solving again at the next time step.
(i)
ak ≤ amax which can be written as a matrix inequality of the
platoon state C. Incorporation of Hybridized Human Driver Model
 
  X̂ We now predict the change in control action from a human
k+j|k
Ǧ g ≤0 driver for use in our platoon controller using a second MPC
1
algorithm. We assume that the human driver will behave reason-
where ably by rarely changing their control action, and will at mini-
⎡ ⎤ mum obey performance (acceleration) constraints of the vehicle,
TM 0 0 1M −1 dmin maintain non-negative velocity and obey the road speed limit.
⎢ ⎥
⎢−TM 0 0 −1M −1 dmax ⎥ The same acceleration and velocity assumptions are applied
  ⎢⎢ 0 −IM 0 1M vmin ⎥
⎥
in our centralized platoon controller. Future change in control
Ǧ g = ⎢
⎢ 0
⎥
⎢ IM 0 −1M vmax ⎥ ⎥
actions from the human driver that violate these constraints could
⎢ ⎥ cause our platoon controller to be infeasible.
⎣ 0 0 −IM 1M amin ⎦
Using our assumption that the human driver will rarely change
0 0 IM −1M amax their control action, we consider that the driver’s control actions
will be constant over the finite horizon prediction time of the
where TM is a size (M − 1) × M Toeplitz matrix with −1 on
platoon controller. We propose the following quadratic finite
the diagonal and 1 on the first upper diagonal, and 1M −1 and 1M
horizon cost function of the human driver
are column vectors of ones of size (M − 1) and M , respectively,
for a total of 6M − 2 constraints for each step of the prediction    T  
J¯ xk , ΔÛk
() () () ()
horizon. = x̂k+N |k P̄ x̂k+N |k
These constraints can be extended over the finite prediction

N −1  T  
horizon () ()
  + x̂k+j|k Q̄ x̂k+j|k
  X j=0
k
Ḡ ḡ ≤0  T  
1 () ()
+ Δûk+j|k rΔ Δûk+j|k
where Ḡ = diag[Ǧ, . . . , Ǧ] and ḡ T = [g T , . . . , g T ]. Using the
prediction model (9), the constraints on the states can be written As we make no assumption on desired state of the human driver,
in terms of ΔÛk we choose no penalty on the state such that Q̄ = P̄ = 0, and
ḠΓ(IN ⊗ αk )ΔÛk the penalty on control action as the same in the platoon model,
  where rΔ is the th diagonal element of RΔ .
≤ −Ḡ ΦXk + λUk−1 + Γ(IN ⊗ ᾱk )ΔÛk − ḡ (17) This is the most uninformative cost function possible as it
assumes the driver will make no changes to their current control
such that the state constraints appear as a boundary on the cost value. Our cost function simplifies to
function (16) [64].
   T
The left of (17) only changes in the event a vehicle leaves the
J¯ xk , ΔÛk
() () () ()
= ΔÛk Ψ̄ΔÛk (19)
platoon and αk = αk−1 . For a finite platoon of M vehicles, it is
possible to pre-compute all variations of ḠΓ(IN ⊗ αk ). How-
ever, the right of (17) is dependent on the current platoon state where Ψ̄ = diag{rΔ , . . . , rΔ }. Clearly, this cost function is
()
Xk and the last control action Uk−1 , and must be recomputed minimized when ΔÛk = 0.
each step. Now we consider the minimum constraints that we assume
For a prediction horizon of length N , there are N × (6M − ()
the human driver obeys of velocity vmin ≤ vk ≤ vmax , and
2) constraints across the 3 M states at each time point, such ()
acceleration amin ≤ ak ≤ amax which can be written as the
that Ḡ is of size N (6M − 2) × 3M N . Increasing the length matrix inequality on the vehicle state at time k + j as
of the prediction horizon, N , increases the number of control
values required by the number of vehicles M but the number of  
  x̂()
applied constraints by 6M − 2. For a large platoon this can be Ǧ g k+j|k
≤0
a significant increase to the computational effort. 1

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
KENNEDY et al.: CENTRALIZED MODEL PREDICTIVE CONTROL WITH HUMAN-DRIVER INTERACTION FOR PLATOONING 12673

where
⎡ ⎤ Algorithm 1: Centralized Platoon Control With Human-
0 −1 0 vmin Driver Interaction.
  ⎢0 0 −vmax ⎥ Input:
⎢ 1 ⎥
Ǧ g =⎢ ⎥. Number of vehicles M ,
⎣0 0 −1 amin ⎦
Mechanical actuation lags τi ,
0 0 1 −amax (i)
Headway times hk and distances ri ,
These constraints can be extended over the finite prediction Desired velocity vd ,
horizon Constraints {dmin , dmax , vmin , vmax , amin , amax },
 
  X () Reference sampling periods km ,
Ḡ ḡ k
≤0 Prediction window N , and
1 Cost function weights {q1 , q2 , q3 , q4 , r}.
1: for time steps k = 0, 1, . . . do
where Ḡ = diag[Ǧ, . . . , Ǧ] and ḡ T = [g T , . . . , g T ]. Using alge-
2: if a vehicle has left the platoon then
braic manipulation with the vehicle dynamics (10), the state
3: Identify human driver controlled vehicle-
constraints can be written as a function of the change in control
() 4: Re-set state reference based on vehicle- position
ΔÛk
and velocity (Section V-A)
  ()
() ()
ḠΓ̄ΔÛk ≤ −Ḡ Φ̄xk + λ̄uk−1 − ḡ.
()
(20) 5: Predict change in human driver action ΔÛk by
minimizing (19) subject to (20)
The quadratic cost function (19) with the linear matrix con- 6: end if
straints (20) can be minimized using standard quadratic pro- 7: Compute state reference χk (Section V-A)
gramming solvers to find the minimum control action that meets 8: Compute platoon change in control ΔÛk in (18)
the constraints. We would only expect to predict a change in subject to state constraints (17)
control action when one of the constraints will be violated in 9: Compute latest control action Uk in (7)
the finite horizon. We take the prediction of the constrained but 10: Step platoon dynamics (8)
minimally penalized control for the human controlled vehicle-, 11: end for
()
ΔÛk , and include this in the computation for the centralized
platoon control action.
positions and velocities. However, it is the initial transient phase
D. Controller Implementation where the impact is most prominent.
We now summarize our proposed control design with human At minimum the position error penalties q1 and q2 are required
driver interaction. In Algorithm 1 we outline the implementation to ensure that the platoon converges to the desired positions. As
of our centralized platoon MPC with the human driver predic- motivated in [21], the absolute position error penalty q2 must be
tion. The inputs to the controller are the number of vehicles, present or convergence is a function of the platoon length. As q1
dynamics (mechanical actuation lags), desired headway times, and q2 affect the same position error states, it is suggested to tune
inter-vehicle distances, platoon velocity, constraints, reference these parameters together. Increasing q2 forces the vehicles to the
sampling periods for acceleration, prediction window length, position reference, with less regard to the relative distance, while
and cost function weights. In normal operation, the state refer- increasing the relative position error penalty q1 preferences the
ence is computed, the platoon change in control is optimized, inter-vehicle distance over the absolute position reference.
the latest control action is computed, and used to update the It is possible to set the velocity (q3 ) and acceleration (q4 )
dynamics. At the start of each time-step, a check is performed penalties to zero. Convergence is natural following the position
to identify if a vehicle has left the platoon. If a vehicle has left errors. Increasing the velocity (acceleration) penalty forces the
the platoon, then the reference parameters are updated based on velocities (accelerations) closer to the desired reference, which
the current state of the human driver controlled vehicle. This is can slow the convergence of the positions.
used to produce a prediction of the human drivers action.
B. Numerical Simulation
VI. NUMERICAL STUDY
We consider a numerical simulation of five (M = 5) vehicles,
In this section, we provide guidance on the tuning of the cost with sampling period of Δt = 0.1 [sec/sample]. We consider
function weights from (13) and illustrate a numerical experi- vehicle lengths as li = 2.5 [m] for all i = {1, 5}, and the vehicle
ment. mechanical lags as τ1 = 0.5, τ2 = 0.2, τ3 = 0.3, τ4 = 0.6, and
τ5 = 0.4 [sec]. The desired velocity is set as highway speed
A. Cost Function Weights limit of 100 [km/h] or vd = 27.78 [m/s]. We consider that the
Increasing all of the penalties, q1 , q2 , q3 , and q4 , substantially drivers individually select their desired standstill distances as
can cause the optimization algorithm to become infeasible. In r1 = 6, r2 = 6, r3 = 5, r4 = 8, r5 = 7 [m], and headway times
(1) (2) (3) (4) (5)
the initial transient phase, it is necessary for the vehicles to as hk = 1, hk = 0.4, hk = 0.2, hk = 0.3, and hk = 1.4
deviate from the desired reference to enable convergence of both [sec]. At steady-state the inter-vehicle distances between the

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
12674 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 10, OCTOBER 2023

Fig. 3. Inter-vehicle Distance from simulation with five vehicles. The inter- Fig. 4. Velocities of the five vehicles from the simulation. The vehicles follow
vehicle distance smoothly converges to the desired values. The distances between the velocity reference and are represented by colors (solid blue, dashed red,
each pair of neighboring vehicles are represented by colors (solid blue, dashed dotted purple, dash-dot brown, solid gold).
red, dash-dot yellow, dotted purple).

vehicles will be d¯k = 19.61, d¯k = 13.06, d¯k = 18.83, and

(2) (3) (4)

d¯k = 20.61 [m].

(5)

At time 100 [sec] the human driver of vehicle-3 performs

an emergency brake to 0 [m/s], then at 150 [sec] increases
speed to 11 [m/s] until 250 [sec] when the driver returns to
platoon control. Following the interruption to platoon automa-
tion from the driver of vehicle-3, the drivers decide to increase
their individual inter-vehicle distances for additional safety and
(2) (3) (4)
set their headway times to hk = 1.9, hk = 1.7, hk = 1.8,
(5)
and hk = 2.0 [sec] at time 320 [sec]. The new steady-state
inter-vehicle distances will then be d¯k = 61.28, d¯k = 54.72,
(2) (3)

d¯k = 60.50, and d¯k = 65.06 [m].

(4) (5) Fig. 5. Accelerations of the five vehicles from the simulation. The vehicles
follow the acceleration reference and are represented by colors (solid blue,
We choose the acceleration constraints as amin = −6 [m/s2 ] dashed red, dotted purple, dash-dot brown, solid gold).
and amax = 3 [m/s2 ], based on the performance of an average
passenger vehicle and comfort of passengers. We choose the
velocity constraints as vmin = 0 [m/s] and vmax = 27.8 [m/s], Fig. 4 shows the velocities and Fig. 5 shows the accelerations
based on the road speed limit. Finally, the minimum inter-vehicle of the five vehicles in the platoon. Our proposed controller
distance dmin = 2 [m] and maximum inter-vehicle distance smoothly converges the vehicles to the target velocity reference.
dmax = 70 [m]. We consider a prediction horizon of 1.5 [sec], We observe that the controller initially accelerates the latter ve-
or 15 samples, with time to reach desired velocity of 40 [sec] hicles in the platoon to converge to the target position references.
which equates to km = 400 [samples]. Our simulation results illustrate that our proposed control de-
We choose the penalty on the relative position errors or sign successfully converges the controlled vehicles to the desired
inter-vehicle distances as q1 = 1, absolute position error or error velocity and inter-vehicle distances. The controller smoothly
to the position reference as q2 = 1, velocity errors as q3 = 1, accelerates the vehicles to the desired position reference, min-
acceleration errors as q4 = 1, change in control as RΔ = 2IM , imizing the absolute and relative distance errors before con-
and the terminal cost P as the solution to the algebraic Riccatti verging the velocity to the reference. The use of a constrained
equation. MPC optimization approach ensures that safety margins on the
We simulate our proposed control design using a standard inter-vehicle are maintained, the velocities are within the road
convex optimizer1 to solve the quadratic program (16) with speed limits, and the commanded accelerations are appropriate
constraints (17), and human driver prediction (19) with (20). for the vehicle and comfortable for passengers.
The inter-vehicle distances are shown in Fig. 3, with constraints Our simulation also illustrates that the use of a human driver
as the solid horizontal black lines. Our proposed controller model enables the centralized controller to operate in the pres-
converges the vehicles to the desired inter-vehicle distances by ence of an unknown human driver. Within one sample, our
converging to the desired position reference and ensures the centralized controller reacts, reducing the speed of the remaining
position constraints are maintained when vehicle-3 is controlled vehicles and ensuring that all vehicles in the platoon reach
by a human driver. zero velocity before the inter-vehicle distance constraints are
violated. As the human controlled vehicle speeds up to a slower
1 We use the MATLAB mpcActiveSetSolver from the MPC toolbox. velocity than desired, the platoon then maintains the desired

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
KENNEDY et al.: CENTRALIZED MODEL PREDICTIVE CONTROL WITH HUMAN-DRIVER INTERACTION FOR PLATOONING 12675

Fig. 6. Inter-vehicle Distance from simulation with five vehicles and distur- Fig. 7. Inter-vehicle Distance from the comparison simulation with five ve-
bance to vehicle-1. The inter-vehicle distances start at steady state and diverge hicles. Without inclusion of constraints in the control design, the inter-vehicle
with application of the disturbance. The impact of the disturbance is reduced distance becomes quite large before vehicles-4 and-5 crash at 25.4 [sec]. The
down the platoon indicating possible string stability. The distances between each distances between each pair of neighboring vehicles are represented by colors
pair of neighboring vehicles are represented by colors (solid blue, dashed red, (solid blue, dashed red, dash-dot yellow, dotted purple).
dash-dot yellow, dotted purple).

standstill distances with additional margin of the constant time inter-vehicle distances reduces down the platoon, illustrating a
headway while matching the lower velocity. As the vehicle key property of string stability. Finally, when the disturbance
returns the platoon smoothly accelerates back to the desired is removed the platoon, the inter-vehicle distances return to the
velocity and returns to the full desired inter-vehicle distances. steady-state values.
Remark: A key contribution of our work is we include a human
driver model within our centralized design (9). Our design avoids
catastrophic safety incidents where the actions of the human C. Comparison Simulation
driver, such as emergency braking, could result in an accident We provide a short comparison example by applying the
where a simpler platoon controller is unable to react to the hu- continuous-time LQR controller of [21] to our scenario above.
man driver actions. For example, during the emergency braking This controller is shown to be string stable and furthers the
scenario a constant gain state feedback controller using only the original centralized platoon control designs in [4] and [71].
current state in (4), would command vehicles-4 and-5 to drive However, the control design does not take into account safety and
through vehicle-3, and give a constraint violating instruction actuation constraints, and for certain initial condition scenarios
to vehicle-3 when it re-joined the platoon. An MPC designed the control action may be larger than the possible actuation [81].
for the platoon without the human driver model, see (6), would In our comparison, we keep all scenario parameters outlined
become infeasible as the actions of the human driver in vehicle-3 above the same, except as the design of [21] does not incorporate
would result in the state being constraint violating. individual inter-vehicle distances, we choose the largest initial
Remark: String stability has been well studied for linear sys- steady-state distance considered above of 49.16 [m] as the
tems and the so-called predecessor-follower topology [2], [6], constant inter-vehicle distance. Additionally, we note that the
[42], [44], [69]. However, string stability for nonlinear systems vehicle model, and hence controller, in [21] does not include
and general topologies is very challenging, and remains an open the acceleration, thus we consider the control action or desired
problem [42]. Due to the inclusion of constraints in the optimiza- velocity as the vehicle acceleration and saturate this control
tion problem (17), the closed-loop dynamics resulting from our signal to the acceleration constraints above. Finally, we note that
proposed controller are nonlinear. Additionally, the centralized there is no acceleration error penalty (q4 ) or change in control
design with all vehicle states forms general design topology. (RΔ ), while there is penalty on the control action which we
As our proposed centralized constrained controller is nonlinear choose as r = 1.
with a general topology, showing string stability analytically Figs. 7 and 8 show the inter-vehicle distances and velocities. It
is a non-trivial problem [42]. String stability properties of our is clear that the constraints, in black, are all violated very quickly.
controller can be observed in simulation. We observe that the inter-vehicle distance between vehicles-
We illustrate in an extension to the above simulation example 4 and-5 becomes negative at 25.4 [sec], indicating that these
of our centralized controller for a platoon of five vehicles. Con- vehicles have crashed. Additionally, the velocities are absurd
sider the period just after the platoon reaches steady state, just with vehicle-5 reaching velocities of −7.14 to 55.02 [m/s] (or
after 40 [sec]. We apply a disturbance to vehicle-1 between 60 −25.68 to 198.08 [km/h]).
[sec] and 120 [sec]. Fig. 6 shows the inter-vehicle distance. The In comparison, beyond incorporating human driving, a key
impact on the inter-vehicle distance between vehicle-1 and-2 contribution of our proposed controller is the inclusion of safety
is clear (solid blue line). However, the impact on the distance and actuation constraints. Our control design converges the pla-
between vehicle-4 and-5 (dotted purple line) is very small. We toon of vehicles to the desired inter-vehicle distances and platoon
observe in simulation that the impact of the disturbance on the velocity without violating the state and actuation constraints.

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
12676 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 10, OCTOBER 2023

⎡    ⎤
1 τi Δt − τi 1 − exp − Δ
Δt t
⎢   τi ⎥
⎢ ⎥
to give A(i) = ⎢0 1 τi 1 − exp − Δ t
⎥
⎣  τi ⎦
0 0 exp − Δ τi
t

⎡     ⎤
Δ2t
−τi Δt − τi 1 − exp −Δ t
+
⎢   τi  2 ⎥
⎢ ⎥
and B (i) =⎢ Δt − τi 1 − exp −Δ t
⎥.
⎣  τi ⎦
1 − exp −Δ τi
t

B. Platoon Dynamics
The block diagonal dynamics matrices of (4) are defined as
Fig. 8. Velocities of the five vehicles from the comparison simulation. The
velocities become extremely large and are represented by colors (solid blue,
⎡ (1,3)
⎤ ⎡ (1,1) ⎤
IM Δt IM AM BM
dashed red, dotted purple, dash-dot brown, solid gold). ⎢ (2,3) ⎥ ⎢ (2,1) ⎥
AM = ⎣ 0 IM AM ⎦ and BM = ⎣BM ⎦
(3,3) (3,1)
0 0 AM BM
Additionally, we also include interruption from a human driver where IM the identity matrix of size M × M , 0 is a matrix of
in our CACC design. zeros of appropriate size, and
    
VII. CONCLUSION (1,3) Δt
AM = diag τ1 Δt − τ1 1 − exp − ,...,
τ1
In this paper we propose a hybrid constrained MPC algorithm    
to control a heterogeneous platoon of vehicles to a desired Δt
τM Δt − τM 1 − exp −
platoon velocity and inter-vehicle distance. The finite horizon τM
cost function of our centralized platoon controller is inspired    
(2,3) Δt
from the infinite horizon cost function of [21] with inclusion AM = diag τ1 1 − exp − ,...,
τ1
of headway times individual to each vehicle and able to be   
changed with time. Through the use of constraints in the op- Δt
τM 1 − exp −
timization, we ensure that the control actions result in safe τM
vehicle behavior. In our approach, we propose the use of a cost     
(3,3) Δt Δt
function to predict the control actions of a human driver that AM = diag exp − , . . . , exp −
takes control of their vehicle, by assuming that the human driver τ1 τM
will only obey at minimum, the legal velocity limits and the     
(1,1) −Δt Δ2
physical performance constraints of their vehicle. We illustrate BM = diag −τ1 Δt − τ1 1 − exp + t,
τ1 2
the performance of our control approach in a numerical study.     
The centralized design allows the controller to utilize all possible −Δt Δ2t
. . . , −τM Δt − τM 1 − exp +
vehicle state information. Future work includes implementing τM 2
distributed state estimation such that the centralized approach    
(2,1) −Δt
can be operated decentralized, with consideration of unreliable BM = diag Δt − τ1 1 − exp ,...,
τ1
and delayed communication between vehicles.   
−Δt
Δt − τM 1 − exp
APPENDIX τM
    
A. Single Vehicle Dynamics (3,1) −Δt −Δt
BM = diag 1 − exp , . . . , 1 − exp
τ1 τM
The continuous-time dynamics and control input matrices of
the dynamics (1) are In the case of a homogeneous platoon, τi = τ , then A(i) = A
⎡ ⎤ ⎡ ⎤ and B (i) = B and the platoon dynamics can be conveniently
0 1 0 0 computed AM = A ⊗ IM and BM = B ⊗ IM , where ⊗ is the
⎢ ⎥ ⎢ ⎥
A(i)
c  ⎣ 0 0 1 ⎦ , and B (i)
c  ⎣ 0⎦. Kronecker operator.
0 0 − τi 1 1
τi
The matrices of the prediction model of the platoon dynamics
(6) are
Following [76], a continuous-time system can be discretized ⎡ ⎤ ⎡ ⎤
with sampling interval Δt [s] using AM A0M BM
⎢ . ⎥ ⎢ .. ⎥
 Δt Φ=⎢ . ⎥
⎣ . ⎦, λ = ⎣
⎢
.
⎥,
⎦
(i) (i) (i)
A = exp(Ac Δt ) and B = exp(A(i)
c m)dm Bc
(i)
AN (A N −1
+ · · · + A0
)B
0 M M M M

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
KENNEDY et al.: CENTRALIZED MODEL PREDICTIVE CONTROL WITH HUMAN-DRIVER INTERACTION FOR PLATOONING 12677

⎡ ⎤
BM ··· 0 Recall that the virtual lead and tail vehicles perfectly follow the
⎢ .. .. ⎥ reference such that ξˆκ = ξˆκ
(0) (M +1) (0)
= 0, and ζ̂κ = ζ̂κ
(M +1)
= 0,
and Γ = ⎢
⎣ .
..
. ⎥
. ⎦ then
N −1
(AM + · · · + A0M )BM ··· BM   2  2
and for the human controlled vehicle (10) are = q1 2 ξˆκ(1) + · · · + ξˆκ(M )
⎡ ⎤ ⎡ ⎤
A() (A() )0 B ()
⎢ ⎥ ⎢ ⎥ − 2ξˆκ(2) ξˆκ(1) − · · · − 2ξˆκ(M ) ξˆκ(M −1)
.. ..
Φ̄ = ⎢
⎣ .
⎥ λ̄ = ⎢
⎦ ⎣ .
⎥,
⎦  2  2  2  2
(A() )N ((A() )N −1 + · · · + (A() )0 )B () + h(1) κ ζ̂κ(1) + · · · + h(M κ
)
ζ̂κ(M )
⎡ ⎤
+ 2h(1) ˆ(1) (1) (M ) ˆ(M ) (M )
B () ··· 0 κ ξκ ζ̂κ + · · · + 2hκ ξκ ζ̂κ
⎢ .. .. ⎥ 
and Γ̄ = ⎢ .. ⎥
. ⎦. −2h(2) ˆ(1) (2) (M ) ˆ(M −1) (M )
⎣ κ ξκ ζ̂κ − · · · − 2hκ ξκ
. . ζ̂κ
((A() )N −1 + · · · + (A() )0 )B () · · · B ()
This can be written in matrix notation as
C. Cost Function Expansion  
In the below, we show the expansion of sums in the cost q1 ξˆκT TM ξˆκ + ζ̂κT Hκ ζ̂κ + ξˆT Tκ ζ̂κ + ζ̂κT TκT ξˆκ
function from (13) to the matrix version (14). Consider (13)  T   
 +1 M  
ξˆκ TM Tκ ξˆκ

N −1 M  (i) 2  2 = q1
J= q1 η̂k+j|k + q2 ξˆk+j|k
(i) ζ̂κ TκT Hκ ζ̂κ
j=0 i=1 i=1
 2  2  2  where TM is a symmetric Toeplitz matrix of size M × M with
(i) (i) (i) the first row of the form [2, −1, 0, . . . , 0], and Tκ and Hκ are
+q3 ζ̂k+j|k + q4 ψ̂k+j|k +r Δuk+j|k
M × M matrices where
 T  
⎡ (1) ⎤
+ X̂k+N |k − Xk+N

Pk+N X̂k+N |k − Xk+N

hκ −hκ
(2)
0 ... 0 0
⎢ (2) (3) ⎥
The first sum in (13) is of the relative position errors η̂κ noting ⎢ 0 hκ −hκ ... 0 0 ⎥
⎢ ⎥
that we substitute κ in place of k + j|k. We start by showing this ⎢ .. .. .. .. ⎥
⎢ . . . . ⎥
in terms of the absolute position errors and velocity errors Tκ = ⎢ ⎢ . .. ⎥
⎥
⎢ .. .. ..

M +1  2 
M +1  2 ⎢ . . . ⎥ ⎥
q1 η̂κ(i) = q1 ξˆκ(i) − ξˆκ(i−1) + h(i) (i) ⎢ (M −1) (M ) ⎥
κ κ ζ̂ ⎣ 0 0 0 . . . hκ −hκ ⎦
i=1 i=1 (M )
0 0 0 ... 0 hκ
+1  2  2  2  2 

M
2  2 
= q1 ξˆκ(i) + ξˆκ(i−1) + h(i)
κ ζ̂κ(i) and Hκ = diag h(1) κ , . . . , hκ
(M )
.
i=1

−2ξˆκ(i) ξˆκ(i−1) + 2h(i) ˆ(i) (i)
κ ξκ ζ̂κ − 2hκ ξκ
(i) ˆ(i−1) (i)
ζ̂κ Returning to the full sum, we can see that the other terms can
 also be written in matrix notation
2  2  2  2
⎡  T   
= q1 ξˆ(1) + ξˆ(0) + h(1) ζ̂ (1)  −1
ξˆk+j|k ξˆk+j|k
κ κ κ κ N
⎣ TM Tk+j
J= q1 T
− 2ξˆκ(1) ξˆκ(0) + 2h(1) ˆ(1) (1) (1) ˆ(0) (1)
κ ξκ ζ̂κ − 2hκ ξκ ζ̂κ j=0
ζ̂k+j|k Tk+j Hk+j ζ̂k+j|k
 2  2  2  2
+ ξˆκ(2) + ξˆκ(1) + h(2) ζ̂κ(2) +q2 ξˆk+j|k
T
IM ξˆk+j|k + q3 ζ̂k+j|k
T
IM ζ̂k+j|k
κ
 T 
T
− 2ξˆκ(2) ξˆκ(1) + 2h(2) ˆ(2) (2) (2) ˆ(1) (2) +q4 ψ̂k+j|k Im ψ̂k+j|k + r ΔUk+j|k IM ΔUk+j|k
κ ξκ ζ̂κ − 2hκ ξκ ζ̂κ
 2  2  2  2
+ · · · + ξˆκ(M ) + ξˆκ(M −1) + h(M )
ζ̂κ(M ) + (X̂k+N |k − Xk+N
)T Pk+N (X̂k+N |k − Xk+N

)
κ
⎡⎡ ⎤T ⎡ ⎤
− 2ξˆκ(M ) ξˆκ(M −1) + 2h(M ) ˆ(M ) (M )
ζ̂κ − 2h(M ) ˆ(M −1) (M )
 −1 ξˆ ξˆk+j|k
κ ξκ κ ξκ ζ̂κ N
⎢⎢ k+j|k ⎥
= ⎢⎣ ζ̂k+j|k ⎦ Qk+j ⎢ ⎥
⎣ ζ̂k+j|k ⎦
 2  2  2  2 ⎣
+ ξˆ(M +1) + ξˆ(M ) + h(M +1)
κ κ ζ̂ (M +1)
κ κ
j=0 ψ̂k+j|k ψ̂k+j|k
 T 
− 2ξˆκ(M +1) ξˆκ(M ) + 2h(M
κ
+1) ˆ(M +1) (M +1)
ξκ ζ̂κ + ΔUk+j|k RΔ ΔUk+j|k

+1) ˆ(M ) (M +1)
−2h(Mκ ξκ ζ̂κ + (X̂k+N |k − Xk+N

)T Pk+N (X̂k+N |k − Xk+N

)

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
12678 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 10, OCTOBER 2023

where from (15) of RΔ = rIM and [21] M. R. Jovanovic and B. Bamieh, “On the ill-posedness of certain vehicular
platoon control problems,” IEEE Trans. Autom. Control, vol. 50, no. 9,
⎡ ⎤ pp. 1307–1321, Sep. 2005.
q1 T M + q2 I M q1 T κ 0
⎢ T ⎥ [22] J. Fleming, X. Yan, and R. Lot, “Incorporating driver preferences into
Qκ = ⎣ q1 T κ q 1 Hκ + q 3 I M 0 ⎦ eco-driving assistance systems using optimal control,” IEEE Trans. Intell.
Transp. Syst., vol. 22, no. 5, pp. 2913–2922, May 2021.
0 0 q4 I M [23] P. Fernandes and U. Nunes, “Platooning with IVC-Enabled autonomous
vehicles: Strategies to mitigate communication delays, improve safety and
then with further simplification of state and reference we find traffic flow,” IEEE Trans. Intell. Transp. Syst., vol. 13, no. 1, pp. 91–106,
the cost function in reduced form (14). Mar. 2012.
[24] X. Liu, A. Goldsmith, S. Mahal, and J. Hedrick, “Effects of communication
delay on string stability in vehicle platoons,” in Proc. IEEE Int. Conf. Intell.
Transp. Syst., Oakland, CA, 2001, pp. 625–630.
REFERENCES [25] C. Sommer, O. K. Tonguz, and F. Dressler, “Traffic information systems:
[1] C. Sommer and F. Dressler, Vehicular Networking. Cambridge, U.K.: Efficient message dissemination via adaptive beaconing,” IEEE Commun.
Cambridge Univ. Press, 2014. Mag., vol. 49, no. 5, pp. 173–179, May 2011.
[2] Z. Wang, Y. Bian, S. E. Shladover, G. Wu, S. E. Li, and M. J. Barth, “A [26] ETSI, “Intelligent transport systems (its); Decentralized congestion con-
survey on cooperative longitudinal motion control of multiple connected trol mechanisms for intelligent transport systems operating in the 5GHz
and automated vehicles,” IEEE Intell. Transp. Syst. Mag., vol. 12, no. 1, range; access layer part,” ETSI, Tech. Rep. TS 102 687 V1.1.1, Jul. 2011.
pp. 4–24, Spring 2020. [27] C. Sommer, S. Joerer, M. Segata, O. K. Tonguz, R. Lo Cigno, and
[3] R. Kunze, R. Ramakers, K. Henning, and S. Jeschke, “Organization and F. Dressler, “How shadowing hurts vehicular communications and how
operation of electronically coupled truck platoons on german motorways,” dynamic beaconing can help,” IEEE Trans. Mobile Comput., vol. 14, no. 7,
in Proc. Autom. Commun. Cybern. Sci. Eng., 2010, pp. 427–439. pp. 1411–1421, Jul. 2015.
[4] W. Levine and M. Athans, “On the optimal error regulation of a string of [28] M. Segata, F. Dressler, and R. Lo Cigno, “Jerk beaconing: A dynamic
moving vehicles,” IEEE Trans. Autom. Control, vol. 11, no. 3, pp. 355–361, approach to platooning,” in Proc. IEEE Veh. Netw. Conf., Kyoto, Japan,
Jul. 1966. 2015, pp. 135–142.
[5] A. Alam, B. Besselink, V. Turri, J. Martensson, and K. H. Johansson, [29] M. Segata, F. Dressler, and R. Lo Cigno, “Let’s talk in groups: A distributed
“Heavy-duty vehicle platooning for sustainable freight transportation: A bursting scheme for cluster-based vehicular applications,” Veh. Commun.,
cooperative method to enhance safety and efficiency,” IEEE Control Syst. vol. 8, pp. 2–12, Apr. 2017.
Mag., vol. 35, no. 6, pp. 34–56, Dec. 2015. [30] M. S. Amjad, T. Hardes, M. Schettler, C. Sommer, and F. Dressler, “Using
[6] B. Besselink et al., “Cyberphysical control of road freight transport,” Proc. full duplex relaying to reduce physical layer latency in platooning,” in Proc.
IEEE, vol. 104, no. 5, pp. 1128–1141, May 2016. IEEE Veh. Netw. Conf., Los Angeles, CA, 2019, pp. 236–239.
[7] D. Swaroop and J. K. Hedrick, “String stability of interconnected systems,” [31] M. Segata et al., “Multi-technology cooperative driving: An analysis
IEEE Trans. Autom. Control, vol. 41, no. 3, pp. 349–357, Mar. 1996. based on PLEXE,” IEEE Trans. Mobile Comput., to be published,
[8] C. Wu, Z. Xu, Y. Liu, C. Fu, K. Li, and M. Hu, “Spacing policies for doi: 10.1109/TMC.2022.3154643.
adaptive cruise control: A survey,” IEEE Access, vol. 8, pp. 50149–50162, [32] G. Cecchini, A. Bazzi, B. M. Masini, and A. Zanella, “Performance
2020. comparison between IEEE 802.11p and LTE-V2V in-coverage and out-
[9] C. Lank, M. Haberstroh, and M. Wille, “Interaction of human, machine, of-coverage for cooperative awareness,” in Proc. IEEE Veh. Netw. Conf.,
and environment in automated driving systems,” Transp. Res. Rec.: J. Turin, Italy, 2017, pp. 109–114.
Transp. Res. Board, vol. 2243, no. 1, pp. 138–145, 2011. [33] V. Vukadinovic et al., “3GPP C-V2X and IEEE 802.11p for Vehicle-to-
[10] J.-J. Martinez and C. Canudas-de Wit, “A safe longitudinal control for Vehicle communications in highway platooning scenarios,” Ad Hoc Netw.,
adaptive cruise control and stop-and-go scenarios,” IEEE Trans. Control vol. 74, no. 7, pp. 17–29, May 2018.
Syst. Technol., vol. 15, no. 2, pp. 246–258, Mar. 2007. [34] A. Bazzi, B. M. Masini, A. Zanella, and I. Thibault, “On the performance
[11] M. Gouy, K. Wiedemann, A. Stevens, G. Brunett, and N. Reed, “Driving of IEEE 802.11p and LTE-V2V for the cooperative awareness of connected
next to automated vehicle platoons: How do short time headways influence vehicles,” IEEE Trans. Veh. Technol., vol. 66, no. 11, pp. 10419–10432,
non-platoon drivers–longitudinal control?,” Transp. Res. Part F: Traffic Nov. 2017.
Psychol. Behav., vol. 27, pp. 264–273, 2014. [35] R. Molina-Masegosa and J. Gozalvez, “LTE-V for sidelink 5G V2X
[12] K. C. Dey et al., “A review of communication, driver characteristics, and vehicular communications: A new 5G technology for short-range vehicle-
controls aspects of cooperative adaptive cruise control (CACC),” IEEE to-everything communications,” IEEE Veh. Technol. Mag., vol. 12, no. 4,
Trans. Intell. Transp. Syst., vol. 17, no. 2, pp. 491–509, Feb. 2016. pp. 30–39, Dec. 2017.
[13] F. Dressler, “Cyber physical social systems: Towards deeply integrated [36] A. Gonzalez, N. Franchi, and G. P. Fettweis, “A feasibility study of LTE-
hybridized systems,” in Proc. IEEE Int. Conf. Comput., Netw. Commun., V2X semi-persistent scheduling for string stable CACC,” in Proc. IEEE
Maui, HI, 2018, pp. 420–424. Wireless Commun. Netw. Conf., Marrakesh, Morocco, 2019, pp. 1–7.
[14] M. S. Amjad, M. Schettler, S. Dimce, and F. Dressler, “Inband full-duplex [37] S. Hegde, O. Blume, R. Shrivastava, and H. Bakker, “Enhanced resource
relaying for RADCOM-based cooperative driving,” in Proc. IEEE Veh. scheduling for platooning in 5G V2X systems,” in Proc. IEEE 5G World
Netw. Conf., 2020, pp. 1–8. Forum, 2019, pp. 108–113.
[15] M. Farina and R. Scattolini, “Distributed MPC for large-scale systems,” in [38] C. Campolo, A. Molinaro, G. Araniti, and A. Berthet, “Better platooning
Handbook of Model Predictive Control, S. V. Raković and W. S. Levine, control toward autonomous driving: An LTE device-to-device commu-
Eds. Cham, Switzerland: Springer, 2019, pp. 239–258. nications strategy that meets ultralow latency requirements,” IEEE Veh.
[16] F. Dressler, F. Klingler, M. Segata, and R. Lo Cigno, “Cooperative driv- Technol. Mag., vol. 12, no. 1, pp. 30–38, Mar. 2017.
ing and the tactile internet,” Proc. IEEE, vol. 107, no. 2, pp. 436–446, [39] M. Giordani, A. Zanella, and M. Zorzi, “Millimeter wave communication
Feb. 2019. in vehicular networks: Challenges and opportunities,” in Proc. Int. Conf.
[17] V. Milanés, S. E. Shladover, J. Spring, C. Nowakowski, H. Kawazoe, Modern Circuits Syst. Technol., Thessaloníki, Greece, 2017, pp. 1–6.
and M. Nakamura, “Cooperative adaptive cruise control in real traffic [40] S. Dimce, M. S. Amjad, and F. Dressler, “mmWave on the road: Inves-
situations,” IEEE Trans. Intell. Transp. Syst., vol. 15, no. 1, pp. 296–305, tigating the weather impact on 60GHz V2X communication channels,”
Feb. 2014. in Proc. IEEE/IFIP Conf. Wireless Demand Netw. Syst. Serv., 2021,
[18] M. Segata et al., “Towards communication strategies for platooning: Sim- pp. 1–8.
ulative and experimental evaluation,” IEEE Trans. Veh. Technol., vol. 64, [41] B. Coll-Perales, J. Gozalvez, and M. Gruteser, “Sub-6GHz assisted MAC
no. 12, pp. 5411–5423, Dec. 2015. for millimeter wave vehicular communications,” IEEE Commun. Mag.,
[19] T. A. Badgwell and S. J. Qin, “Model-predictive control in practice,” vol. 57, no. 3, pp. 125–131, Mar. 2019.
in Encyclopedia of System and Control, J. Baillieul and T. Samad, Eds. [42] S. Feng, Y. Zhang, S. E. Li, Z. Cao, H. X. Liu, and L. Li, “String stability
London, U.K.: Springer, 2015, pp. 756–760. for vehicular platoon control: Definitions and analysis methods,” Annu.
[20] S. J. Qin and T. A. Badgwell, “A survey of industrial model predictive Rev. Contr., vol. 47, pp. 81–97, 2019.
control technology,” Control Eng. Pract., vol. 11, no. 7, pp. 733–764, [43] R. Rajamani, Vehicle Dynamics and Control, 2nd ed.. Berlin, Germany:
2003. Springer, 2012.

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
KENNEDY et al.: CENTRALIZED MODEL PREDICTIVE CONTROL WITH HUMAN-DRIVER INTERACTION FOR PLATOONING 12679

[44] S. Knorn, A. Donaire, J. C. Agüero, and R. H. Middleton, “Passivity-based [68] Y. Zhou, M. Wang, and S. Ahn, “Distributed model predictive control
control for multi-vehicle systems subject to string constraints,” Automat- approach for cooperative car-following with guaranteed local and string
ica, vol. 50, no. 12, pp. 3224–3230, 2014. stability,” Transp. Res. Part B: Methodol., vol. 128, pp. 69–86, 2019.
[45] J. Hu, P. Bhowmick, F. Arvin, A. Lanzon, and B. Lennox, “Cooperative [69] E. van Nunen, J. Reinders, E. Semsar-Kazerooni, and N. van de Wouw,
control of heterogeneous connected vehicle platoons: An adaptive leader- “String stable model predictive cooperative adaptive cruise control for het-
following approach,” IEEE Robot. Autom. Lett., vol. 5, no. 2, pp. 977–984, erogeneous platoons,” IEEE Trans. Intell. Veh., vol. 4, no. 2, pp. 186–196,
Apr. 2020. Jun. 2019.
[46] H. Zhou, R. Saigal, F. Dion, and L. Yang, “Vehicle platoon control [70] K.-Y. Liang, J. Martensson, and K. H. Johansson, “Heavy-duty vehicle
in high-latency wireless communications environment: Model predictive platoon formation for fuel efficiency,” IEEE Trans. Intell. Transp. Syst.,
control method,” Transp. Res. Rec.: J. Transp. Res. Board, vol. 2324, no. 1, vol. 17, no. 4, pp. 1051–1061, Apr. 2016.
pp. 81–90, 2012. [71] S. M. Melzer and B. C. Kuo, “Optimal regulation of systems described
[47] M. Wang, “Infrastructure assisted adaptive driving to stabilise heteroge- by a countably infinite number of objects,” Automatica, vol. 7, no. 3,
neous vehicle strings,” Transp. Res. Part C: Emerg. Technol., vol. 91, pp. 359–366, 1971.
pp. 276–295, 2018. [72] R. Curtain, O. Iftime, and H. Zwart, “A comparison between LQR control
[48] S. Santini, A. Salvi, A. S. Valente, A. Pescapè, M. Segata, and R. Lo for a long string of SISO systems and LQR control of the infinite spatially
Cigno, “A consensus-based approach for platooning with inter-vehicular invariant version,” Automatica, vol. 46, no. 10, pp. 1604–1615, 2010.
communications and its validation in realistic scenarios,” IEEE Trans. Veh. [73] M. R. Jovanovic and B. Bamieh, “On the ill-posedness of certain vehicular
Technol., vol. 66, no. 3, pp. 1985–1999, Mar. 2017. platoon control problems,” in Proc. IEEE Conf. Decis. Control, 2004,
[49] G. Giordano, M. Segata, F. Blanchini, and R. Lo Cigno, “The joint pp. 3780–3785.
network/control design of platooning algorithms can enforce guaranteed [74] F. Lin, M. Fardad, and M. R. Jovanovic, “Optimal control of vehicular for-
safety constraints,” Ad Hoc Netw., vol. 94, Nov. 2019, Art. no. 101962. mations with nearest neighbor interactions,” IEEE Trans. Autom. Control,
[50] C. C. Chien and P. Ioannou, “Automatic vehicle-following,” in Proc. Amer. vol. 57, no. 9, pp. 2203–2218, Sep. 2012.
Contr. Conf., Chicago, IL, 1992, pp. 1748–1752. [75] P. Barooah, P. G. Mehta, and J. P. Hespanha, “Mistuning-based control
[51] P. A. Ioannou and C. C. Chien, “Autonomous intelligent cruise control,” design to improve closed-loop stability margin of vehicular platoons,”
IEEE Trans. Veh. Technol., vol. 42, no. 4, pp. 657–672, Nov. 1993. IEEE Trans. Autom. Control, vol. 54, no. 9, pp. 2100–2113, Sep. 2009.
[52] D. Swaroop, J. K. Hedrick, C. C. Chien, and P. Ioannou, “A comparision of [76] K. J. Åström and B. Wittenmark, Computer-Controlled Systems: Theory
spacing and headway control laws for automatically controlled vehicles,” and Design, 3rd ed. Hoboken, NJ, USA: Prentice Hall, Inc., 1997.
Veh. Syst. Dyn., vol. 23, no. 1, pp. 597–625, 1994. [77] L. Xiao and F. Gao, “Practical string stability of platoon of adaptive
[53] D. Yanakiev and I. Kanellakopoulos, “Nonlinear spacing policies for cruise control vehicles,” IEEE Trans. Intell. Transp. Syst., vol. 12, no. 4,
automated heavy-duty vehicles,” IEEE Trans. Veh. Technol., vol. 47, no. 4, pp. 1184–1194, Dec. 2011.
pp. 1365–1377, Nov. 1998. [78] D. E. Quevedo, R. P. Aguilera, M. A. Perez, P. Cortes, and R. Lizana,
[54] D. Yanakiev and I. Kanellakopoulos, “Longitudinal control of heavy-duty “Model predictive control of an AFE rectifier with dynamic references,”
vehicles for automated highway systems,” in Proc. Amer. Contr. Conf., IEEE Trans. Power Electron., vol. 27, no. 7, pp. 3128–3136, Jul. 2012.
Seatle, WA, 1995, pp. 3096–3100. [79] P. O. M. Scokaert and J. B. Rawlings, “Constrained linear quadratic
[55] J. Ploeg, B. Scheepers, E. van Nunen, N. van de Wouw, and H. Nijmei- regulation,” IEEE Trans. Autom. Control, vol. 43, no. 8, pp. 1163–1169,
jer, “Design and experimental evaluation of cooperative adaptive cruise Aug. 1998.
control,” in Proc. IEEE Int. Conf. Intell. Transp. Syst., Washington, D.C., [80] B. D. O. Anderson and J. B. Moore, Optimal Control: Linear Quadratic
2011, pp. 260–265. Methods. Hoboken, NJ, USA: Prentice-Hall, 1990.
[56] S. Tak, S. Kim, and H. Yeo, “A study on the traffic predictive cruise control [81] M. R. Jovanovic, J. M. Fowler, B. Bamieh, and R. DAndrea, “On avoiding
strategy with downstream traffic information,” IEEE Trans. Intell. Transp. saturation in the control of vehicular platoons,” in Proc. Amer. Contr. Conf.,
Syst., vol. 17, no. 7, pp. 1932–1943, Jul. 2016. 2004, pp. 2257–2262.
[57] J. Wang and R. Rajamani, “Should adaptive cruise-control systems be [82] S. Boyd and L. Vandenberghem, Convex Optimization. Cambridge, U.K.:
designed to maintain a constant time gap between vehicles?,” IEEE Trans. Cambridge Univ. Press, 2004.
Veh. Technol., vol. 53, no. 5, pp. 1480–1490, Sep. 2004.
[58] G. F. Silva, A. Donaire, A. McFadyen, and J. J. Ford, “String stable inte-
Justin M. Kennedy (Member, IEEE) received the
gral control design for vehicle platoons with disturbances,” Automatica,
B. Eng (Electrical)/B. Maths, and Ph.D. degrees in
vol. 127, 2021, Art. no. 109542.
estimation and control of marine craft in the presence
[59] F. Gao, S. E. Li, Y. Zheng, and D. Kum, “Robust control of heterogeneous
of environmental disturbances from the Queensland
vehicular platoon with uncertain dynamics and communication delay,”
University of Technology (QUT), Brisbane, QLD,
IET Intell. Transp. Syst., vol. 10, no. 7, pp. 503–513, 2016.
Australia, in 2016 and 2022, respectively. He is cur-
[60] A. Alam, A. Gattami, K. H. Johansson, and C. J. Tomlin, “Guaranteeing
rently a Postdoctoral Researcher with the School
safety for heavy duty vehicle platooning: Safe set computations and
of Electrical Engineering and Robotics, QUT. His
experimental evaluations,” Control Eng. Pract., vol. 24, pp. 33–41, 2014.
research interest include the application of mathe-
[61] V. Turri, B. Besselink, and K. H. Johansson, “Cooperative look-ahead
matical and control system tools to solve network
control for fuel-efficient and safe heavy-duty vehicle platooning,” IEEE
engineering problems. Dr. Kennedy is a member of
Trans. Control Syst. Technol., vol. 25, no. 1, pp. 12–28, Jan. 2017.
the IEEE Control Systems Society (CSS) and Society for Industrial and Applied
[62] A. Alam, J. Mårtensson, and K. H. Johansson, “Experimental evaluation
Mathematics (SIAM).
of decentralized cooperative cruise control for heavy-duty vehicle platoon-
ing,” Control Eng. Pract., vol. 38, pp. 11–25, 2015.
[63] G. Fiengo, D. G. Lui, A. Petrillo, S. Santini, and M. Tufo, “Distributed Julian Heinovski (Student Member, IEEE) is cur-
robust PID control for leader tracking in uncertain connected ground vehi- rently working toward the Ph.D. degree with the
cles with V2V communication delay,” IEEE/ASME Trans. Mechatronics, School of Electrical Engineering and Computer
vol. 24, no. 3, pp. 1153–1165, Jun. 2019. Science, TU Berlin, Germany. He is also a Re-
[64] J. Maciejowski, Predictive Control: With Constraints. Hoboken, NJ, USA: searcher with the Telecommunications Networks
Prentice Hall, 2002. Group (TKN), School of Electrical Engineering and
[65] R. Amrit, J. B. Rawlings, and D. Angeli, “Economic optimization using Computer Science, TU Berlin. He received the B.Sc.
model predictive control with a terminal cost,” Annu. Rev. Control, vol. 35, and M.Sc. degrees from the Department of Computer
no. 2, pp. 178–186, 2011. Science, Paderborn University, Paderborn, Germany,
[66] D. He, T. Qiu, and R. Luo, “Fuel efficiency-oriented platooning control in 2016 and 2018, respectively. Mr. Heinovski is also
of connected nonlinear vehicles: A distributed economic MPC approach,” a Student Member ACM and the Member of IEEE
Asian J. Control, vol. 22, no. 4, pp. 1628–1638, 2020. Intelligent Transportation Systems Society (ITSS) and IEEE Vehicular Tech-
[67] Y. Zheng, S. E. Li, K. Li, F. Borrelli, and J. K. Hedrick, “Distributed nology Society (VTS). He is a Reviewer of manuscripts in the field of Vehicular
model predictive control for heterogeneous vehicle platoons under unidi- Networks and Intelligent Transportation Systems. His research interest include
rectional topologies,” IEEE Trans. Control Syst. Technol., vol. 25, no. 3, cooperative driving and intelligent transportation systems, mainly focusing on
pp. 899–910, May 2017. platooning.

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.
12680 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 72, NO. 10, OCTOBER 2023

Daniel E. Quevedo (Fellow, IEEE) received Inge- Falko Dressler (Fellow, IEEE) received the M.Sc.
niero Civil Electrónico and M.Sc. degrees from Uni- and Ph.D. degrees from the Department of Computer
versidad Técnica Federico Santa María, Valparaíso, Science, University of Erlangen, Erlangen, Germany,
Chile, in 2000, and the Ph.D. degree from the Uni- in 1998 and 2003, respectively. He is currently a
versity of Newcastle, Callaghan NSW, Australia, in Full Professor and Chair of Telecommunication Net-
2005. He is currently a Professor of cyberphysical works with the School of Electrical Engineering and
systems with the School of Electrical Engineering Computer Science, TU Berlin, Berlin, Germany. His
and Robotics, Queensland University of Technology research interests include adaptive wireless network-
(QUT), Brisbane, QLD, Australia. Before joining ing (sub-6GHz, mmWave, visible light, molecular
QUT, he established and led the Chair in automatic communication) and wireless-based sensing with ap-
control with Paderborn University, Paderborn, Ger- plications in ad hoc and sensor networks, the Internet
many. His research interests include networked control systems and cyberphys- of Things, and Cyber-Physical Systems. Dr. Dressler is also an Associate Editor-
ical systems security. He was the Co-recipient of the 2018 IEEE TRANSACTIONS in-Chief of IEEE TRANSACTION ON MOBILE COMPUTING and Elsevier Com-
ON AUTOMATIC CONTROL George S. Axelby Outstanding Paper Award. Prof. puter Communications and the Editor of journals such as IEEE/ACM TRANS-
Quevedo is also an Associate Editor for IEEE CONTROL SYSTEMS. From 2015 ACTION ON NETWORKING, IEEE TRANSACTION ON NETWORK SCIENCE AND
to 2018, he was the Chair of the IEEE Control Systems Society, Technical ENGINEERING, Elsevier Ad Hoc Networks, and Elsevier Nano Communication
Committee on Networks & Communication Systems. Networks. He has been Chairing conferences such as IEEE INFOCOM, ACM
MobiSys, ACM MobiHoc, IEEE Vehicular Networking Conference (VNC),
IEEE GLOBECOM. He authored the textbooks Self-Organization in Sensor and
Actor Networks published by Wiley & Sons and Vehicular Networking published
by Cambridge University Press. He has been an IEEE Distinguished Lecturer
and an ACM Distinguished Speaker. Dr. Dressler is an ACM Distinguished
Member. He is the Member with the German National Academy of Science and
Engineering (acatech). He has been serving on the IEEE COMSOC Conference
Council and the ACM SIGMOBILE Executive Committee.

Authorized licensed use limited to: UNIVERSITY OF BIRMINGHAM. Downloaded on October 23,2023 at 17:20:35 UTC from IEEE Xplore. Restrictions apply.