Вы находитесь на странице: 1из 7

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/327092371

Online Learning-based Receding Horizon Control of Tilt-rotor Tricopter: A


Cascade Implementation

Conference Paper · June 2018


DOI: 10.23919/ACC.2018.8430814

CITATION READS

1 56

2 authors:

Mohit Mehndiratta Erdal Kayacan


Nanyang Technological University Aarhus University
9 PUBLICATIONS   30 CITATIONS    149 PUBLICATIONS   2,290 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Automated Construction Quality Assessment Robot System View project

Special Session on " Advances to Type-2 Fuzzy Logic Control", WCCI- 2018 View project

All content following this page was uploaded by Mohit Mehndiratta on 02 April 2019.

The user has requested enhancement of the downloaded file.


Online Learning-based Receding Horizon Control of Tilt-rotor
Tricopter: A Cascade Implementation
Mohit Mehndiratta and Erdal Kayacan

Abstract— This study manifests a learning-based cascade the literature and have been successfully implemented in
nonlinear model predictive control (NMPC) algorithm for numerous ground and aerial robotic applications [8]–[10].
the trajectory tracking of a tilt-rotor tricopter UAV; wherein While EKF is based on the linearization of a nonlinear
two time-varying aerodynamic parameters, thrust and drag-
moment coefficients, are estimated online incorporating nonlin- system at the current estimate, MHE exploits the past mea-
ear moving horizon estimation method. Since the performance surements available over a window and subsequently solves
of a model-based controller is guaranteed for an accurate an optimization problem. In this study, we prefer to utilize
mathematical model of the system to be controlled, it is indeed MHE-based learning over EKF-based learning due to the
important to estimate the changing dynamics in order to make contained estimation ability of the former.
NMPC adaptive – and therefore robust – to the time-varying
operational disturbances. To further illustrate the tracking A tilt-rotor tricopter is far more maneuverable than a usual
capability of learning-based cascade NMPC, a complex square- quadrotor UAV. However, this additional maneuverability
shaped trajectory is flown and is observed to be well tracked. comes at the cost of significant nonlinearities and cross-
To the best of our knowledge, this is the first application couplings that make its control a challenging task. This paper
of an online learning-based cascade NMPC to a complicated employs a cascade NMPC structure with a position control
aerospace system. Moreover, owing to ACADO toolkit, the
overall execution time of the closed-loop is below 4 milliseconds, at high-level (H-NMPC) and an attitude control at low-level
which eventually demonstrates the real-time potential of the (L-NMPC). The importance of utilizing a cascade NMPC
presented control framework. structure over a full-state controller can be acknowledged in
terms of: (i) ability to distribute the overall control task, such
I. I NTRODUCTION that, the two NMPCs can run on different low-cost embedded
Model predictive control (MPC), which was initially de- hardware, instead of running one but on an expensive on-
veloped for process control, has recently gained popular- board computer, and (ii) flexibility to run the two loops at
ity for the control of fast aerospace systems, including different frequencies. This is an important aspect for UAV
multicopter unmanned aerial vehicles (UAVs) [1]–[3]. The control as the low-level controller providing stability has to
reason for its widespread use is its ability to: (i) handle run at a higher rate. In addition, we incorporate nonlinear
multi-input-multi-output systems, (ii) systematically handle MHE (NMHE) in order to learn online the two crucial aero-
system constraints, and (iii) optimize system performance dynamic parameters, i.e thrust and drag-moment coefficients.
via recursive online optimization. Besides linear MPC, the It is indeed important to estimate these parameters because
technological advancements in modern computation power apart from system identification errors, they can vary in-flight
have also paved the way for nonlinear MPC (NMPC), where due to wind disturbance or propeller degradation.
the optimization problem becomes non-convex and hence, This study is organized as follows: Section II starts with
gets difficult to solve. Moreover, the applications involving the tilt-rotor tricopter dynamic model. In Section III, NMPC
aggressive maneuvers of UAVs excite the nonlinear system and NMHE problem formulations are presented. Thereafter,
dynamics, and when approximated by a linear model, results the obtained numerical results are discussed in Section IV.
in a substandard performance. Thanks to the availability of Lastly, in Section V, some conclusions are drawn.
recent fast solvers, like ACADO toolkit [4], NMPC has also II. T ILT- ROTOR T RICOPTER M ODEL
been applied for the tracking control of various ground as
The tilt-rotor tricopter is considered to be a rigid body
well as aerial robots [5]–[7].
having two stationary rotors – right rotor (RR) rotating
Since the performance of MPC is only guaranteed for an
clockwise and left rotor (LR) rotating counter-clockwise –
accurate mathematical model of the system to be controlled,
and one tilting rotor – back rotor (BR) rotating counter-
uncertainties would lead to a suboptimal performance. There-
clockwise – as shown in Fig. 1.
fore, state estimators such as extended Kalman filter (EKF)
and moving horizon estimator (MHE) are often employed A. Kinematic Equations
in conjunction with MPC so that it can adapt according to The kinematic equations for the tilt-rotor tricopter are
the changing operational conditions. The resulting learning- obtained by using the transformation from the body frame
based MPC framework is referred as adaptive MPC in (FB ) to the Earth-fixed frame (FE ). They can be written as:
2 3 2 3 2 ˙ 3 2 3
Mohit Mehndiratta and Erdal Kayacan are with the School of ẋ u p
Mechanical and Aerospace Engineering, Nanyang Technological 4 ẏ 5 = REB 4 v 5 , 4 ✓˙ 5 = TEB 4 q 5 , (1)
University, 639798, Singapore. Email: mohit005@e.ntu.edu.sg, ˙
erdal@ntu.edu.sg
ż w r
𝐹1 𝒛𝑬
𝒙𝑬 where Fx , Fy , Fz are the total external forces and ⌧x , ⌧y ,
Ω1
⌧z are the total external moments acting in frame FB . In
𝑙2 𝜏1 addition, Ixx , Iyy , Izz and Ixz represent the moments of
𝓕𝑬
𝑙3 inertia of the whole tricopter along axes FBx , FBy , FBz and
𝐹2 𝒛𝑩 𝒚𝑬
Ω2 𝒙𝑩 FBxz , respectively. One may note that the tilt-rotor tricopter
𝑙4 𝜏𝑧
𝜏𝑥 𝑙1 only has a single plane of symmetry, and hence, the effect
𝜏2 𝓕𝑩
𝜇 of asymmetric moment Ixz is explicitly considered in (5).
𝜏𝑦 𝐹3
𝒚𝑩 Ω3 The external aerodynamic forces and moments are gener-
ated when a rotor is propelled through the surrounding air.
𝜏3 Using the momentum theory, steady state thrust and drag-
moment generated by a hovering rotor can be modelled as:

Fig. 1: Coordinate frame and sign conventions of the con- Fi = KF ⌦2i and ⌧i = K⌧ ⌦2i , (6)
sidered tilt-rotor tricopter UAV.
where Fi and ⌧i are the force and drag-moment generated by
where x, y, z and , ✓, are the translational position and ith rotor with ⌦i angular velocity. In addition, KF and K⌧
rotational attitude, respectively, that are defined in frame FE ; are the force and drag-moment coefficients, respectively, of
u, v, w and p, q, r are the translational and rotational ve- the rotor. Moreover, with respect to the tilt-rotor tricopter
locities that are defined in frame FB ; REB is the translation configuration shown in Fig. 1, the expressions for total
transformation matrix between frames FE and FB , whereas external force (Fext ) and total external moment (⌧ext ) acting
TEB maps the rotational velocity component from FB to FE in FB frame, are written as:
(c : cos, s : sin, t : tan): 2 3 2 3
2 3 Fx 0
c✓c s s✓c s c c s✓c + s s Fext = 4 Fy 5 = 4 F3 sin(µ) 5, (7)
REB = 4 c✓s s s✓s + c c c s✓s s c 5, Fz F1 + F2 + F3 cos(µ)
s✓ s c✓ c c✓ 2 3
F2 l 4 F 1 l 2
(2) ⌧ext = 4 (F3 cos(µ))l1 (F1 + F2 )l3 + ⌧3 sin(µ) 5 , (8)
2 3
1 s t✓ c t✓ ⌧1 ⌧2 ⌧3 cos(µ) + (F3 sin(µ))l1
TEB = 4 0 c s 5. (3)
0 s c where µ is the tilting angle of the back rotor and ⌧ext is
c✓ c✓
composed of [⌧x , ⌧y , ⌧z ]T .
B. Rigid-body Equations The constant intrinsic parameters for a custom-designed
Similar to the quadrotor UAV in [11], the rigid-body tilt-rotor tricopter UAV in our laboratory are listed in Table
dynamic equations of the tilt-rotor tricopter are derived based I. While parameters such as mass (m) and the arm lengths
on Newton-Euler formulation in the body coordinate system. (l1 ,l2 ,l3 ,l4 ) are directly measured, the moment of inertia
Within these equations, the tricopter is assumed to be a point values are obtained from its SolidWorks CAD model. On
mass, wherein all the forces act at the center of gravity. In the other hand, the thrust and drag-moment coefficients
general, they are categorized as: are evaluated based on a simple experimentation, details of
Force Equations which are given in [12].
1 For the real-time dynamic optimization framework, the
u̇ = rv qw + gsin(✓) + Fx , (4a) nonlinear tricopter model is discretized based on direct
m
1 multiple shooting method, utilizing a shooting grid size
v̇ = pw ru gsin( )cos(✓) + Fy , (4b)
m of ts = 0.01s. In addition, the explicit Runge-Kutta
1 4th order integrator, with 2 steps per shooting interval is
ẇ = qu pv gcos( )cos(✓) + Fz , (4c)
m incorporated. Moreover, in order to model the reality in this
Moment Equations
⇣ 1 ⌘h TABLE I: Tilt-rotor tricopter intrinsic parameters.
ṗ = 2
{ pq(Ixz ) + qr(Iyy Izz )}Izz
Ixx Izz Ixz
i Par. Description Value
{qr(Ixz ) + pq(Ixx Iyy )}Ixz + ⌧x (Izz ) ⌧z (Ixz ) , m Mass of tricopter UAV 1.412 kg
l1 Arm length for rotor 3 along FBx 0.357 m
(5a) l2 Arm length for rotor 1 along FBy 0.2845 m
⇣I I ⌘ ⇣I ⌘ ⇣ 1 ⌘ l3 Arm length for rotor 1, 2 along FBx 0.149 m
zz xx xz
q̇ = pr (r2 p2 ) + ⌧y , (5b) l4 Arm length for rotor 2 along FBy 0.2755 m
Iyy Iyy Iyy Ixx Moment of inertia about FBx 0.0149 kg-m2
⇣ 1 ⌘h
Iyy Moment of inertia about FBy 0.0219 kg-m2
ṙ = {qr(Ixz ) + pq(Ixx Iyy )}Ixx
Ixx Izz Ixz 2 Izz Moment of inertia about FBz 0.0344 kg-m2
i Ixz Moment of inertia about FBxz 0.0202 kg-m2
{ pq(Ixz ) + qr(Iyy Izz )}Ixz + ⌧z (Ixx ) ⌧x (Ixz ) , KF Aerodynamic force coefficient 6.85 ⇥ 10 6 N-s2
K⌧ Aerodynamic drag-moment coefficient 3.35 ⇥ 10 7 Nm-s2
(5c)
numerical study, the effects of unmodelled uncertainties and uref
k , respectively; the terminal state reference is represented
measurement errors in the form of process and measurement by xref
N c ; Wx 2 R
nx ⇥nx
, Wu 2 Rnu ⇥nu and WNc 2 Rnx ⇥nx
noise, respectively, are included in the final model: are the corresponding weight matrices, which are assumed
constant in this study. Furthermore, xk,min  xk,max 2 Rnx
xk+1 = fd (xk , uk , p) + wk , (9)
and uk,min  uk,max 2 Rnu , specify the lower and upper
zk = h(xk , uk ) + ⌫k , (10) bounds on the states and control inputs, respectively.
where the state vector x 2 Rnx , control vector u 2 Rnu and 1) High-level NMPC for Position Tracking: Utilizing the
parameter vector p 2 Rnp are composed of: feedback of other states at current time instant, the H-NMPC
computes the optimized solution in terms of total thrust
x = [x, y, z, u, v, w, , ✓, , p, q, r]T , (11) and attitude angles, required to follow the given position
u = [⌦1 , ⌦2 , ⌦3 , µ] , T
(12) trajectory, as can be visualized in Fig. 2. Therefore, the state
and control vectors for H-NMPC are composed of:
p = [KF , K⌧ ]T , (13)
T
xH-NMPC = [x, y, z, u, v, w] , (15a)
and z 2 Rnx is the measurement vector. The state and ⇤ ⇤ ⇤ T
measurement functions are denoted by fd (·, ·, ·):Rnx ⇥Rnu ⇥ uH-NMPC = [ ,✓ , , Fz⇤ ] . (15b)
Rnp ! Rnx and h(·, ·):Rnx ⇥ Rnu ! Rnx , respectively. In addition, the following state and control reference trajec-
Also, the process and measurement noise are defined by the tories are selected for the optimization problem of H-NMPC:
covariance matrices Q and R, respectively:
xref ref T
w ⇡ (0, Q), ⌫ ⇡ (0, R), E[w⌫ T ] = 0. H-NMPC = xNc ,H-NMPC = [xr , yr , zr , 0, 0, 0] , (16a)
uref
T
H-NMPC = [ 0.0414, 0, 0, 0.5mg] . (16b)
III. R ECEDING H ORIZON C ONTROL A ND E STIMATION
A. Nonlinear Model Predictive Control One may note that for the OCP formulation of H-NMPC,
the parameterization of the nonlinear model (in translation)
(N)MPC is an advanced, optimization-based strategy for
is done with respect to p, q and r. That is why, the three
feedback control which computes an optimum control action
rotational rates are fed to H-NMPC along with the other
by optimizing the predicted system’s behaviour over a finite
states, as also illustrated in Fig. 2.
window, often recalled as prediction horizon (Nc ). This
In order to achieve a stable response from the low-level
optimal forecast of the system’s behaviour is induced from
controller, following constraints are imposed in H-NMPC:
an open-loop online optimization which is expressed in the

form of a constrained, finite horizon, optimal control problem 45 ( )   45 ( ), (17a)
(OCP). The open-loop OCP, when solved for the current state 45 ( )  ✓  45 ( ) ⇤
(17b)
of the system and at the current time instant, gives an optimal 0.01mg (N)  Fz⇤  mg (N). (17c)
sequence of open-loop control actions; the first term of which
is then regarded as the optimal control action [13]. Since the Also, its weight matrices are selected as:
prediction window recedes forward in time, (N)MPC is also Wx,H-NMPC = diag(35, 35, 20, 3.5, 5, 3), (18a)
referred as the receding horizon control technique.
Wu,H-NMPC = diag(30, 30, 50, 0.005), (18b)
Similar to other tracking applications [14], we formulate
the discrete time parametric OCP for NMPC in the form WNc ,H-NMPC = diag(50, 50, 30, 4, 6, 4), (18c)
of a least square function that penalizes the deviations based on trial-and-error method, such that, a smooth closed-
of predicted state (xk ) and control (uk ) trajectories from loop response of the UAV is achieved. Furthermore, the
their specified references, over the given prediction horizon prediction window Nc = 30, is selected for the H-NMPC.
window (tj  t  tj+Nc ): Remark 1: One may notice a control allocation block after
⇢ j+N
X c 1⇣ ⌘ the H-NMPC in Fig. 2. It computes the rotor velocities that
1 2 2
min xk xrefk Wx + uk uref
k Wu +
are required to climb/descent, based upon the total thrust
xk ,uk 2 commanded by the H-NMPC:
k=j
r r s
2
xNc xref
Nc WNc
(14a) ⇤
⌦1 =
Fz⇤ ⇤
, ⌦2 =
Fz⇤ ⇤
, ⌦3 =
Fz⇤
. (19)
3KF 3KF 3KF cos(µ)
s.t. xj = x̂j , (14b)
xk+1 = fd (xk , uk , p), k = j, · · · , j + Nc 1, Process and
measurement
(14c) 𝐹𝑧∗ Control Ω1∗ , Ω∗2 ,Ω∗3 𝑇 noise
𝑥𝑟
xk,min  xk  xk,max , k = j, · · · , j + Nc , (14d) 𝑦𝑟
High-Level Allocation
Tilt-rotor
Controller + 𝑇
𝑧𝑟
Low-Level + Ω1 , Ω2 ,Ω3 ,𝜇
uk,min  uk  uk,max , k = j, · · · , j + Nc 1, (H-NMPC) 𝜑 ∗ ,𝜃 ∗ ,𝜓 ∗ 𝑇 Controller + + + + 𝑇+ Tricopter
(L-NMPC) Ω1 , Ω2 ,Ω3 ,𝜇 Dynamics
(14e) 𝜑,𝜃,𝜓, 𝑝, 𝑞, 𝑟, 𝐾𝐹 , 𝐾𝜏 𝑇
(time-varying)
Estimator
where xk 2 Rnx is the differential state, uk 2 Rnu is the con- 𝑥,𝑦,𝑧, 𝑢, 𝑣, 𝑤, 𝑝, 𝑞, 𝑟 𝑇 (NMHE)

trol input, and x̂j 2 Rnx is the current state estimate; time-
varying state and control references are denoted by xref Fig. 2: Closed-loop control diagram.
k and
2) Low-level NMPC for Attitude Tracking: The optimized parameter (p̂) at time tj using the available measurements
solution of H-NMPC is subsequently passed to the L-NMPC, within the horizon, is of the form [15]:
which computes the optimized solution to an attitude tracking ⇢ 2 X j
problem directly in terms of actuator inputs. These actuator x̂L x̄L 2
min + kzk h(x̂k , uk )kV
inputs are given to the UAV after a summation with the inputs x̂k ,p̂ p̂ p̄L P
L k=L
coming from the H-NMPC, as can be visualized in Fig. 2. j 1
X 2
Therefore, the state, control and parameter vectors for L- + kwk kW (24a)
NMPC are composed of: k=L
T s.t. x̂k+1 = fd (x̂k , uk , p) + wk , k = L, · · · , j 1,
xL-NMPC = [ , ✓, , p, q, r] , (20a)
⇥ ⇤ (24b)
+ T
uL-NMPC = ⌦+ + +
1 , ⌦ 2 , ⌦3 , µ , (20b) x̂k,min  x̂k  x̂k,max , k = L, · · · , j, (24c)
T
pL-NMPC = [KF , K⌧ ] , (20c) p̂min  p̂  p̂max , (24d)
and the following state and control reference trajectories are where x̂k,min  x̂k,max and p̂min  p̂max specify the lower
selected for L-NMPC: and upper bounds on the estimated states and parameter,
respectively. Also, PL , V and W are the respective weight
xref ref ⇤ T
L-NMPC = xNc ,L-NMPC = [ , ✓⇤ , ⇤
, 0, 0, 0] , (21a) matrices that are usually interpreted as the inverse of covari-
uref
T
L-NMPC = [595, 605, 544, 0.173] , (21b) ance matrices [15]:
1 1 1
where ⌦’s are in rad/s and µ is in radian. Additionally, PL = Q 0 2 , V =R 2 , W =Q 2 , (25)
the following control constraints are imposed in L-NMPC, where Q0 is the initial covariance matrix (incorporating state
such that, even after a full actuation command from the L- and parameter, both), R is the measurement noise covariance
NMPC, the UAV still has some room to follow H-NMPC’s matrix and Q is the process noise covariance matrix. In
climb/descent command without saturating the rotors: addition, x̄L and p̄L denote the estimated state and parameter
0 (rad/s)  ⌦+ (22a) values (arrival cost data) at the start of estimation horizon,
1  700 (rad/s),
i.e. at tL [15].
0 (rad/s)  ⌦+  700 (rad/s), (22b)
2 As new measurement becomes available, the measurement
0 (rad/s)  ⌦+
3  700 (rad/s), (22c) present at the beginning of the estimation window (zL ) is
+
25 ( )  µ  25 ( ). (22d) dropped. In this way, the estimation window moves forward
in time, which gives this technique the name moving horizon
Also, after utilizing trial-and-error method, the L-NMPC is estimation. Another implication of this moving horizon is
designed with the following weight matrices: the boundedness of the estimation window size. In general,
Wx,H-NMPC = diag(25, 35, 22, 0.2, 0.2, 0.2), (23a) the performance of NMHE is dependent on the choice of
4 4 5
estimation window length M , which is in fact problem-
Wu,H-NMPC = diag(1 ⇥ 10 , 1.3 ⇥ 10 , 1.3 ⇥ 10 , 20), specific. It basically represents a trade-off between compu-
(23b) tational liability and estimation accuracy that simultaneously
WNc ,H-NMPC = diag(30, 40, 25, 1, 1, 1), (23c) grow with M . Therefore, for the fast dynamical systems
like UAVs, we cannot indefinitely increase M as limited
and with an equal prediction window as that for the H-
computational power is available on-board.
NMPC, i.e., Nc = 30.
NMHE for State and Parameter Estimation: For the
B. Nonlinear Moving Horizon Estimation considered tricopter model in this study, NMHE is designed
with the state, control and parameter vectors specified in (11),
Typically, (N)MHE is considered as a dual problem of (12) and (13), respectively. In addition, the weight matrices
(N)MPC as they both exploit the same optimization problem PL , V and W, are chosen to be:
structure [14], [15]. Therefore, in a similar manner, the
(N)MHE scheme is also formulated using a least square PL = diag(0.12 , 0.12 , 0.12 , 0.12 , 0.12 , 0.12 , 0.12 , 0.12 , 0.12 ,
function to penalize the deviation of estimated outputs 0.12 , 0.12 , 0.12 , 0.0322 , 0.0322 )
1
2 , (26a)
h(·, ·) from measurements (z). It includes an estimation
V = diag(0.032 , 0.032 , 0.032 , 0.045 , 0.045 , 0.0452 ,
2 2 2 2 2
horizon containing M measurements zL , · · · , zj , taken at
time tL < · · · < tj , where the length of the horizon is given 0.0222 , 0.0222 , 0.0222 , 0.2242 , 0.2242 , 0.2242 ,
def 1
by TE = tj tL , and j M + 1 = L is defined. As the 0.0032 , 0.0032 , 0.0032 , 0.0022 ) 2 , (26b)
performance of NMHE also relies on the availability of an 2 2 2 2
W = diag(0.1 , 0.1 , 0.14 , 0.14 , 0.14 , 0.14 , 0.14 , 2 2 2
accurate state model, a suitable component (arrival cost) is 1

included in the final optimization problem formulation of 0.142 , 0.142 , 0.172 , 0.172 , 0.172 , 0.892 , 0.892 ) 2 .
NMHE in order to prevent destabilization of the closed-loop (26c)
[15]. Finally, the discrete time dynamic optimization problem The above values of weight matrices are decided based upon
to estimate the constrained states (x̂) as well as the unknown the definition in (25) while exploiting the availability of the
true values of noise covariance matrices (Q and R). Also, Reference
Measurement
the arrival cost is initialized with the following state and 2.5 NMHE est.
parameter vectors:
2
6 7 T
x̄L = 010 , p̄L = [5 ⇥ 10 , 3 ⇥ 10 ] , (27)

z-axis (m)
1.5
where 010 represents a zero vector of length 10. Moreover, 1
the following constraints are imposed for achieving a con-
strained estimation of the unknown parameters: 0.5

0
4.11 ⇥ 10 6
(N-s2 )  KF  9.59 ⇥ 10 6
(N-s2 ), (28a)
1
2.47 ⇥ 10 7
(Nm-s2 )  K⌧  5.75 ⇥ 10 7
(Nm-s2 ). 0 1
(28b) -1 -1
0

y -axis (m) x -axis (m)


Furthermore, the estimation window length M is selected to
be 30, in order to impose a decent computational burden and 2

x -axis (m)
thus, facilitate the real-time applicability of this framework. 1

IV. N UMERICAL R ESULTS 0

In this section, we present the simulation results for the 0 10 20 30 40 50 60


proposed learning-based cascade NMPC which is employed 2
y -axis (m)
for trajectory tracking of a tilt-rotor tricopter UAV. Owing to 1
the similarity between the optimization problems of NMPC 0
and NMHE, defined in (14) and (24), they are solved using 0 10 20 30 40 50 60
the direct multiple shooting method and real-time iteration
(RTI) approach incorporated in ACADO toolkit. It is an
z-axis (m)

2
open-source C++ based software environment that creates 1
self-contained C codes, utilizing its code generation package
0
[4]. These C codes can be exported and executed on any 0 10 20 30 40 50 60
computational platform including low-cost embedded hard- Time (s)

ware. In our application, we utilize these C codes in order Fig. 3: Position tracking performance of the tilt-rotor tri-
to create MATLAB/Simulink based S-functions. Moreover, copter UAV for a square-shaped reference trajectory.
all the simulations are performed in MATLAB/Simulink
installed on a hardware consisting of 3.4-GHz Intel Core i7 The control outputs for the high- and low-level NMPCs
processor and 8 GB RAM. For initializing the simulations, are shown in Figs. 4 and 5, respectively. As evident, none of
the UAV is considered to be on the ground with zero state the constraints are violated by both H-NMPC and L-NMPC,
and control vectors, and with the parameters given in Table I. which in-turn supports the real-time applicability of the
A fixed sampling time of t = 0.01 s is maintained for the presented control framework. Additionally, the RMSE values
entire simulation. for L-NMPC in following the ⇤ , ✓⇤ and ⇤ commands of
In the subsequent part, we provide the position tracking H-NMPC are 3.92 , 5.74 and 2.32 , respectively.
results for a square-shaped reference trajectory. The mo- The results of parameter estimation by NMHE for square-
tivation behind incorporating this aggressive trajectory is like varying parameters are shown in Fig. 6. Owing to the
to further excite the existing nonlinear and highly coupled learning ability of NMHE, both the parameters are well
dynamics of the tilt-rotor tricopter and thus, examine the assessed, without any constraint violation, even when they
ability of cascade NMPC in handling it. In addition, NMHE are aggressively varying. In addition, one may appreciate
is employed to learn the change in UAV dynamics, which is the noise filtering ability of NMHE in Figs. 3 and 4, as it
explicitly induced by varying the aerodynamic parameters successfully caters for the present measurement noise.
KF and K⌧ by ±20% from their true values (listed in Lastly, the average execution times taken by NMHE,
Table I). H-NMPC and L-NMPC are 2.0 ms, 0.6 m and 0.7 ms,
respectively. Moreover, the combined average execution time
Square-shaped Trajectory Tracking of the closed-loop is approximately 3.3 ms, which is sub-
A discrete square-shaped trajectory of 1.5 m length is stantially less than the adopted sampling time of 10 ms,
given in terms of way points along x- and y-directions, after and thus, validates the real-time feasibility of the closed-loop
the UAV climbs to 1.5 m in z-direction. For the weights spec- framework.
ified in (18) and (23), the tracking performance of cascade
NMPC for the square-shaped trajectory can be visualized in V. C ONCLUSIONS
Fig. 3. In addition, the position tracking RMSE values for The learning-based cascade NMPC framework for the
the square-shaped reference along x-, y- and z-directions are trajectory tracking application of tilt-rotor tricopter UAV has
0.3494 m, 0.3071 m and 0.2848 m, respectively. been elaborated in this study. A time-varying dynamical
20
H-NMPC Constraints model is considered, on the top of already complex tricopter
F *z (N)

10 dynamics, in order to investigate the online learning capabil-


0 ity of NMHE. The results have illustrated a precise tracking
0 10 20 30 40 50 60 of the utilized square-shaped trajectory by the control frame-
100
H-NMPC Measurement NMHE Constraints
work. Thanks to RTI scheme of ACADO toolkit, a combined
execution time below 4 milliseconds is achieved that exhibit
* ( o)

50
0 the real-time feasibility of the presented closed-loop.
-50
0 10 20 30 40 50 60
ACKNOWLEDGMENT
100
The authors would like to acknowledge the support pro-
* ( o)

50
0
vided by Singapore Ministry of Education (RG185/17).
-50
0 10 20 30 40 50 60 R EFERENCES
10 [1] G. Garimella and M. Kobilarov, “Towards model-predictive control
for aerial pick-and-place,” in 2015 IEEE International Conference on
* ( o)

0 Robotics and Automation (ICRA), May 2015, pp. 4692–4697.


-10 [2] C. Papachristos, K. Alexis, and A. Tzes, “Dual–Authority Thrust–
Vectoring of a Tri–TiltRotor employing Model Predictive Control,”
0 10 20 30 40 50 60
Journal of Intelligent & Robotic Systems, vol. 81, no. 3, pp. 471–504,
Time (s)
2016.
Fig. 4: Controller outputs for H-NMPC with the actual [3] H. Seo, S. Kim, and H. J. Kim, “Aerial grasping of cylindrical object
using visual servoing based on stochastic model predictive control,”
response of the UAV. in 2017 IEEE International Conference on Robotics and Automation
(ICRA), May 2017, pp. 6362–6368.
(RPM)

6000
4000 [4] D. Ariens, B. Houska, H. Ferreau, and F. Logist, ACADO: Toolkit for
2000 L-NMPC Automatic Control and Dynamic Optimization, 1st ed., Optimization
+
1

0 in Engineering Center (OPTEC), 2010, http://www.acadotoolkit.org/.


0 10 20 30 40 50 60 [5] E. Kayacan, J. M. Peschel, and E. Kayacan, “Centralized, decentral-
ized and distributed nonlinear model predictive control of a tractor-
(RPM)

6000
4000 trailer system: A comparative study,” in 2016 American Control
2000 Conference (ACC), July 2016, pp. 4403–4408.
+
2

0 [6] J. Trachte, F. Gonzalez, and A. McFadyen, “Nonlinear model predic-


0 10 20 30 40 50 60
tive control for a multi-rotor with heavy slung load,” in International
(RPM)

6000 Conference on Unmanned Aircraft Systems (ICUAS), 2014, May 2014,


4000 pp. 1105–1110.
2000 [7] M. Neunert, C. de Crousaz, F. Furrer, M. Kamel, F. Farshidian,
+
3

0
0 10 20 30 40 50 60
R. Siegwart, and J. Buchli, “Fast nonlinear model predictive control
for unified trajectory optimization and tracking,” in 2016 IEEE Inter-
20 national Conference on Robotics and Automation (ICRA), May 2016,
(deg)

0 pp. 1398–1404.
[8] P. Bouffard, A. Aswani, and C. Tomlin, “Learning-based model predic-
+

-20
0 10 20 30 40 50 60 tive control on a quadrotor: Onboard implementation and experimental
Time (s) results,” in 2012 IEEE International Conference on Robotics and
Automation (ICRA), May 2012, pp. 279–284.
Fig. 5: Controller outputs for L-NMPC in terms of rotor [9] E. Kayacan, E. Kayacan, H. Ramon, and W. Saeys, “Learning in
RPMs and tilting angle for back rotor. centralized nonlinear model predictive control: Application to an
autonomous tractor-trailer system,” IEEE Transactions on Control
10 -6 Systems Technology, vol. 23, no. 1, pp. 197–205, Jan 2015.
12
NMHE True [10] M. Mehndiratta and E. Kayacan, “Receding horizon control of a 3
10 DOF helicopter using online estimation of aerodynamic parameters,”
8 Proceedings of the Institution of Mechanical Engineers, Part G:
Journal of Aerospace Engineering, 2017.
6 [11] S. Bouabdallah, “Design and control of quadrotors with application to
4 autonomous flying,” PhD dissertation, EPFL, 2007.
[12] B. Li, W. Zhou, J. Sun, C. Wen, and C. Chen, “Model predictive
0 10 20 30 40 50 60 control for path tracking of a VTOL tailsitter UAV in an HIL
10 -7 simulation environment,” in 2018 AIAA Modeling and Simulation
Technologies Conference. American Institute of Aeronautics and
6 Astronautics, 8-12 January 2018, p. 1919.
[13] U. Eren, A. Prach, B. B. Koçer, S. V. Raković, E. Kayacan, and
4 B. Açikmeşe, “Model predictive control in aerospace systems: Current
state and opportunities,” Journal of Guidance, Control, and Dynamics,
vol. 40, no. 7, pp. 1541–1566, 2017.
2 [14] M. Vukov, S. Gros, G. Horn, G. Frison, K. Geebelen, J. Jørgensen,
0 10 20 30 40 50 60 J. Swevers, and M. Diehl, “Real-time nonlinear MPC and MHE for
Time (s) a large-scale mechatronic application,” Control Engineering Practice,
vol. 45, pp. 64 – 78, 2015.
Fig. 6: True values of aerodynamic parameters and their [15] P. Kühl, M. Diehl, T. Kraus, J. P. Schlöder, and H. G. Bock, “A real-
estimation utilizing NMHE. time algorithm for moving horizon state and parameter estimation,”
Computers & Chemical Engineering, vol. 35, no. 1, pp. 71 – 83,
2011.

View publication stats

Вам также может понравиться