IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 31, NO.

3, MAY 2023 1395

Fast Tube Model Predictive Control for Driverless

Cars Using Linear Data-Driven Models
Bernardo A. Hernandez Vicente , Paul A. Trodden , Member, IEEE, and Sean R. Anderson

Abstract— Model predictive control (MPC) has been widely several control techniques have been proven successful at some
applied to different aspects of autonomous driving, typically using aspect of autonomous driving [4], [5], [6], [7]. Among them,
nonlinear physically derived models for prediction. However, model predictive control (MPC) is appealing due to its intrinsic
feedback control systems inherently correct for model errors,
and thus in many applications it is sufficient to use a linear time- handling of multivariable systems, constraints, and its inherent
invariant (LTI) model for control design, especially when using robustness [8]. Indeed, MPC has been shown to successfully
robust control methods. This philosophy of approach appears handle varied autonomous driving requirements [9], [10], [11].
to have been neglected in current driverless car research and MPC controllers for driverless cars fall into different classes
is the research gap that we aim to address here. Namely, depending on the goal of the controller, which in turn defines
instead of deriving meticulous nonlinear physical models of
vehicle dynamics and solving a correspondingly complex optimal the type of vehicle model used to make predictions. If the
control problem (OCP), we identify a low-order data-driven objective is to design high-level path planner–followers, then
LTI model and handle its uncertainty via robust linear MPC Lima et al. [9], Liu et al. [12], and Sun et al. [13] rely
methods. We develop a two-step control scheme for driverless on kinematic models which ignore the transient dynamics
cars based on tube MPC (TMPC), which introduces structural and provide a simple framework to design controllers for
robustness, ensuring constraint compliance despite modeling
error in the data-driven prediction model. Furthermore, we use a range of vehicles. Since they act in higher layers, these
fast optimization methods designed to exploit the special structure MPC controllers usually require low-level controllers to drive
of the linear MPC problem. We evaluate the proposed control the physical vehicle actuators. The simplicity of kinematic
scheme using a vehicle model identified from real-world data and models, even if they are nonlinear as in [13] or time-
simulations in IPGCarmaker, where the model of the vehicle varying as in [9], allows for the design of MPC controllers
under control is inherently nonlinear and uses detailed 3-D
physics. Our results show that an LTI model can be effectively with strong theoretical guarantees despite the presence of
used for the task of lane-keeping, that TMPC can prevent lane constraints.
departure and possible collisions due to model uncertainty, and Dynamic models are usually required for the purpose of
that linear models allow for several algorithmic improvements low-level vehicle control, and most authors use nonlinear
that can decrease computation time by an order of magnitude physically derived dynamic models for use in either the non-
compared with naive MPC implementations.
linear MPC (NMPC) or linear time-varying MPC (LTV-MPC)
Index Terms— Autonomous driving, data-driven model predic- schemes. In [14], for example, the nonlinear models of the
tive control (MPC), fast MPC, lane-keeping, linear MPC, robust lateral dynamics are used in NMPC to perform lane change
control.
maneuvers, while in [11] and [15] NMPC is used with a
I. I NTRODUCTION nonlinear four-wheel dynamic model to solve the tracking
problem. A similar architecture is found in [10] and [16] where

F ULLY autonomous driving has the potential to greatly

reduce travel time, emissions, and accidents, which is
why efforts to deploy (semi) autonomous vehicles in the
NMPC is used for kinematic path-planning with obstacle
avoidance. To reduce the complexity of the MPC problem,
yet capture the entire driving envelope, several authors use
past two decades have steeply increased [1], [2]. The design the LTV models, including lateral-only dynamics [17], [18]
of a reliable control system is of paramount importance to and lateral–longitudinal dynamics [19]. Parameter estimation
guarantee safety of the vehicle and its passengers [3], and modules are included in [18] and [19] to track model varia-
tions; however, the changing nature of the MPC’s prediction
Manuscript received 30 September 2021; revised 24 June 2022; accepted
13 November 2022. Date of publication 29 November 2022; date of current model results in that even for a fixed reference, theoretical
version 25 April 2023. This work was supported by the European Union guarantees are entirely forgone.
(EU) H2020 under Grant 731593 (Dreams4Cars). Recommended by Associate Although prevalent, the use of the nonlinear dynamical
Editor A. Vahidi. (Corresponding author: Bernardo A. Hernandez Vicente.)
Bernardo A. Hernandez Vicente is with the Department of Mechanical models—and furthermore physically derived models—poses
Engineering, University of Concepcion, Concepcion 4070386, Chile (e-mail: several challenges in the design and deployment of MPC
[email protected]). controllers for driverless cars. A nonlinear prediction model
Paul A. Trodden and Sean R. Anderson are with the Depart-
ment of Automatic Control and Systems Engineering, The University results in the optimization associated with the MPC controller
of Sheffield, S1 3JD Sheffield, U.K. (e-mail: [email protected]; being a nonlinear (nonconvex) program [20], which may pose
[email protected]). unachievable computational demands for its on-line solution.
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TCST.2022.3224089. Moreover, the accuracy of the prediction model plays a major
Digital Object Identifier 10.1109/TCST.2022.3224089 role on the performance and theoretical guarantees of MPC
1063-6536 © 2022 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: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
1396 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 31, NO. 3, MAY 2023

controllers [21], and although inherent robustness due to (as opposed to MPC) can prevent departure from the lane
feedback may be present, uncertainties in the nonlinear models and potentially avoid collisions; and 3) the use of fast TMPC
may result in large prediction errors if unaccounted for [18]. can decrease computation time by an order of magnitude,
The NMPC and LTV-MPC techniques with explicitly built- which provides a good baseline for future deployment on target
in robustness provide mitigation for the latter and have been hardware.
devised for the purpose of autonomous driving [22], [23]; The rest of the article is organized as follows. The objec-
however, they are tailored to specific model structures and are tive and scope of the controller are discussed in Section II,
more complex than their linear counterparts, both in the design followed by a description of the kinematic and data-driven
and implementation stages. Furthermore, most robust MPC models used. Section III describes our proposed controller in
controllers require some knowledge about the magnitude of detail, including a brief account of TMPC. In Sections IV–VI,
modeling uncertainty. For the physically derived models, this we present the modeling results and discuss in detail the per-
implies generating a measure of confidence for each model formance of our proposed controller, both with respect to the
parameter, which may require large-scale testing (for a single control objectives and its computational capabilities. We final-
vehicle) [22] and extensive a priori knowledge. ize with some directions for future work in Section VII.
While the nonlinear physical models are the current prevail-
ing choice in MPC for driverless cars, linear time-invariant II. C ONTROL A RCHITECTURE AND S COPE
(LTI) identified models have a track record of successful In this article, we aim to solve the lane-keeping problem
use in a wide range of control applications, particularly in for an autonomous road vehicle. More precisely, we seek
the industrial and process control domains. Much of this to design a control system that provides low-level control
success is built on the fact that the LTI models are often actions to keep the vehicle as close to the lane center as
sufficiently accurate for control purposes even if the true possible, while traveling at a desired speed and respecting
dynamics are nonlinear, which itself may be a consequence of comfort and safety constraints normally associated with road
the tacit observation that feedback control is linearizing [24]. driving.
However, the use of the data-driven models in autonomous We propose a two-step controller to achieve this aim. The
driving applications is uncommon [25], [26], particularly in first step generates references for the vehicle in the veloc-
the context of MPC [27]. The physically derived linear models ity space, using a simple kinematic model in a curvilinear
share many drawbacks of their nonlinear counterparts when it coordinate frame. The second step provides tracking of this
comes to assessing their fidelity, particularly since linearization reference, using a linear robust MPC controller fit with an
may skew a priori knowledge. The data-driven LTI models, LTI data-driven model of the vehicle’s dynamics. The latter is
on the other hand, are built to fit the available data, and hence also expressed in the velocity space, meaning that its outputs
their accuracy depends on model design choices rather than are the velocities (linear and angular) of the vehicle’s center of
(the lack of) a priori knowledge [28]. Moreover, we have gravity (CoG). The choice of velocity space follows previous
recently demonstrated that vehicle dynamics can be well- work showcasing its prediction capabilities [29].
described using the data-driven LTI models in comparison to The architecture of our proposed control system is depicted
the nonlinear and linearized physical models [29]. in Fig. 1. We assume we obtain measurements for the vehicle’s
In this article, therefore, we propose a novel two-step MPC velocity, yaw rate, and global position, which are preprocessed
controller for the purpose of lane-keeping, fit with an LTI to obtain the position of the car in a curvilinear frame of
data-driven model for predicting vehicle dynamics. Our pro- reference. We also assume that some information about the
posed controller performs both high-level path-planning using road’s shape is known a priori, which is a common assumption
spatial-based optimal control and low-level vehicle dynamics in the path-following algorithms [9], [10], [13], [15]. In the
control using linear tube MPC (TMPC) [30]. Our architecture remainder of this section, we introduce the models that are
is similar in structure to that in [10], [11], and [16], but with going to be used by both the steps of our controller, and the
several key differences. First, we use the approach in [29] methods by which they are obtained.
to identify an LTI dynamic model of the vehicle in velocity
space for MPC predictions. Using the LTI models for MPC
predictions, we are able to tackle some of the challenges A. Curvilinear Kinematic Model
previously discussed. We take modeling error explicitly into The first step of our controller uses a simple kinematic
account via built-in robustness in the form of TMPC, and model of the vehicle’s CoG described in a curvilinear reference
we develop the TMPC optimization algorithm for low-level frame with its axes oriented alongside the lane’s center line
vehicle control based on the fast MPC principles described (also known as the Frenet reference frame). This allows for a
in [31], which we apply here to TMPC. simple description of the distance to center of the lane, which
To demonstrate the effectiveness of our proposed method, is one of our main regulation objectives and hence will drive
we conduct simulation experiments on the dynamic models the reference generation.
derived from real-world car data and also using simulation in The reference frame is depicted in Fig. 2, where s(t) is the
IPGCarMaker [32]. These experiments show that: 1) the LTI distance traveled along the curve, yd (t) is the perpendicular
data-driven models are sufficiently accurate for the task of distance to the curve, and the pair [V (t), ψ(t)] describes
path-following in normal driving conditions and may provide the vehicle’s CoG velocity in magnitude and orientation. The
an alternative to nonlinear solutions; 2) the use of TMPC variable γ (s) is the path’s curvature (or heading) with respect

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
HERNANDEZ VICENTE et al.: FAST TMPC FOR DRIVERLESS CARS USING LINEAR DATA-DRIVEN MODELS 1397

Fig. 1. Architecture of the two-step lane-keeping controller. The first step generates velocity references, which are then tracked via TMPC in the second
step.

and steering wheel. We also include road slope as an input in

the modeling stage to reduce model uncertainty, but we do not
use it for regulation in the MPC context.
The states of the model (contained in x) do not necessarily
have any physical meaning and are defined via a subspace
identification method to best fit the available data: the number
of states, n, is a design variable and should be kept small to
avoid overfitting. Finally, the process noise μ represents the
uncertainty in our model and allows to capture the effects the
Fig. 2. Curvilinear reference frame used for reference generation. nonlinearities on the system. In a physical modeling context,
this term can be parameterized as done in [17]. Our data-driven
to some fixed reference frame (X, Y ), and we assume it is approach does not allow for such parameterization, and thus
perfectly known. we assess the impact of μ in prediction via simulation.
Define ω(t) as the vehicle’s yaw rate, and ˙ and  as deriv- The model in (2) is continuous time, but for the purpose
atives with respect to t and s, respectively. In the curvilinear of estimating its parameters we use uniformly sampled data
frame of reference, the kinematic model of the vehicle is with sampling frequency Fs , total samples Nd , and sampling
described by times tk with k ∈ [0, Nd ]; the sampling details are described
in Section IV-A. We estimate the state-space matrices in (2) in
V (t) cos(ψ(t) − γ (s(t)))
ṡ(t) = (1a) two steps. First, we use a closed-form subspace identification
1 − yd (t)γ  (s(t)) method [33] to obtain a set of matrices A◦ , B◦ , and C◦ that
ẏd (t) = V (t) sin(ψ(t) − γ (s(t))) (1b) produce a simulated output, say ŷ that best fits the data
ψ̇(t) = ω(t). (1c) in a least-squares sense [33]. The second step consists of
a refinement of the parameters in A◦ , B◦ , and C◦ via the
In this article, as in [10], we use the Bezier curves to obtain
following nonlinear optimization:
an analytical description for γ (·) and its derivative. Note,
however, that given the nature of our data, we test our 
Nd

path-following algorithm against real-world roads, and hence, min (y(tk ) − ŷ(tk , θ )) ϒ(y(tk ) − ŷ(tk , θ )) (3)
θ
we use Bezier curves of different orders as necessary. k=1

where θ is a vector that contains all the parameters in A◦ ,

B. Data-Driven LTI Model B◦ , and C◦ , and ϒ is a weight used to normalize the different
In the following we describe our selection of the data-driven outputs. The optimization (3) is nonlinear in the parameters,
model and the estimation procedure (the data used are depicted despite (2) being a linear model, because no model is given
in Section IV-A). We propose to use a black-box linear model to the noise ν [33].
in state-space form, that is, The performance of the data-driven models is assessed via
a normalized fit metric associated with the cost in (3)
1 ẋ(t) = Ax(t) + Bu(t) + μ(t) (2a)
F = 100(1 − || yi − ŷi ||2/|| yi − ȳi ||2 )
y(t) = C x(t) + ν(t) (2b)
where x ∈ Rn is the state, u ∈ Rm is the input, y ∈ R p is where the subindex i implies elementwise operations, the bold
the output, ν ∈ R p is the measurement noise, μ ∈ Rn is the variables represent the entire time series, and the bar denotes
process noise, and the matrices A, B, and C are of appropriate a mean value. A value of F = 100% indicates a perfect fit.
dimension. The output vector is composed of the variables we We also analyzed the variance accounted for metric (VAF)
are interested in regulating to reference values, which in this  
V = 100 1 − var( yi − ŷi )/var( yi )
case are linear velocity in the plane V (t) and yaw rate ψ̇(t).
As inputs we choose a subset of low-level inputs available for for which a value of V = 100% indicates that the model is
autonomous driving: pedals (gas and brake), engine torque, able to explain the entire variance of the measurements. This,

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
1398 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 31, NO. 3, MAY 2023

however, is not expected since we do not include a model for constant. Formally, given i ∈ [1, Nc ] and N̄ = (Ng /Nc ), then
measurement noise and hence cannot expect to explain it. (V (k|t), ω(k|t)) = (V̄i , ω̄i ) for all k ∈ [t +(1+(i −1) N̄)Ts , t +
Remark 1: TMPC for linear models is a mature technique i N̄ Ts ].
that enjoys an array of theoretical guarantees provided the The cost function in (4) is defined as
model meets certain conditions. In particular, we require
the pair (A, B) to be stabilizable and the noise/disturbances Jg (V (t), ω(t), V (t), ω(t), s(t), yd (t), ψ(t))
t+Ng Ts 
affecting the model to be bounded. The closed-form subspace   2 
identification method is implemented with a constraint that = Q y yd2 (k|t) + Q ẏ ẏd2 (k|t) + Q ṡ ṡ(k|t) − Vg
forces A to be stable, as we do not expect instability given k=t

our input–output pairs, but we can only guarantee boundedness (5)

of the uncertainty in a probabilistic sense, which we discuss where Q y , Q ẏ , and Q ṡ are the weights. In the cost
in Section IV-C. function (5), the first term penalizes lateral deviations from
the path, the second term promotes the vehicle’s alignment
C. Nonlinear Physics-Based Model with the path, and the third term penalizes deviations from a
To properly frame our contribution, we will also discuss target velocity Vg . Finally, the optimized variables are defined
the performance of our controller with respect to NMPC, by V  (t) = [V  (t +Ts |t), V  (t +2Ts |t), . . . , V  (t + Ng Ts |t)] .
fit with a nonlinear physical model for prediction (as is In what follows, we refer to (4) and (5) as a receding horizon
prevalent in the literature [15], [22], [34]). To represent the nonlinear optimal control problem (RHNOCP).
nonlinear dynamics of the vehicle, we implement a coupled Given a fixed horizon Ng , dividing it into fewer blocks
(longitudinal–lateral) four-wheeled model as described in [29] reduces the complexity of the RHNOCP; however, it may
(referred to as NLPM henceforth). The key features of NLPM result in poor performance and lose of maneuverability. Con-
are one-way coupling from longitudinal to lateral dynamics sider the extreme case where Nc = 1, if large changes in
and the use of the magic formula for tire force estimation. The the path’s curvature are present over the prediction horizon,
unknown parameters of NLPM are fit following the procedure a single value of yaw rate applied across the entire horizon
described in Section II-B. could result in inaccurate tracking of the path. A small
number of blocks should be paired with short prediction
horizons; however, this could result in sharp corners being first
III. TMPC FOR ROBUST PATH -F OLLOWING
encountered in the planning horizon only when the vehicle
A. Reference Generation is in close proximity to them, not allowing enough time
The first step of our controller defines the references in the for the vehicle to slow down appropriately. To tackle these
velocity space. To do so, we use the kinematic model in (1) to issues and allow for reduced complexity of the optimization,
predict the vehicle’s movement and optimize its velocity and we propose to modulate the target velocity Vg with respect
yaw rate to minimize its deviation from the path and adhere to the change in curvature of the oncoming road further from
to the required velocity specifications. At this stage, V (t) and the prediction capabilities of the optimization horizon. The
ψ̇(t) act as the inputs for the kinematic model, so to optimize modulated velocity is defined by
them we propose the following MPC-like discrete optimization 
problem to be solved in a receding horizon fashion at each e−α(s)/ᾱ Vg , e−α(s)/ᾱ Vg ≥ v g
V f (s) =
sampling time t: vg , otherwise
P Nc (V (t), ω(t), s(t), yd (t), ψ(t)) : where α(s) is the total change in curvature throughout a fixed
min JNc (V (t), ω(t), V (t), ω(t), s(t), yd (t), ψ(t)) (4) and arbitrary length of road ahead of the current position of the
V (t),ω(t) vehicle, ᾱ is an arbitrary amount of allowable curvature change
where V (t) and ω(t), respectively, are the predicted velocity within this length of road, and v g is an arbitrarily defined
and yaw rate profiles made at sampling instant t, that is, minimum required velocity.
V (t) = [V (t + Ts |t), V (t + 2Ts |t), . . . , V (t + Ng Ts |t)] , Ng is The RHNOCP is similar to the approach described in [10]
the prediction horizon, Ts the sampling time, and the notation for reference generation, but with some key differences. The
(k|t) indicates prediction k made at time t. Optimization (4) is approach in [10] implements an explicit method to solve
subject to constraints that include a forward Euler discretiza- the reference generation problem, placing the emphasis on
tion of the kinematic equations (1), initial conditions for the the inclusion of collision avoidance (non-convex) constraints.
states s(t), yd (t), and ψ(t), and initial condition for the inputs Nevertheless, the explicit solution is inherently suboptimal,
V (t) and ω(t). The latter explains why V (t|t) is missing from and the subsequent reference tracking is implemented via
V (t), since it is defined by measurements and not available proportional controllers over a bicycle vehicle model. We,
for optimization. on the other hand, do not consider collision avoidance and
The kinematic model is nonlinear, and hence (4) is a use the move-blocking MPC paradigm, which allows longer
nonlinear program (NLP). To reduce computational complex- prediction horizons and the use of general-purpose solvers.
ity, we use the concept of move-blocking MPC [35]. The The latter, in turn, permits the following.
prediction horizon is evenly divided in Nc blocks during 1) To introduce input feedback in the reference gener-
which the optimization variables V (k|t) and ω(k|t) remain ation. Indeed, although linear velocity and yaw rate

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
HERNANDEZ VICENTE et al.: FAST TMPC FOR DRIVERLESS CARS USING LINEAR DATA-DRIVEN MODELS 1399

are considered inputs of the kinematic model, they dynamics in a low-order model, for which the outputs are the
are indeed the dynamic response of the vehicle, and states (see Section IV). By choosing such model, we remove
hence, their value at the current sampling time is not the need for state estimation and the additional machinery in
manipulable. We acknowledge and handle this via the TMPC to deal with estimation error.
inclusion of initial condition constraints in (4) for V (t) Definition 1 (Positive Invariant (PI) Set): A set T ⊂ Rn is
and ω(t), thereby enhancing the kinematic model with a PI set for the dynamics x(t + 1) = Ā K x(t) if Ā K T ⊆ T.
some degree of inertia. Definition 2 (Robust PI (RPI) Set): A set S ⊂ Rn is an RPI
2) To initialize the vehicle at arbitrary conditions (position set for the dynamics x(t +1) = Ā K x(t)+w(t) with w(t) ∈ W
and velocity), and with any target velocity, unlike [10] if Ā K S ⊕ W ⊆ S.
whose bespoke optimization algorithm requires initial- Consider a discrete time version of (2)
ization on the path, at rest, and with a null target velocity.
x(t + 1) = Ad x(t) + Bd u(t) + w(t) (6)
B. Solution of the RHNOCP where Ad and Bd depend on A, B, and the discretization
To solve RHNOCP, we use the NLP solver known as time Ts , and w(t) is composed of the modeling uncertainties
IPOPT [36], via one of its MATLAB interfaces [37]. IPOPT and measurement noise. Furthermore, suppose that states and
is a general-purpose solver that implements an interior point inputs are subject to constraints in the form of compact and
line search method to find a local solution of the NLP. convex sets X ⊂ Rn and U ⊂ Rm , respectively, and the
In the context of RHNOCP (4), i.e., without constraints, the disturbances are bounded inside a compact and convex set
implementation of IPOPT requires the user to provide the W ⊂ Rn . Consider now an undisturbed representation of (6)
value of the cost (5), its gradient, and Hessian, at each iteration z(t + 1) = Ad z(t) + Bd v(t)
of the solver. We estimate the gradient of (5) numerically,
while the Hessian is estimated via a quasi-Newton method and a certain steady-state pair (z ss , v ss ) that is to be tracked.
native to IPOPT. IPOPT uses different termination criteria to The resulting error system is
decide whether the current iteration corresponds to a local ẑ(t + 1) = Ad ẑ(t) + Bd v̂(t) (7)
minima, including iteration number and tolerances on various
indicators such as cost function change. The specific values where ẑ(t) = z(t) − z ss and v̂(t) = v(t) − v ss . Note
used are presented in Section V. that the model matrices (Ad , Bd ) are uncertain, and thus,
Remark 2: The RHNOCP has no hard constraints, nor ter- the computation of v ss —for given a target z ss —also carries
minal ingredients, which are common in MPC. Thus, there are uncertainty. The control law used to regulate the disturbed
no guarantees on the behavior of the optimal cost function, i.e., plant is
whether the optimized trajectory [ψ(t), yd (t), s(t)] remains u(t) = κ̂(x(t), ẑ(t)) = κ(ẑ(t)) + v ss + K T (x(t) − z(t)) (8)
close to the lane’s center. Formal guarantees are currently
under investigation; however, we do observe an acceptable where K T is stabilizing for (Ad , Bd ), and κ(z(t)) is the
tracking behavior in simulation. receding horizon control law that stems from a nominal
MPC controller designed to regulate the error model in (7),
C. Reference Tracking via TMPC subject to tightened and translated versions of the constraints
represented by the sets Z ⊂ Rn and V ⊂ Rm .
In the second step of our controller, we track the references The OCP to be solved at each sampling time is defined as
defined by RHNOCP. To do so, we implement the structurally
robust MPC controller known as TMPC, which is able to P N (ẑ(t)) : min JN (ẑ, v̂) (9)
v̂
take into account external perturbations that may affect the
car and also the modeling error inherent to the data-driven subject to (for k = t, . . . , t + N − 1)
model used to make predictions. For completeness, we now ẑ(k + 1|t) = Ad ẑ(k|t) + Bd v̂(k|t) (10a)
recall some standard definitions and present a brief description
(ẑ(k|t), v̂(k|t)) ∈ (Z  {z ss } × V  {v ss }) (10b)
of the TMPC optimal control problem (OCP) applied in our
framework, but the reader is referred to [38, Ch. 3] for a ẑ(N|t) ∈ Z f (z ss , v ss ) (10c)
detailed description of the technique. In what follows, “⊕” where the cost function JN (ẑ, v̂) is defined as the usual
denotes the Minkowski sum of sets and “” the Pontryagin quadratic cost with terminal penalty
set subtraction [39].
Note that we describe the TMPC technique in a state- −1
t+N
JN (ẑ, v̂) = l f (ẑ(t + N|t)) + l(ẑ(k|t), v̂(k|t))
feedback setting, although the prediction models introduced
k=t
in Section II-B are in output-feedback form. We do this to
simplify the exposition by reducing notation; nevertheless, the with
extension to output feedback is easily achieved by including a l(ẑ(k|t), v̂(k|t)) = ẑ(k|t)T Q ẑ(k|t) + v̂(k|t)T R v̂(k|t) (11a)
suitable state estimation technique (such as the Kalman filter),
l f (ẑ(t + N|t)) = ẑ(t + N|t) T P ẑ(t + N|t). (11b)
and following the methodology described in [40]. Furthermore,
when processing the data through our modeling approach, In (11), Q, R, and P are, respectively, the state, input, and
we find a good tradeoff between simplicity and match to terminal penalties. The controller’s prediction horizon is N,

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
1400 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 31, NO. 3, MAY 2023

and the optimization variable v̂ represents the sequence of Newton process, for which it is possible to leverage warm-
control actions throughout the prediction horizon, i.e., v̂ = starts, where the control solution from the previous time step
{v̂(t|t), v̂(t + 1|t), . . . , v̂(t + N − 1|t)}. The solution to (9) is used to initialize the current one. The use of warm-starts
and (10) is a sequence of optimal inputs v ∗ (ẑ(t)), the set can reduce the number of Newton iterations per MPC step by
Z N (z ss , v ss ) is the set of all states ẑ such that such a solution an order of magnitude.
exists for a given pair (z ss , v ss ), and the implicit nominal The final part of the process is to use early termination
control law is defined as the first control action of the optimal via a fixed iteration limit for the Newton method. Although
sequence κ(ẑ(t)) = v̂ ∗ (t|t). the optimization is then not necessarily solved to optimality,
this does not appear to degrade the control solution. The
explanation for this is likely due to the fact that MPC is largely
D. Fast TMPC
a planning exercise, where future control inputs are discarded.
The optimization problem of standard MPC implementa- Hence, solving the optimization to full accuracy places an
tions is typically formulated as a quadratic program (QP), emphasis on optimizing future control inputs that are never
which can be slow to solve using general-purpose methods. used.
A collection of methods for fast MPC is demonstrated in [31], When all these steps for fast TMPC are combined, the
including: 1) exploiting the structure of the MPC problem solution of the TMPC problem can be obtained orders of
to reduce the computational complexity from cubic, O(N 3 ), magnitude faster than a naive implementation (with the speed-
to linear, O(N), in the planning horizon N; 2) warm-starting up largely depending on the size of the horizon N). These steps
the optimization, where the control solution from the previous are critical to obtaining a computationally efficient controller,
time step is used to initialize the current time step; and 3) enabling the practical implementation of TMPC in driverless
early termination of the optimization. cars where the control loops operate at relatively high sample
One of the advantages of linear TMPC is that its associated rates. Note also that the use of the LTI data-driven models of
optimization problem (9)and (10) is comparable to that of vehicle dynamics, as proposed here, enables us to leverage the
standard MPC, and hence, the methods in [31] can be exploited efficient solution of the approximate primal barrier method.
in our setup. The approach is based on an infeasible start
primal barrier method, where the OCP defined in (9) and (10)
is formulated as a QP with a barrier to replace the inequality E. Stability of the Fast TMPC
constraints, leading to the approximate problem
The following result summarizes the main properties of the
min z  H z + λφ(z) TMPC controller described above.
z
s.t. z=b Assumption 1: The steady-state pair (z ss , v ss ) is such that
Z  {z ss }, V  {v ss }, and Z f (z ss , v ss ) have non-empty interiors,
where and Zn (z ss , v ss ) = ∅. The set S ⊂ Rn is an RPI set for the
closed-loop Ad + Bd K T in the presence of disturbances W.
z = (v̂(k), ẑ(k + 1), . . . , v̂(k + N − 1), ẑ(k + N))
The tightened constraint sets are such that Z ⊆ X  S and
H = blkdiag{R, Q, R, Q, . . . , R, P} V ⊆ U  K T S. The set Z f is a PI set for Ad,K = Ad + Bd K f
⎛ ⎞
−B I 0 0 ··· 0 0 0 such that Z f ⊆ Z  {z ss } and K f Z f ⊆ V  {v ss }. The LQR
⎜ 0 −A −B · · · 0⎟ cost weights fulfill Q ≥ 0, R > 0, and A d,K P A d,K + (Q +
⎜ I 0 0 ⎟
⎜ 0 0 0 −A · · · 0 0 0⎟ K f RK f ) ≤ P.
⎜ ⎟
=⎜ . . . . . . .. .. ⎟
⎜ .. .. .. .. .. .. . .⎟ Theorem 1 (Stability1 ): Suppose Assumption 1 holds. If
⎜ ⎟
⎝ 0 0 0 0 ··· I 0 0⎠ (a) the steady-state pair (z ss , v ss ) is constant, and (b) the
0 0 0 0 · · · −A −B I nominal system is initialized such that x(0) ∈ {z(0)} ⊕
b = (Aẑ(k), 0, . . . , 0) S ⊂ Z N ⊕ S, then the OCP (9) and (10) is feasible at all
times, the true state and input constraints are met at all times,
and where λ is a barrier parameter, and φ(z) is a log barrier and the set S × {z ss } is exponentially stable for (6) when in
function. This barrier method approximates the original opti- closed loop with (8).
mization problem and converges to the optimal solution as the In our architecture, the steady state to be tracked is set to
parameter λ approaches zero. the velocity and yaw rate references generated by RHNOCP,
Wang and Boyd [31] propose a solution to the primal barrier i.e., z ss = [V  (t + 1|Ts ), ω (t + 1|Ts )] , and the corresponding
method defined above using an infeasible start Newton method input v ss is computed via the prediction model. The RHNOCP
and demonstrate how to exploit the structure in H and in a updates the velocities to be tracked at each time instant to
fast Newton update that has linear computational complexity account for current vehicle attitude, which is necessary due
in the horizon N (precise details are omitted here for brevity). to modeling error and the different modeling paradigms used
A further step enables an even greater computational saving: in each step of the controller. This results in the need to
a typical primal barrier method is solved for a decreasing continuously update the terminal constraint set, which could
sequence of λ values; however, they show it is possible to use be dealt with as shown in [42] and [43]; however, it also breaks
just a single value of λ without loss of performance (provided a key assumption necessary for Theorem 1 to hold, namely,
some offline tuning of λ). This turns each MPC step into a the invariance of the steady-state pair to be tracked.

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
HERNANDEZ VICENTE et al.: FAST TMPC FOR DRIVERLESS CARS USING LINEAR DATA-DRIVEN MODELS 1401

There exist some MPC approaches that enjoy stability analysis in MPC assumes that the global minimum is achieved
guarantees in the context of driverless cars, like [17], where each time the problem is solved. Added to this is the fact that
the LTV models for obstacle avoidance are studied. However, the true dynamics are unknown and the identified dynamics
it is not clear whether this result remains valid in view of are merely an approximation, with inevitable modeling error.
a changing reference, and feasibility of the optimization is While several results are available establishing robust stabil-
only assumed. Another interesting approach is [13], where ity of the closed loop in response to both the issues (e.g., [48],
an input-to-state stability result is obtained for a unicycle [49], [50], we do not attempt to adapt them to our setting
robot, yet in the context of kinematic models only. MPC for because of their reliance on assumptions on the true dynamics
tracking changing references has also been studied without a and design conditions on the costs and constraints. Instead,
specific application attached. In [44], for example, the tracking we note that the two-step architecture of our controller is
of periodically changing references is studied, and a result such that the reference used in the TMPC is computed via its
of asymptotic stability is provided. A more general issue own MPC-like problem, and hence, it enjoys some preview
is studied in [20] and [45], where references that change of the road ahead which promotes the smoothness of the
via the system dynamics (or more specifically, the prediction reference profile. In other words, no large changes in the
model) are analyzed. The proposed controllers enjoy stability reference values are observed over one time step, which is in
results that depend on the length of the prediction horizon, line with the results in [47]. Furthermore, the controller shows
removing the need for online (or offline) computation of a good tracking performance in practice, particularly for long
terminal constraint sets. horizons, which is in line with the results shown in [20] and
The RHNOCP could indeed be modified to guarantee that [45]. Therefore, given that the objective of this article is not to
its resulting references are reachable via the prediction model; develop a new tracking MPC technique alongside a manifold
however, the techniques in [20] and [45] also require that the of theoretical guarantees, we forgo from further modifications
reference is entirely determined at initialization (at least for of the TMPC problem.
a duration equal to the prediction horizon). This assumption
is unachievable in our application, since the references are
F. Unconstrained Reference Generation
updated at each time instant by RHNOCP to take into account
the current attitude of the vehicle, which does not match A vehicle’s attitude is usually bound to lie within a certain
predictions due to inherent modeling error. driving envelope defined by regulations and safety require-
The problem of tracking a reference that is randomly ments (particularly during normal driving conditions). This
changing within a compact set is studied in [46]. The proposed envelope is represented by the velocity and yaw rate constraint
controller is shown to drive the state toward a neighborhood sets we enforce in the reference tracking step; however, the
of a set of possible references in finite time. This is akin RHNOCP is an unconstrained optimization problem, hence
to our own setup, since from the perspective of TMPC the the generated references could lie outside of the driving
reference is indeed randomly changing. Nevertheless, a fixed envelope.
(and possibly tight) bound on the reference would constrain the IPOPT allows for the inclusion of constraints in the opti-
pool of applicability to maneuvers with low levels of curvature. mization; however, we have decided to keep the RHONCP
An analogous issue is studied in [47], where the stabilizing constraint-free to avoid creating an undesirable interdepen-
ingredients of a standard MPC controller are modified to dency between the reference generator and the vehicle’s
account for a disturbance whose prediction is available, but dynamic model. Indeed, a constrained RHNOCP would guar-
may be different at each time instant. The modifications antee references within the constraint sets of the TMPC; how-
proposed in [47] result in finite time convergence of the closed ever, the latter do not depend solely on the driving envelope but
loop to a sublevel set of the MPC’s value function, whose also on the TMPC controller parameters (such as the tube gain
size depends on a measure of the allowable change in the used for tightening) and the identified dynamic model. By not
disturbance prediction. In our problem, the error system could including constraints in RHNOCP, we favor a modular design
be recast as and allow for further development of the reference track-
ing step without requiring modifications upstream. Moreover,
ẑ(t + 1) = Ad ẑ(t) + Bd v̂(t) + ŵ(t) we identify RHNOCP as the most critical point in computation
where ŵ(t) represents the change in references from one time (see Section V-C), and therefore, omitting constraints also
instant to the next, that is, ŵ(t) = Ad (z ss (t − 1) − z ss (t)) + alleviates the processing load associated with RHNOCP.
Bd (v ss (t − 1) − v ss (t)), allowing for such an implementation To guarantee that the references to be tracked are within the
to be used. Nevertheless, the authors acknowledge the limited constraint sets of the TMPC, we implement a projection step
practical use of the approach due to the numerous parameters between Steps 1 and 2. Our simulations show that the scheme
that need to be computationally estimated. works well with an unconstrained RHNOCP, with the velocity
Furthermore, we attempt to solve the optimization associ- references violating the TMPC bounds by a maximum of 2%.
ated with the TMPC via the fast MPC algorithm described in
Section III-D, which has early termination as a key feature. IV. DATA -D RIVEN M ODELING R ESULTS
In practice, this results in suboptimality of the MPC solution We first present the data-driven modeling results, since the
and raises the well-known issue of whether the controller characteristics of the prediction model are key in designing
provides stability in view of this, since conventional stability and tuning MPC controllers.

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
1402 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 31, NO. 3, MAY 2023

Fig. 3. Experimental data for the Lancia Delta. (a) Driving route, the gray line indicates the split of the route between training data (above) and validation
data (below). (b) Longitudinal velocity V . (c) Yaw rate ω. (d) Gas pedal position. (e) Brake pressure P. (f) Road slope angle φ. (g) Steering angle δ.

A. Experimental Data TABLE I

The experimental data were gathered through 108 km of P ERFORMANCE M ETRIC S UMMARY FOR THE L ANCIA D ELTA M ODELING
driving on public roads in the Piemonte region in Italy,
on a Lancia Delta car. The route was designed to include a
typical selection of extra-urban and urban roads, motorways,
roundabouts, and intersections, but the driving itself was
unscripted since it had to accommodate for real-time traffic
conditions. The driving route is depicted in Fig. 3(a) and it
was circumnavigated twice. The entire set of data was divided
into training and validation portions as shown by the gray
line in Fig. 3(a), with 2501 s of data for training and 2383 s
Table I shows the performance metrics computed over the
for validation. The data were sampled at 100 Hz; however,
validation portion of the data for several modeling configu-
the power of relevant signals was found to be concentrated
rations, including NLPM. We first obtained models with a
at very low frequencies. In view of this, we resampled it at
(V, ω) ← (T, P, δ, φ) input–output configuration and state
20 Hz, to decrease the computational load of the estimation
orders ranging from 2 to 4. The performance of the estimated
and control algorithms. Note also that we use Ts = 0.05 s
state-space models (see Table I) does not change significantly
for discretization purposes of all our models. In what fol-
throughout the different state orders, albeit the number of
lows, we refer to post-gearbox torque, brake master cylinder
parameters grows from 16 to 40. In view of the above, and to
pressure, steering angle, and road inclination by T , P, δ,
avoid overfitting, we decide to set the state order to n = 2. This
and φ, respectively. Fig. 3(b) and (c) shows the measurements
also has the advantage of avoiding state estimation, allowing
used to drive the identification procedure for a part of the
us to use the state-feedback TMPC described in Section III
training portion, while Fig. 3(d)–(g) shows the corresponding
directly. The performance of NLPM is also shown in Table I.
inputs.2
In normal driving, however, it is unexpected to experience
2 For commercial confidentiality, we report gas pedal and not post-gearbox
acceleration and braking simultaneously, so we proceed to
torque. We use road inclination to improve model performance but is later merge these two into a single input. We define the new merged
dropped in the controller analysis due to it being an uncontrolled input. input as T P = r T − P, where r is the scaling factor that

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
HERNANDEZ VICENTE et al.: FAST TMPC FOR DRIVERLESS CARS USING LINEAR DATA-DRIVEN MODELS 1403

TABLE II TABLE III

P ERFORMANCE M ETRIC S UMMARY FOR THE M EGANE M ODELING O NE -S TEP A HEAD E RROR B OUNDS C OMPUTED VIA S IMULATION FOR THE
M ODELS OF THE L ANCIA D ELTA AND M EGANE

was tuned with respect to model fit metrics. Furthermore,

from the cross terms in the state and input matrices of
error. Indeed, 60% fit in yaw rate may seem unacceptably
(V, ω) ← (T, P, φ, δ), it appears that the coupling between
low to guarantee safety. Nevertheless, the TMPC step of our
longitudinal and lateral dynamics is weak, at least within
controller is able to deal with some modeling error as long as
the available data. In view of the above is that we estimate
it can be bounded inside the perturbation set W. To quantify
uncoupled models, whose performance metrics differ from
the modeling error as an uncertainty like the one described
those of the fully coupled model by less than 2% (see Table I).
in (6), we compute the one-step ahead prediction error of our
Finally, note that NLPM has poorer performance metrics
model for each data point in the validation dataset and fit a
than any of the linear data-driven models. A detailed discus-
normal distribution to the obtained values. We then define the
sion pertaining these results can be found in [29]; however,
bounds on the modeling error as two standard deviations of
the main reason for this is the numerical complexity involved
the fit normal distribution, which account for a 95% of the
in an accurate estimation of the NLPM parameters.
error values. With this approach, we obtain the error bounds
depicted in Table III, where LB and UB stand for lower and
B. CarMaker Data upper bounds, respectively.
The IPGCarMaker software provides high-fidelity simula-
tions of vehicle dynamics through a 3-D physics environ-
V. PATH -F OLLOWING A LGORITHM P ERFORMANCE
ment and allows flexible coupling with software such as
Simulink for measurement acquisition and controller testing. A. Nonlinear Baseline
IPGCarMaker allows for a meticulous vehicle design process To acknowledge the prevalence of nonlinear solutions in
that would allow us to replicate the Lancia Delta car in the autonomous driving, we begin with a performance comparison
software. However, the main reason for using a data-driven between linear data-driven and nonlinear physics-based MPC.
model of the vehicle dynamics is that we do not know (or have The latter is an NMPC fit with the NLPM as a prediction
no means to obtain) several of the many physical parameters model. Note, however, that the NLPM has a poorer fit to data
that define the vehicle. than the linear data-driven models, and thus, to avoid reaching
To test our algorithm with a high-fidelity simulator, we use biased conclusions, we will temporarily consider NLPM as the
a vehicle already loaded in the IPGCarMaker database. true plant (i.e., NLPM will be used to simulate the feedback,
We chose a Renault Megane for its similarities to the Lancia resulting in perfect prediction of NMPC).
Delta. To replicate the conditions under which we obtained Note, however, that we do not attempt to produce a lin-
the Lancia results (see Table I), we drive the Megane with the ear controller that outperforms carefully designed nonlinear
same set of inputs the Lancia was driven. We feed gas pedal controllers, nor do we make such claims; rather, we propose
G, brake pedal B, and steering angle to the IPGCarMaker a novel linear data-driven solution in the context of MPC
simulation environment, but we allow the internal driver to for autonomous driving. Thereby we use a general NMPC
choose the appropriate gear. Furthermore, since the vehicles solution provided by the nlmpc object in MATLAB. The
are not identical, their response will necessarily be different. comparison is based on the driving performance of our two-
In view of this, we do not load any particular route, but let step controller over certain cornering maneuvers, with the
the vehicle to be driven over a flat square of pavement, 60 km reference tracking step being solved by the NMPC and linear
in length. data-driven MPC (a non-structurally robust implementation,
The results of this modeling experiment are shown in which does not tighten constraints nor allocates part of the
Table II. Note that although we fed gas and brake pedal control authority for a corrective action).
signals, we still use post-gearbox torque and master cylinder At each time instant, the NMPC solves the OCP PnN (x(t)) :
pressure as inputs to our model. We do so to avoid the minu JNn (u) subject to (for k = t, . . . , t + N − 1)
nonlinearities that most likely exist in the engine dynamics.
The performance metrics are similar to those from the Lancia x(k + 1|t) = f d (x(k|t), u(k|t))
Delta car both in magnitude and shape, and hence, we also (x(k|t), u(k|t)) ∈ (X × U)
move forward with the uncoupled models for the design of
the TMPC (see Appendix for the model parameter values). where f d represents a discretization of the NLPM, and
−1
t+N
C. Modeling Error Quantification JNn (u) = x̂(k|t)T Q x̂(k|t) + u(k|t)T Ru(k|t)
It follows from the fit metrics that the data-driven models, k=t
for both the Lancia and the Megane, present some modeling x̂(k|t) = x(k|t) − x ss(t), u(k|t) = u(k|t) − u(k − 1|t).

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
1404 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 31, NO. 3, MAY 2023

TABLE IV TABLE VI
S TEP 1 PARAMETERS (RHNOCP) FOR THE L ANCIA D ELTA AND M EGANE P ERFORMANCE C OMPARISON (MPC/NMPC) FOR THE L ANCIA D ELTA

TABLE VII
S TEP 2 PARAMETERS (TMPC) FOR THE L ANCIA D ELTA

TABLE V
S TEP 2 PARAMETERS (MPC/NMPC) FOR THE L ANCIA D ELTA

defined as

Tf
C= x̂(t)T Q x̂(t).
t=0
Table VI summarizes the results. The difference in C across
the maneuvers is at most 4%, indicating that the linear data-
driven MPC performs comparably to the NMPC at tracking
the velocity, despite the model mismatch. Furthermore, MPC
is able to keep the car 40% closer to the lane’s center line
when compared with NMPC (see ȳ). We conclude from these
simulations that the performance of the linear data-driven MPC
Note that input tracking is replaced by input rate weighting. is within the neighborhood of the NMPC performance, and
This is due to the complexity of the nonlinear model, which thus provides an alternative to nonlinear autonomous driving
hinders the online computation of steady-state inputs. solutions.
The Lancia Delta model was obtained with data gathered
during normal driving on public roads, and hence, we assess B. Disturbance Rejection (Robustness)
our algorithm’s capabilities against cornering maneuvers found We now demonstrate the capabilities of our proposed path-
in the validation portion of our dataset. Fig. 4(a) and (f) following algorithm with respect to disturbance rejection.
presents the test cornering maneuvers and their optimized To do so, we perform simulations in which the feedback is
Bezier curves, running through the center line of the driving generated by the linear data-driven model solved over the
lane. These will be referred to as Maneuvers I and II hence- sampling period. We do this only on the Lancia Delta model
forth. These were driven with a road target velocity of 40 for brevity.
and 60 km/h, respectively, using ᾱ = 2π and v g = 10 km/h As evidenced by the constraint tightening in Assumption 1,
as modulating parameters. These targets and modulation para- the RPI set S takes away some autonomy from MPC and
meters are chosen given the understeering characteristics of allocates it to the linear feedback gain K T . It is usual then to
the Lancia Delta. try to find the smallest possible RPI set given a disturbance set
The parameters associated with Step 1 (reference genera- W [30]. Finding the minimal RPI set is usually computation-
tion via RHNOCP) can be found in Table IV. For Step 2, ally expensive and seldom provides a sensible result; however,
we consider box constraint for states and inputs with VLB ≤ given the simplicity of our model (two single-input single-
V (t) ≤ VUB and analogous bounds for the remaining variables. output models stacked), we can easily find said RPI set for any
Table V lists the parameters chosen for both the controllers. pair K T and W. Note that it depends on K T whether Z and
Note also that we consider a negative lower bound on velocity; V are nonempty. Table VII lists the new/updated parameters
however, this is only to accommodate for theoretical require- chosen for the TMPC (the rest can be found in Table V).
ments of MPC that require the origin to be in the interior The bounds on the acceptable disturbance are larger than
of the constraint sets. We do not expect the vehicle to go in the quantified model error to also allow for some external
reverse. perturbations in the built-in robustness of TMPC.
The control objectives are to stay in the middle of the We first discuss the advantages of built-in robustness as
lane and travel at a certain speed, and hence, performance is opposed to relying on inherent robustness. Fig. 4(a)–(e)
assessed with respect to the maximum lateral deviation from presents a comparison in performance of the proposed con-
the lane’s center line, defined by ȳ, and the tracking cost is troller for when Step 2 is executed by vanilla MPC (inherent

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
HERNANDEZ VICENTE et al.: FAST TMPC FOR DRIVERLESS CARS USING LINEAR DATA-DRIVEN MODELS 1405

Fig. 4. (a)–(e) Cornering Maneuver I (Lancia). (a) Map and Bezier curve optimized to follow the center of the rightmost lane. (b) Lateral deviation from the
center line of the vehicle with MPC in Step 2. (c) Lateral deviation from the center line of the vehicle with TMPC in Step 2. (d) Histogram of the velocity V
measurement. (e) Map and trajectories of several TMPC simulations. (f)–(i) Cornering Maneuver II. (f) Map and Bezier curve optimized to follow the center
of the rightmost lane, (g) Unmodulated target velocity and true vehicle velocity for MPC and TMPC. (h) Lateral deviation from the center line of the vehicle
with MPC and TMPC. (i) Map and trajectories of MPC and TMPC simulations.

robustness) and TMPC (built-in robustness). Note that by Maneuver I was driven with a road target velocity of 100 km/h
vanilla MPC, we refer to a standard non-structurally robust without modulation, which amounts to drive at the velocity
MPC implementation, which does not tighten constraints nor constraint (see Table IV).
allocates part of the control authority for a corrective action. Fig. 4(b) and (c) shows the lateral deviation from the
The parameters for this standard MPC controller are, where center line obtained with MPC and TMPC (respectively) for
applicable, as depicted in Table VII. For this simulation, 100 independent and random realizations of a disturbance

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
1406 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 31, NO. 3, MAY 2023

sequence (the same 100 realizations are fed to the algorithm that similar results are obtained when the feedback is realized
with MPC and TMPC for fairness of comparison). The width with the NLPM (allowing for reduced target velocities and
of the vehicle is 1.8 m, and the estimated lane width is 3.5 m; modulation); however, they are omitted here for brevity.
it follows then that both the controllers are able to keep the Remark 3: The source of the high-frequency lateral vibra-
vehicle within acceptable distance from the center line with tion shown in Fig. 4(b) and (c) is mainly due to the persistent
deviations of less than 0.4 m in any direction. When compared perturbation we include in this simulation. The latter is not
with TMPC, MPC is able to complete the maneuver in shortest necessarily representative of real driving conditions, and thus,
time (about 14 s or 24% faster) and also reduce the lateral we defer passenger comfort analysis for Section VI, where a
deviation by about 15 cm. This is because MPC does not need high-fidelity suite is used for simulating the vehicle dynamics
to allocate control authority to a linear gain, which ultimately (providing a more realistic model uncertainty setup).
modulates the control action for disturbance rejection pur-
poses, but also because TMPC needs to restrict the reachable C. Computational Performance
states (constraint tightening) thereby restricting the maximum Control systems designed for autonomous driving must be
achievable velocity to be 98.6 km/h in this case. However, able to produce control actions at sufficient sampling rates
since MPC does not explicitly account for the disturbance, using the available hardware. Our controller, as depicted in
and the target is to drive exactly at the velocity constraint, Section III, generates a new control action at each time instant
the vehicle’s velocity does violate the imposed constraint in via the solution of two optimization problems, the RHNOCP
several occasions when controlled by vanilla MPC, as depicted solved via IPOPT and the TMPC solved by means of the
in the histogram in Fig. 4(d). Nevertheless, given the nature of fast MPC algorithm, henceforth referred to as FMPC. IPOPT
the disturbance and the move-blocking MPC technique which and FMPC have been chosen due to the array of algorithmic
provides smoother references, the constraint is violated in only features they enjoy, which make them computationally more
0.7% of the sampling times, and by a maximum of 0.2 m/s or efficient when compared with other solvers [31], [36]; never-
0.7%. Finally, Fig. 4(e) shows the first ten TMPC trajectories theless, the real-time feasibility of any algorithm is contingent
plotted against the map, which confirms the capability of to the hardware on which it will be implemented. To give
TMPC to follow the reference generated in Step 1 and keep an approximate estimate of computational performance on a
the vehicle close to the lane’s center line. realistic embedded system, such as would be used in a driver-
Given the results in Fig. 4(b)–(d), the built-in robustness less car, we evaluated a performance benchmark scaled to
might seem an unnecessary trade-off in the form of removing the NVIDIA Jetson TX2 [51] low-power-embedded processing
control authority from the MPC controller, but the next set of board, which uses the same Tegra X2 chipset as the NVIDIA
results, depicted in Fig. 4(f)–(i), show that this is not the case. Drive PX2 [52] made specifically for the driverless car market.
Maneuver II, shown in Fig. 4(f), represents a sharper corner We do not attempt to claim, at this stage, that our control
when compared with Maneuver I, and thus it requires a much architecture is real-time capable, since it is out of the scope
lower velocity to be driven safely. Maneuver II was driven with of this article to deploy the algorithm in target hardware
a target velocity of 25 m/s without modulation, and a single akin to that used in autonomous driving applications. Instead,
realization of the disturbance which was set to w(t) = [0.20] , we perform a computational performance analysis in line with
which is a fixed maximum allowable disturbance in velocity the original message in [31] and [36], that is, to study the effect
and zero disturbance in yaw rate. This type of disturbance of the algorithmic improvements that IPOPT and FMPC have
could represent an unexpected road slope or an unaccounted over alternative, possibly naive, solutions.
actuator response (model uncertainty). Fig. 4(g) shows the We study the processing time of our algorithm under the
target velocity and the velocity profiles obtained by MPC following conditions.
and TMPC. Both the controllers converge to a value with
1) Step 1 solved via
a steady-state error; however, MPC relies solely on inherent
robustness, and hence it converges to 14.7 m/s or 52.9 km/h, a) The MATLAB function fminsearch, which imple-
more than three times the true target velocity. TMPC, on the ments a direct search method.
other hand, uses the linear gain K T to actively correct for the b) The IPOPT algorithm as obtained from [37].
disturbance and hence converges with a steady-state error of 2) Step 2 solved via
5.71%. This, in turn, means that MPC is not able to complete a) The MATLAB function quadprog which is opti-
the cornering maneuver safely, as evidenced by the lateral mized to solve optimization problems such as (9)
deviation of almost 4 m depicted in Fig. 4(h). On the other and (10).
hand, TMPC completes the maneuver with a maximum lateral b) The FMPC algorithm as obtained from [53].
deviation of 10.29 cm. The quadprog function is the go-to option for solving quadratic
The results depicted in Fig. 4 show that in many situations programming problems in MATLAB, although it is not
the inherent robustness of MPC might be enough to provide necessarily optimized for real-time capabilities. The FMPC
satisfactory constraint handling and better performance than algorithm [53] has several limitations when compared with
TMPC due to the retention of full control authority. However, quadprog, being a key one that an arbitrary terminal set is not
there exist scenarios in which the lack of built-in robustness allowed, that is, Z f (z ss , v ss ) = Z  {z ss }. Nevertheless, it has
could result in lane departure and possible collision. In view of been shown [20] that terminal constraints can be replaced by
this, we proceed with TMPC for the remaining analysis. Note a long enough prediction horizon, which we assume to be our

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
HERNANDEZ VICENTE et al.: FAST TMPC FOR DRIVERLESS CARS USING LINEAR DATA-DRIVEN MODELS 1407

TABLE VIII
S CALED P ROCESSING T IME (M ANEUVER I—L ANCIA D ELTA —P ERTURBED ). C OMPARISON B ETWEEN fminsearch-quadprog AND IPOPT–FMPC

case with N = 40. Finally, we fed the solution at the previous obtained via the implementation of FMPC is able to heavily
sampling time as warm-start for all the solvers. counter the increased load of IPOPT, resulting in overall
All the processing time results were obtained by running faster computations for the IPOPT–FMPC pair, even when the
our proposed control algorithm in MATLAB R2017b, on an horizon of the RHNOCP is split into more blocks.
Intel3 Core4 i7-4700MQ CPU at 2.40-GHz processor, fit with In summary, the algorithmic improvements provided by
L1, L2, and L3 cache memories of 0.256, 1, and 6 MB the IPOPT–FMPC pair result in maximum processing times
respectively. An approximate scaling is then performed to of 41% of the sampling time, a decrease of 17% when
estimate the processing times on Jetson TX2 based on the compared with fminsearch–quadprog. These results are well
results of running the Netlib Whetstone Benchmark [54] on below the 100% sampling time threshold imposed by Berntorp
the desktop computer, and Jetson TX2. and Magnusson [34], and in line with similar approaches
We assessed the performance of our controller with a sim- such as Jalali et al. [19], which reported maximum process-
ulation of Maneuver I under a single realization of the distur- ing times of 42% in experiments, and [15], [55], which in
bance. Different prediction horizons were used for RHNOCP; simulation achieved average processing times of 50% and
however, fminsearch was able to find a stabilizing solution 28% (respectively).
only when a single block was used, and hence, an addi- Furthermore, note that to achieve these processing capa-
tional simulation is performed for the pair IPOPT–FMPC with bilities, [19], [55] resort to the LTV models and LTV-MPC
multiple blocks. Table VIII summarizes the computational algorithms with relatively short prediction horizons (3 and 10,
performance of our algorithm in terms of mean processing respectively). NMPC (based on a nonlinear physical model)
time, its standard deviation, and maximum processing time (all is solved in [15] with an off-the-shelf tool; however, reporting
as a percentage of the sampling time). As expected, increasing relatively long processing times (50% of the sampling time in
the prediction horizon of the RHNOCP (from 10 to 18) average) and in a race-car driving setup in which deviations
increases the time needed to obtain a reference, independent of over 5 m from the center line seem to be acceptable.
of the strategy used to solve the optimization. Moreover, Our approach, on the other hand, is able to leverage on the
although small, there exist variations in the time taken to solve mathematical simplicity of the uncoupled data-driven model
Step 2 for different horizons of RHNOCP (via either quadprog to achieve processing times under 30% of the sampling time
or FMPC). The reason for these is that different horizons (in average) and to provide safe driving in public roads.
in RHNOCP will unavoidably result in different references,
hence modifying the TMPC optimization problem. VI. C AR M AKER S IMULATION
The average and maximum time used to solve Step 1 The results in Section V show the ability of our pro-
increases by a factor of 1.3 when IPOPT is implemented; posed controller to safely steer a vehicle throughout corner-
however, using a larger number of blocks does not necessarily ing maneuvers for a perfect prediction model. To analyze
result in further increments. Indeed, solving RHNOCP via the performance of our control architecture in the presence
IPOPT with (Ng , Nc ) = (18, 1) takes a maximum of 40% of of modeling uncertainty, we use the high-fidelity simulator
the sampling time (and an average of 21%); however, when IPGCarMaker. Note, however, that the IPGCarMaker software
three blocks are used, the RHNOCP is solved in a maximum does not allow to define post-gearbox torque externally (i.e.,
of 34% of the sampling time (and an average of 28%). as a control input), since it would require bypassing sev-
Oppositely, solving Step 2 with FMPC results in a reduction of eral engine modules for axle torque computation. In view
an order of magnitude in the average and maximum processing of this, we map our input TP to gas and brake pedals
times, when compared with the usage of quadprog. as follows:
The bottleneck seems to be solving the RHNOCP (Step 1), 
which can take between 31% and 40% of the sampling (TP(t)/TPUB , 0), TP(t) ≥ 0
time depending on the horizon length and number of blocks. (G(t), B(t)) =
(0, TP(t)/TPLB ), TP(t) < 0.
However, the implementation of IPOPT allows for the usage
of more blocks in the solution of RHNOCP, resulting in This mapping neglects engine and brake system dynamics and
references with better characteristics (as observed in the com- hence introduces uncertainty. A possible solution to this issue
fort analysis). Furthermore, the reduction in processing time is to use the Hammerstein–Wiener models to find a (static)
nonlinear map between the pedals and the chosen inputs;
3 Registered trademark. however, this remains as future work. For the following results,
4 Trademarked. we allow for increased uncertainty bounds to account for this.

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
1408 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 31, NO. 3, MAY 2023

Fig. 5. CarMaker simulation of cornering Maneuver I (Megane). (a) Velocity profile for v g = 8.33 m/s. (b) Velocity profile for v g = 13.89 m/s. (c) Lateral
deviation from the lane center line. (d) Map and trajectories of the IPGCarMaker simulation.

TABLE IX of 1.8 m. Finally, Fig. 5(d) shows both the trajectories plotted
C ONTROLLER PARAMETERS FOR THE M EGANE against the map. The maximum difference in lateral deviation
with respect to the s coordinate (i.e., when passing through
the same patch of track) is under 30 cm, which explains the
apparent superimposition of the trajectories in view of the map
resolution.

A. Passenger Comfort Analysis

To assess the performance of our controller in terms of
passenger comfort, we use measures of local and average
comfort. In particular, we observe maximum lateral acceler-
ation ā y , maximum lateral jerk J¯y , and frequency content of
the lateral acceleration within certain ranges, measured via
Table IX shows all the controller parameters that changed the power spectrum PSDa y , which measure local and average
from the Lancia simulation (see Tables IV and VII). Fig. 5 comfort [56]. Table X shows the acceptable limits of these
shows the results of driving the Megane through Maneuver I indicators, alongside their values for Maneuver I simulated
with a road target velocity Vg of 80 km/h and modulating with v g = 8.33 m/s. It is clear from Table X that although
velocities v g of 30 and 50 km/h. Fig. 5(a) and (b) shows no specific comfort requirement has been included in the
how the modulated target velocity V f (green-dotted line) design of our two-step controller, the instantaneous measure
changes depending on the modulating velocity v g and the of comfort related to the lateral acceleration is met; however,
road curvature it can be seen that the reference velocity is the maximum lateral jerk is an order of magnitude over the
tracked accurately while it experiences slow changes (for comfort limit.
example, in the interval from initialization until 30 s), despite The source of this excessive jerk is twofold, the first being
the modeling error described in II. the corrective action of the TMPC. The second contributor
Fig. 5(c) shows the lateral deviation from the center line for is the constant update of the reference yaw rate by RHNOCP,
both the modulating velocities. A larger modulating velocity which is necessary to account for the modeling dichotomy that
results in larger target velocities throughout the maneuver, thus exists between the two steps of our controller. To improve the
in the vehicle being driven faster. This, in turn, results in a comfort performance of our controller, we include a filter to
larger lateral deviation [as depicted in Fig. 5(c)], particularly attenuate the high-frequency content present in the reference
when the vehicle is in the middle of the roundabout. Neverthe- generated by RHNOCP. We implement a first-order low-pass
less, the maximum deviation remains under acceptable values filter with a cutoff frequency of 1.5 Hz. Table X shows that this
given an estimated lane width of 3.5 m and a vehicle width does result in a decrease in the maximum lateral jerk, however,

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
HERNANDEZ VICENTE et al.: FAST TMPC FOR DRIVERLESS CARS USING LINEAR DATA-DRIVEN MODELS 1409

TABLE X R EFERENCES
PASSENGER C OMFORT C OMPARISON B ETWEEN F ILTERED AND
U NFILTERED R EFERENCES FOR M ANEUVER I [1] J. Levinson et al., “Towards fully autonomous driving: Systems and
W ITH v g = 8.33 m/s (M EGANE ) algorithms,” in Proc. IEEE Intell. Vehicles Symp., Baden, Germany,
Jun. 2011, pp. 163–168.
[2] M. Doumiati, A. Victorino, A. Charara, and D. Lechner, “Estimation
of vehicle lateral tire-road forces: A comparison between extended
and unscented Kalman filtering,” in Proc. Eur. Control Conf. (ECC),
Aug. 2009, pp. 4804–4809.
[3] M. M. Shirazi and A. B. Rad, “L1 adaptive control of vehicle lateral
dynamics,” IEEE Trans. Intell. Vehicles, vol. 3, no. 1, pp. 92–101,
Mar. 2018.
[4] E. H. M. Lim and J. K. Hedrick, “Lateral and longitudinal vehicle control
coupling for automated vehicle operation,” in Proc. IEEE Amer. Control
Conf., Jun. 1999, pp. 3676–3680.
at the expense of worsened average comfort (as measured by [5] M. H. Lee et al., “Lateral controller design for an unmanned vehicle via
the acceleration frequency content). Kalman filtering,” Int. J. Automot. Technol., vol. 13, no. 5, pp. 801–807,
2012.
Given that the main source of the large jerk is RHNOCP [6] G. Tagne, R. Talj, and A. Charara, “Design and comparison of robust
optimization, we also perform a simulation with a reduced nonlinear controllers for the lateral dynamics of intelligent vehicles,”
amount of blocks in the move-blocking MPC strategy, IEEE Trans. Intell. Transp., vol. 17, no. 3, pp. 796–809, Mar. 2015.
[7] C. A. Lua and S. Di Gennaro, “Nonlinear adaptive tracking for ground
to reduce the frequency content of the optimal reference passed vehicles in the presence of lateral wind disturbance and parameter
on to TMPC. This results, as portrayed in Table X, in a variations,” J. Franklin Inst., vol. 354, no. 7, pp. 2742–2768, May 2017.
considerable reduction of the maximum jerk and a slight [8] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert, “Con-
strained model predictive control: Stability and optimality,” Automatica,
improvement in all the other comfort measures. Although vol. 36, no. 6, pp. 789–814, 2000.
the maximum jerk is still over the recommended limit, this [9] P. F. Lima, G. C. Pereira, J. Martensson, and B. Wahlberg, “Experimental
threshold is surpassed only during 14% of the total maneuver validation of model predictive control stability for autonomous driving,”
Control Eng. Pract., vol. 81, pp. 244–255, Dec. 2018.
duration (around 9 s in aggregate), and the mean jerk is [10] U. Rosolia, S. de Bruyne, and A. G. Alleyne, “Autonomous vehicle
0.57 m/s3 , a reduction of 2.3 times when compared with the control: A nonconvex approach for obstacle avoidance,” IEEE Trans.
unfiltered maneuver. Control Syst. Technol., vol. 25, no. 2, pp. 469–484, Mar. 2017.
[11] Y. Gao et al., “Spatial predictive control for agile semi-autonomous
ground vehicles,” in Proc. 11th Int. Symp. Adv. Vehicle Control, 2012,
VII. C ONCLUSION pp. 1–6.
[12] C. Liu, S. Lee, S. Varnhagen, and H. E. Tseng, “Path planning for
In this article, we have proposed a two-step robust MPC autonomous vehicles using model predictive control,” in Proc. IEEE
controller for the purpose of path-following. The key advan- Intell. Vehicles Symp. (IV), Jun. 2017, pp. 174–179.
[13] Z. Q. Sun, L. Dai, K. Liu, Y. Q. Xia, and K. H. Johansson, “Robust MPC
tage of our controller is the use of uncoupled LTI data- for tracking constrained unicycle robots with additive disturbances,”
driven models for the prediction of lateral and longitudinal Automatica, vol. 90, pp. 172–184, Apr. 2018.
dynamics, which show good fit to data and small one-step [14] P. Falcone, F. Borrelli, J. Asgari, H. E. Tseng, and D. Hrovat, “Predictive
active steering control for autonomous vehicle systems,” IEEE Trans.
ahead prediction error. These characteristics allow for a good Control Syst. Technol., vol. 15, no. 3, pp. 566–580, May 2007.
control performance with linear MPC with built-in robust- [15] A. Buyval, A. Gabdulin, R. Mustafin, and I. Shimchik, “Deriving
ness, tested against a high-fidelity simulator in IPGCarMaker. overtaking strategy from nonlinear model predictive control for a race
car,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Sep. 2017,
Furthermore, the implementation of linear MPC allows for pp. 2623–2628.
computationally efficient algorithms for solving the online [16] S. Arrigoni, F. Braghin, and F. Cheli, “Decentralized fast-MPC path
optimization problem, achieving processing times which show planner for automated vehicles,” in Proc. Int. Conf. Electr. Electron.
Technol. Automot., Jun. 2017, pp. 1–7.
their advantage over naive solvers, and could provide an [17] P. Falcone, F. Borrelli, H. E. Tseng, J. Asgari, and D. Hrovat, “Linear
advantageous baseline for future implementation in target time-varying model predictive control and its application to active
hardware. steering systems: Stability analysis and experimental validation,” Int.
J. Robust Nonlinear Control, vol. 18, no. 8, pp. 862–875, 2008.
[18] C. E. Beal and J. C. Gerdes, “Model predictive control for vehicle
A PPENDIX stabilization at the limits of handling,” IEEE Trans. Control Syst.
Technol., vol. 21, no. 4, pp. 1258–1269, Jul. 2012.
A. Uncoupled Linear Models Used for Predictions [19] M. Jalali, S. Khosravani, A. Khajepour, S.-K. Chen, and B. Litkouhi,
“Model predictive control of vehicle stability using coordinated active
steering and differential brakes,” Mechatronics, vol. 48, pp. 30–41,
Dec. 2017.
[20] L. Grüne and J. Pannek, Nonlinear Model Predictive Control: Theory
Algorithms, 1st ed. London, U.K.: Springer, 2011.
[21] J. Liu, P. Jayakumar, J. L. Stein, and T. Ersal, “A study on model
fidelity for model predictive control-based obstacle avoidance in high-
speed autonomous ground vehicles,” Vehicle Syst. Dyn., vol. 54, no. 11,
pp. 1629–1650, Nov. 2016.
[22] Y. Gao, A. Gray, H. E. Tseng, and F. Borrelli, “A tube-based robust non-
linear predictive control approach to semiautonomous ground vehicles,”
Veh. Syst. Dyn., vol. 52, no. 6, pp. 802–823, Jun. 2014.
ACKNOWLEDGMENT [23] E. Alcala, V. Puig, J. Quevedo, and O. Sename, “Fast zonotope-tube-
The authors would like to thank A. Saroldi and Centro based LPV-MPC for autonomous vehicles,” 2020, arXiv:2009.02248.
[24] S. A. Eker and M. Nikolaou, “Linear control of nonlinear systems:
Ricerche Fiat for their contribution in collecting the experi- Interplay between nonlinearity and feedback,” AIChE J., vol. 48, no. 9,
mental data. pp. 1957–1980, Sep. 2002.

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.
1410 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 31, NO. 3, MAY 2023

[25] S. M. Yoon, K. S. Lee, S. Y. Kim, J. H. Kang, and M. H. Lee, “Lateral [48] P. O. M. Scokaert, D. Q. Mayne, and J. B. Rawlings, “Suboptimal model
control of an UCT(unmanned container transporter) using ultrasonic predictive control (feasibility implies stability),” IEEE Trans. Autom.
satellite system and system identification,” in Proc. Int. Conf. Control, Control, vol. 44, no. 3, pp. 648–654, Mar. 1999.
Autom. Syst., Oct. 2008, pp. 296–300. [49] G. Pannocchia, J. B. Rawlings, and S. J. Wright, “Is suboptimal
[26] N. Liu and A. G. Alleyne, “Iterative learning identification applied to nonlinear MPC inherently robust?” in Proc. 18th IFAC World Congr.,
automated off-highway vehicle,” IEEE Trans. Control Syst. Technol., 2011, pp. 7981–7986.
vol. 22, no. 1, pp. 331–337, Jan. 2014. [50] D. A. Allan, C. N. Bates, M. J. Risbeck, and J. B. Rawlings, “On the
[27] M. Canale, L. Fagiano, and M. C. Signorile, “Nonlinear model predictive inherent robustness of optimal and suboptimal nonlinear MPC,” Syst.
control from data: A set membership approach,” Int. J. Robust Nonlinear Control Lett., vol. 106, pp. 68–78, Aug. 2017.
Control, vol. 24, no. 1, pp. 123–139, Jan. 2014. [51] NVIDIA. (2022). Jetson TX2 Module. Accessed: Jun. 17, 2022. [Online].
[28] M. Gevers, “Identification for control: From the early achievements to Available: https://developer.nvidia.com/embedded/jetson-tx2
the revival of experiment design,” Eur. J. Control, vol. 11, nos. 4–5, [52] NVIDIA. (2018). 5.x_Linux_DPX_SDK. Accessed: Jun. 17,
pp. 335–352, Jan. 2005. 2022. [Online]. Available: https://docs.nvidia.com/drive/active/
[29] B. A. H. Vicente, S. S. James, and S. R. Anderson, “Linear system 5.0.10.3L/nvvib_docs/
identification versus physical modeling of lateral–longitudinal vehi- [53] Y. Wang and S. Boyd. (2008). Fast Model Predictive Control Using
cle dynamics,” IEEE Trans. Control Syst. Technol., vol. 29, no. 3, Online Optimization V Alpha [Source Code]. [Online]. Available:
pp. 1380–1387, May 2021. https://web.stanford.edu/~boyd/papers/fast_mpc.html
[30] D. Q. Mayne, M. M. Seron, and S. V. Raković, “Robust model predic- [54] R. Painter. C Converted Whetstone Double Precision Benchmark.
tive control of constrained linear systems with bounded disturbances,” Accessed: May 15, 2022. [Online]. Available: http://www.netlib.org/
Automatica, vol. 41, no. 2, p. 219–224, 2005. benchmark/whetstone.c
[31] Y. Wang and S. Boyd, “Fast model predictive control using online [55] K. Liu, J. Gong, A. Kurt, H. Chen, and U. Ozguner, “Dynamic modeling
optimization,” IEEE Trans. Control Syst. Technol., vol. 18, no. 2, and control of high-speed automated vehicles for lane change maneuver,”
pp. 267–278, Mar. 2010. IEEE Trans. Intell. Veh., vol. 3, no. 3, pp. 329–339, Sep. 2018.
[32] IPG Automotive. CarMaker: Virtual Testing of Automobiles and Light- [56] J. Eriksson and L. Svensson, “Tuning for ride quality in autonomous
Duty Vehicles. [Online]. Available: https://ipg-automotive.com/products- vehicle application to linear quadratic path planning algorithm,” Ph.D.
services/simulation-software/carmaker/ dissertation, Dept. Inf. Technol., Uppsala Univ., Uppsala, Sweden, 2015.
[33] L. Ljung, System Identification Theory for the User, 2nd ed.
Upper Saddle River, NJ, USA: Prentice-Hall, 1999.
[34] K. Berntorp and F. Magnusson, “Hierarchical predictive control for
ground-vehicle maneuvering,” in Proc. Amer. Control Conf. (ACC), Bernardo A. Hernandez Vicente received the
Jul. 2015, pp. 2771–2776. M.Eng. degree from the University of Concepcion,
[35] R. C. Shekhar and J. M. Maciejowski, “Robust variable horizon MPC Concepcion, Chile, in 2013, and the Ph.D. degree in
with move blocking,” Syst. Control Lett., vol. 61, no. 4, pp. 587–594, model predictive control theory from the Department
Apr. 2012. of Automatic Control and Systems Engineering, The
[36] A. Wächter and L. T. Biegler, “On the implementation of an interior- University of Sheffield, Sheffield, U.K., in 2018.
point filter line-search algorithm for large-scale nonlinear programming,” Since 2021, he has been an Assistant Professor
Math. Program., vol. 106, no. 1, pp. 25–57, May 2006. with the Department of Mechanical Engineering,
[37] E. Bertolazzi. (2015–2021). Mexipopt V1.1.2 [Source Code]. [Online]. University of Concepcion.
Available: https://github.com/ebertolazzi/mexIPOPT
[38] J. B. Rawlings and D. Q. Mayne, Model Predictive Control: Theory
Design. Madison, Wisconsin: Nob-Hill, 2014.
[39] K. Ilya and E. G. Gilbert, “Theory and computation of disturbance
invariant sets for discrete-time linear systems,” Math. Problems Eng., Paul A. Trodden (Member, IEEE) received the
vol. 4, no. 4, pp. 317–367, 1998. M.Eng. degree in engineering science from the
[40] D. Q. Mayne, S. V. Raković, R. Findeisen, and F. Allgöwer, “Robust University of Oxford, Oxford, U.K., in 2003, and
output feedback model predictive control of constrained linear systems,” the Ph.D. degree in aerospace engineering from the
Automatica, vol. 42, no. 7, pp. 1217–1222, Jul. 2006. University of Bristol, Bristol, U.K., in 2009.
[41] B. A. H. Vicente and P. A. Trodden, “Stabilizing predictive control with Since 2012, he has been a Lecturer with the
persistence of excitation for constrained linear systems,” Syst. Control Department of Automatic Control and Systems
Lett., vol. 126, pp. 58–66, Apr. 2019. Engineering, The University of Sheffield, Sheffield,
[42] P. A. Trodden and J. M. Maestre, “Distributed predictive control with U.K. His research interests include robust and dis-
minimization of mutual disturbances,” Automatica, vol. 77, pp. 31–43, tributed model predictive and optimization-based
Mar. 2017. control, and control and optimization for power and
[43] D. Simon, J. Lofberg, and T. Glad, “Reference tracking MPC using energy systems.
dynamic terminal set transformation,” IEEE Trans. Autom. Control,
vol. 59, no. 10, pp. 2790–2795, Oct. 2014.
[44] D. Limon, M. Pereira, D. M. D. La Pena, T. Alamo, C. N. Jones, and Sean R. Anderson received the M.Eng. degree in
M. N. Zeilinger, “MPC for tracking periodic references,” IEEE Trans. control systems engineering and the Ph.D. degree in
Autom. Control, vol. 61, no. 4, pp. 1123–1128, Apr. 2016. nonlinear system identification and predictive con-
[45] J. Kohler, M. A. Müller, and F. Allgower, “Nonlinear reference tracking: trol from the Department of Automatic Control and
An economic model predictive control perspective,” IEEE Trans. Autom. Systems Engineering, The University of Sheffield,
Control, vol. 64, no. 1, pp. 254–269, Jan. 2019. Sheffield, U.K., in 2001 and 2005, respectively.
[46] P. Falugi, “Model predictive control for tracking randomly vary- From 2005 to 2010, he was a Research Asso-
ing references,” Int. J. Control, vol. 88, no. 4, pp. 745–753, ciate with the Neural Algorithms Research Group,
Apr. 2015. The University of Sheffield. From 2010 to 2011,
[47] P. R. B. Monasterios and P. A. Trodden, “Model predictive control he was a Research Associate with the Department
of linear systems with preview information: Feasibility, stability, and of Automatic Control and Systems Engineering, The
inherent robustness,” IEEE Trans. Autom. Control, vol. 64, no. 9, University of Sheffield, where he became a Lecturer in 2012 and has been a
pp. 3831–3838, Sep. 2019. Senior Lecturer since 2015.

Authorized licensed use limited to: Harbin Engineering Univ Library. Downloaded on May 02,2024 at 13:52:35 UTC from IEEE Xplore. Restrictions apply.