Gomaa Et Al - 2022 - Nonlinear MPC Without Terminal Costs or Constraints For Multi-Rotor Aerial

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

440 IEEE CONTROL SYSTEMS LETTERS, VOL.

6, 2022

Nonlinear MPC Without Terminal Costs or


Constraints for Multi-Rotor Aerial Vehicles
Mahmoud A. K. Gomaa , Graduate Student Member, IEEE, Oscar De Silva , Member, IEEE,
George K. I. Mann , and Raymond G. Gosine

Abstract—This letter proposes a novel NMPC for multi- alleviate this issue some have introduced a terminal cost and/or
rotor aerial vehicles which is designed without stabilizing terminal constraints to the optimization problem [6], [8] (for a
terminal costs or constraints in its cost function for sta- thorough review see [9]). The terminal constraint reduces the
bilization. A growth bound sequence is derived from a region of attraction, i.e., the feasible operating region of the
tailored running cost to ensure the closed-loop stability
and provide a measure of the performance of the proposed
MPC scheme (refer [10, Sec. 7.4]).
NMPC scheme. Furthermore, it facilitates the computa- Pioneering work reported in [11], [12] introduces a novel
tion of a stabilizing prediction horizon that guarantees method to design a finite optimal control problem without
the asymptotic stability of the system. The performance using the stabilizing terminal costs or constraints. This par-
of the proposed scheme is investigated through two sets ticular design provides a larger stability region and improved
of numerical simulations and compared against the tradi- closed-loop performance. This method adopts a Lyapunov
tional NMPC scheme for the application as proposed in function to represent the finite horizon value function without
(Kamel et al., 2017). The results show superior performance terminal conditions, which allows the stability to be implied by
of the proposed NMPC scheme in terms of tracking accu-
racy, convergence rate, and computation time. the monotonicity of the value function [12]. Grimm et al. [12]
showed that the asymptotic stability of the closed-loop system
Index Terms—Micro aerial vehicle, nonlinear model can be guaranteed for a sufficiently long prediction horizon.
predictive control, nonlinear systems, stability analysis. However, this initial study did not provide explicit bounds
for the length of the prediction horizon. A series of recent
studies [13]–[15] has ensured the closed-loop asymptotic sta-
I. I NTRODUCTION bility relying on the concept of cost controllability, i.e., local
ONLINEAR model predictive control (NMPC) has controllability condition (see [16], [17] for thorough details).
N received increased popularity for controlling fast dynamic
robotic devices, including aerial vehicles [1]–[3]. NMPC tends
This particular feature allows the NMPC to be implemented
at a faster rate for nonlinear robotic systems while maintain-
to be computationally demanding and implementing this con- ing closed-loop stability ([10, Sec. 7.4]). Continued with this
troller real-time on a dynamically fast system (e.g., robotic pioneering work, research work in [2], [18] effectively demon-
devices) is quite challenging. Generally, fast solvers such strated a real-time implementation of this method to stabilize
as CasADi [4] are required to be incorporated for efficient a ground robot system to operate fast on a given trajectory.
deployment of NMPC algorithms. Additionally, the closed- This ground robot uses the non-holonomic kinematic model,
loop stability of receding horizon problems is not guaranteed which is generally open-loop stable.
except for special cases such as when the prediction horizon The control of a quadrotor aerial vehicle is quite challeng-
tends to infinity under some controllability assumptions such ing mainly because the system is open-loop unstable, has 3D
as there exists a suitable upper bound on the optimal value geometric nonlinearity invoked by the rigid body dynamics,
function [5]. Having an infinite (or very large) prediction hori- and is under-actuated, i.e., it has only four control actions
zon is not practical due to the computational complexity, which (inputs) from the four rotors and six degrees of freedom state
increases exponentially with the length of the prediction hori- parameterizing the quadrotor pose in space. Applying constant
zon and the number of the manipulated variables [6], [7]. To motor inputs does not achieve stable flight, e.g., hovering,
thus sufficiently fast control actions are necessary to stabi-
Manuscript received March 4, 2021; revised April 12, 2021; accepted lize and also to maintain smooth flights [19]. For trajectory
April 27, 2021. Date of publication May 10, 2021; date of cur- control of these vehicles, the controller requires to perform
rent version June 25, 2021. This work was supported in part by all computations of the control algorithm in real-time using
the Natural Sciences and Engineering Research Council of Canada
(NSERC), and in part by Memorial University of Newfoundland, Canada.
resource-constrained onboard computers. Furthermore, the
Recommended by Senior Editor S. Tarbouriech. (Corresponding author: mathematical model describing the system dynamic ignores
Mahmoud A. K. Gomaa.) many other aero-dynamical interactions and therefore the con-
The authors are with the Intelligent Systems Lab, Memorial trol method should have sufficient robustness to compensate
University of Newfoundland, St. John’s, NL A1B 3R5, Canada
(e-mail: makamel@mun.ca; oscar.desilva@mun.ca; gmann@mun.ca;
the unmodelled dynamics of the MPC model [20]. Recent
rgosine@mun.ca). work in [21] explains an NMPC scheme without terminal
Digital Object Identifier 10.1109/LCSYS.2021.3078809 constraints. However, the controller contained a feedback
2475-1456 
c 2021 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: Xian Univ of Architecture and Tech. Downloaded on September 10,2021 at 01:29:30 UTC from IEEE Xplore. Restrictions apply.
GOMAA et al.: NONLINEAR MPC WITHOUT TERMINAL COSTS OR CONSTRAINTS FOR MULTI-ROTOR AERIAL VEHICLES 441

linearized system that requires the vehicle to be at an ini- with an optimal value function VN : X → R≥0 defined as
tial position closer to the equilibrium point in order for the
system to be stable. VN (xk ) := inf JN (xk , u) (3)
u∈U N (xk )
Motivated by the research work in [2], [13], [14], this let-
ter provides the stability analysis based design of a novel for N ∈ N∪{∞}. Solving the OCP while attaining the infimum
computationally-efficient NMPC scheme (with improved sta- in (3) induces an optimal control function u∗ = u∗ (·, xk ) with
bility characteristics) with a tailored objective function without VN (xk ) = JN (xk , u∗ ). The NMPC algorithm aims at computing
terminal costs or constraints for control of a quadrotor type a feedback law μN : X → U (defined as μN (k, x) = u∗ (0, xk ))
aerial vehicle. Finally, this letter provides numerical valida- such that for each xk ∈ X the resulting closed-loop trajectory
tion of the robustness of the proposed scheme at various initial xu (.; xk ) generated by system model (1), where xu (0; x0 ) = x0 ,
conditions and system configurations and shows that it outper- satisfies the constraints xu (k; x0 ) ∈ X and μN (xu (k; x0 )) ∈ U
forms the traditional NMPC scheme used in [1]. The work in for all k ∈ N0 , and is asymptotically stable.
[1] was considered in the comparative study as it proposes an
NMPC with terminal costs for the control of the full dynamics A. Stability of MPC Without Terminal Costs or
(translation and rotation) of the quadrotor. Constraints
Notations: Q, R, and N denote the sets of rational, real, Neither the asymptotic stability of system (1) nor the recur-
and natural numbers, respectively, and Q≥0 , R≥0 , and N0 := sive feasibility can be guaranteed under the NMPC without
N ∪ {0} are the sets of non-negative rational, real, and integer stabilizing conditions incorporated in its OCP. This section
numbers, respectively. A function ξ : R≥0 → R≥0 belongs to explains the main findings from [2], [13], [14] that ensure
class K if it is continuous and strictly increasing with ξ(0) = 0. the asymptotic stability by computing a stabilizing prediction
In addition, if ξ ∈ K is unbounded it is of class K∞ . A horizon N through deriving a growth bound sequence that
function ζ : R≥0 × R≥0 → R≥0 belongs to class KL if it is satisfies the cost controllability condition [16]. A system is
continuous and ζ (·, k) ∈ K∞ ∀k ∈ R≥0 and ζ (h, ·) is strictly asymptotic stable (e.g., at the origin) if there exists a function
monotonically decaying to zero for each h > 0. β ∈ KL such that the closed-loop trajectory xμN (k, x0 ) has
the following condition, i.e.,
II. MPC AND A SYMPTOTIC S TABILITY W ITHOUT  
xμ (k, x0 ) ≤ β( x0 , k), ∀k ∈ N0 , ∀x0 ∈ X. (4)
S TABILIZING C OSTS OR C ONSTRAINTS N

Consider the following discrete-time nonlinear system, Firstly, the following assumptions are formulated, which are
crucial for proving the asymptotic stability.
x+ = fδ (x, u), (1) Assumption 1: Assume that there exist.
where fδ : Rn × Rm → Rn is a discrete-time analytic mapping, A1.1: A monotonically increasing and bounded sequence
δ is the sampling time, x ∈ X ⊆ Rn is the state vector, x+ (γi )i∈N such that
is the next evolution of x, u ∈ U ⊆ Rm is the control input Vi (x0 ) ≤ γi x0 − xr 2Q ∀i ∈ N (5)
vector, n is the number of states, and m is the number of
control inputs. The objective is to drive the system from an holds for each x0 ∈ X, where x 2Q represents the quadratic
initial state x0 to a reference state xr using least amount of form xT Qx and Q ∈ Rn×n is the weight matrix of the state
control effort and minimum tracking error. For this purpose, error and selected such that Q 0.
the NMPC scheme is defined with a predefined running cost A1.2: Two functions α1 , α2 ∈ K∞ satisfying
(x, u) : Rn × Rm → R≥0 that uses the nonlinear model (1)
to compute a feedback control input. The admissibility of an α1 ( x − xr ) ≤ ∗ (x) ≤ α2 ( x − xr ) ∀x ∈ X (6)
input function is defined as follows. where, ∗ (x) := infu∈U 1 (x) (x, u).
Definition 1: For given admissible sets of states X and con- The performance bound of NMPC closed-loop can be
trol values U, and for initial states x0 ∈ X and N ∈ N, a defined as follows.
sequence of control values u = (u(0), u(1), . . . , u(N − 1)) ∈ Proposition 1: Let the closed-loop control sequence
UN is said to be admissible if the state trajectory μ (k, x) be admissible and given, and J∞ cl (x, μ ) :=
N∞ N
xu (·; x0 ) = (xu (0; x0 ), xu (1; x0 ), . . . , xu (N; x0 )), k=0 (xμN (k, x), μN (k, x)) be the closed-loop costs on the
infinite horizon. Then, the performance bound of the proposed
which is iteratively generated by the system model and pro- NMPC scheme can be defined as follows,
ceeded from xu (0; x0 ) = x0 , satisfies xu (k; x0 ) ∈ X ∀k ∈
{0, 1, . . . , N}. We denote this admissible control sequence for
cl
J∞ (x, μN ) ≤ αN−1 V∞ (x) (7)
x0 up to time N by u ∈ U N (x0 ). where, αN is the performance index, i.e., degree of
The NMPC algorithm solves iteratively an optimal control suboptimality.
problem (OCP) at each time instant with sampling time δ > 0 Hereinafter, the asymptotic stability is ensured by the
and prediction horizon N ∈ N≥2 . The solution of the OCP following theorem.
includes the minimization of a cost function JN : X × UN → Theorem 1: Consider an NMPC problem satisfying
R≥0 that sums up the running costs (x, u) along the predicted Assumption 1 and let the performance index αN be
trajectories. The cost function is defined as governed by

k+N−1 
JN (xk , u) := (xu (r, xk ), u(r)) (γN − 1) N k=2 (γk − 1)
(2) αN := 1 − N N . (8)
r=k k=2 γk − k=2 (γk − 1)

Authorized licensed use limited to: Xian Univ of Architecture and Tech. Downloaded on September 10,2021 at 01:29:30 UTC from IEEE Xplore. Restrictions apply.
442 IEEE CONTROL SYSTEMS LETTERS, VOL. 6, 2022

Kφ 1
φ̇(t) = φcmd (t) − φ(t)
τφ τφ
Kθ 1
θ̇ (t) = θcmd (t) − θ (t) (10)
τθ τθ
where, p and R := Rz (ψ) · Ry (θ ) · Rx (φ) are the vehicle posi-
tion vector and rotation matrix of the frame {B} respectively,
defined relative to {G}; v is the quadrotor velocity vector of
{B} relative to {G}; g is the gravitational acceleration; T is the
Fig. 1. Coordinate frames of the quadrotor system. mass normalized thrust; τφ and τθ are the time constants of
the inner-loop first-order model for the roll and pitch angles,
respectively; Kφ and Kθ are the gains of the same model;
and ē3 is the third standard basis vector. For simplicity, let
K
K̄φ = τφφ , τ̄φ = τ1φ , K̄θ = Kτθθ , and τ̄θ = τ1θ . Assuming
piecewise constant control inputs on each sampling interval
δ (in seconds), model (10) can be discretely expressed using
forward Euler integration as
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
x13 x46 03×1
⎢x ⎥ ⎢ −gē3 ⎥ ⎢Rē u ⎥
x+ = fδ (x, u) = ⎣ 46 ⎦ + δ · ⎣ + δ · ⎣ ¯ 3 1 ⎦ (11)
Fig. 2. Inner-outer loop control scheme for quadrotor system. ζ is the x7 −τ̄φ x7 ⎦ Kφ u2
Euler angles vector, ω is the angular velocity vector, Mφ , Mθ , and Mψ
are the applied torques for the three Euler angles, and the subscript r
x8 −τ̄θ x8 K¯θ u3
denotes the reference value.
where, x13 := p, x46 := v, x7 := φ, x8 := θ , u1 := T, u2 :=
φcmd , and u3 := θcmd . The state vector could be constrained
by ximin ≤ xi (k) ≤ ximax for all i = 1, . . . , 8 and k ∈ N0 . The
Then, if αN > 0, the relaxed Lyapunov inequality
control input constraints are defined as
VN (fδ (x, μN (x))) ≤ VN (x) − αN (x, μN (x)) (9) ⎧ ⎛ ⎞ ⎛ ⎞⎫
⎨ u1min u1max ⎬
holds for all x ∈ X and the NMPC closed-loop with horizon U := u ∈ Rm | ⎝−u2max ⎠ ≤ u(k) ≤ ⎝u2max ⎠ (12)
⎩ −u3max u3max ⎭
N is asymptotically stable. Additionally, Inequality (7) holds.
For the proof of Proposition 1 and Theorem 1, see [13] and
∀k ∈ N0 , where 0 < u1min < u1max , and u = [u1 , u2 , u3 ]T .
([10, Ch. 6]). The upper and lower bounds in A1.2 can be
simply guaranteed as given in [10], however, ensuring A1.1 is
not straightforward since the derivation of the growth bound B. Derivation of the Growth Bound
γi , i ∈ N is generally sophisticated as shown in Section III-B. Without lossof generality, the desired
T state is chosen as the
hovering xr = xr,1 xr,2 xr,3 01×5 , where 01×5 is a 1 × 5
III. S TABILITY A NALYSIS OF THE Q UADROTOR V EHICLE zero matrix. For this purpose, the following running cost is
AND THE G ROWTH F UNCTION
proposed, which is the key contribution in our work to derive
the growth bound in (5) and ensure the stability.
A. Quadrotor Vehicle Dynamics  2
Fig. 1 shows the body frame trajectory {B} attached to the (x, u) = x − xr 2Q + u − ũR (13)
u
center of gravity of a quadrotor with reference to a global
The first term penalizes the tracking error while the second
inertial frame {G}. The frame {B} orientation can be defined
term is a penalty related to the control inputs. Where, Q and
using Euler angles, roll (φ), pitch (θ ) and yaw (ψ), since the
Ru are the weight matrices of the tracking error and control
pitch and roll angles in this work are limited to zones which
actions, respectively. Expanding (13) yields the following
do not trigger the singularity in the system. The inner-outer
loop cascade control structure [22] shown in Fig. 2 is adopted 
8
 2
for this work. The inner control loop (low-level attitude con- (x, u) = qi xi − xr,i + r1 (u1 − ũ1 )2
troller) is modelled as a first-order system for the motor control i=1
as proposed in [1]. The inputs for the inner control loop are + r2 (u2 − ũ2 )2 + r3 (u3 − ũ3 )2 (14)
the commanded roll (φcmd ) and pitch (θcmd ) and the reference
yaw (ψr ) angles, and this loop generates the necessary control which is designed to satisfy the asymptotic stability conditions
inputs (torques) for the propeller motors. The outer trajec- presented in Theorem 2, where
tory tracking control loop receives the feedback information to g τ̄φ τ̄θ
generate the feed-forward command signals for the inner-loop ũ1 = , ũ2 = x7 , ũ3 = x8 ,
cx7 cx8 K̄φ K̄θ
controller. The first-order model sufficiently represents the
behavior of the inner loop and its parameters can be identified cx7 := cos(x7 ), and cx8 := cos(x8 ).
through classical system identification techniques [1]. Using (10) and (14) it can be found that (u1 − ũ1 ) penalizes
The nonlinear model of the quadrotor is defined as follows: the acceleration along ZB , (u2 − ũ2 ) penalizes the roll rate,
(u3 − ũ3 ) penalizes pitch rate, qi ∈ R≥0 , i = 1, 2, . . . , 8 are
ṗ(t) = v(t) the ith diagonal elements of the matrix Q, r1 , r2 , and r3 ∈ R≥0
v̇(t) = Rē3 T − gē3 are the ith diagonal elements of the matrix Ru .

Authorized licensed use limited to: Xian Univ of Architecture and Tech. Downloaded on September 10,2021 at 01:29:30 UTC from IEEE Xplore. Restrictions apply.
GOMAA et al.: NONLINEAR MPC WITHOUT TERMINAL COSTS OR CONSTRAINTS FOR MULTI-ROTOR AERIAL VEHICLES 443

The growth bound γi given in Assumption 1 can be obtained The second control input u2 was given in (11) as
by constructing a summable sequence cj ⊆ R≥0 , j ∈ N0 ,
 i−1 x7 [j + 1] = (1 − δ τ̄φ )x7 [j] + δ K¯φ u2 [j].
where ∞ j=0 cj < ∞, such that γi = j=0 cj , i ∈ N≥2 . For an
admissible control action ux0 = ux0 (j), j ∈ N0 , the sequence
cj satisfies the following inequality Thus, u2 is calculated as follows:
 ρ   ρ 
(xux0 (j; x0 ), ux0 (j)) ≤ cj x0 − xr 2Q λ − (j + 1)ρ λ − jρ
(15) x0,7 − (1 − δ τ̄φ ) x0,7 = δ K¯φ u2
λρ λρ
for all j ∈ N0 and x0 ∈ X, see [13] for more details. The
computation of the sequence cj and thus the growth bound γi The above is simplified to obtain u2 as follows:
is governed by the following theorem.
Theorem 2: Consider the system model (11) and running −Cρ + δ τ̄φ (λρ − jρ )
u2 = x0,7 . (21)
costs (14). Let the penalty parameters in (14) are given as δ K¯φ λρ
σ
r1 ≤ q6 , r2 ≤ σ K̄φ2 q7 , r3 ≤ σ K̄θ2 q8 (16) The third input u3 is obtained similar to u2 as
16
with weighting ratio σ ∈ Q, and let θ and φ ∈ [−60◦ , 60◦ ]. −Cρ + δ τ̄θ (λρ − jρ )
u3 = x0,8 . (22)
Then, condition (5) holds with γi = j=0 i−1
cj , i ∈ N≥2 , where δ K¯θ λρ
sequence cj is governed by
⎡ Applying (18), (20), (21), and (22) into (14) yields the running
⎛ ⎞2 ⎤
 ρ ρ 2 ρ−1

costs (23) along the resulting open-loop trajectories.
⎢ λ −j σ ρ ⎠ ⎥
cj := ⎣ + 2 2⎝ Ci ⎦ (17)
λ ρ δ λ (xux0 (j; x0 ), ux0 (j))
i=0
 
  8 Cρ 2
for j ∈ {0, 1, . . . , λ − 1}, where λ ∈ N≥2 is the number of λρ − jρ 2   2 δλρ
= qi x0,i − xr,i + r1 x2
steps required to perform any given maneuver, the exponent λρ c2 x7 c2 x8 0,6
ρ ∈ Q≥0 is adjusted for different trajectory shapes, and ρ Ci = 
i=1
2  2
ρ! −Cρ −Cρ
i!(ρ−i)! . Also, there exists a prediction horizon N ∈ N such + r2 2
x0,7 + r3 2
x0,8 (23)
that condition (8) holds and the NMPC closed-loop with N is δ K¯φ λρ δ K¯θ λρ
asymptotically stable.
The quadrotor trajectories can be chosen as straight lines, To this end, the bounding sequence cj can be found by
curves, or lattice shapes [23] based on the initial and final bounding the running costs (23) such that condition (15) is
states, prediction horizon, and sampling time. Therefore, for satisfied. We use the limit on the attitude angles as θ and φ ∈
each trajectory the initial state x0 ∈ X to the reference state [−60◦ , 60◦ ] as this bound captures the nominal non-aggressive
xr can be defined using a function given by; flight trajectories of the quadrotors. As a result, the second
 ρ   ρ term in (23) can be bounded as
λ − jρ j
x[j] = ρ
x0 + xr  2  ρ−1 
λ λρ Cρ
ρ i 2
 ρ δλρ 16 i=0 Ci j
j r1 x ≤ r1 2
2 2
x0,6
= x0 + (xr − x0 ), ∀j ∈ {0, 1, .., λ − 1}. (18) cos2 (x7 ) cos2 (x8 ) 0,6 δ λρ
λρ
Substitute (11) into (18) in order to find the required open- where, 1 cos2 (x7 ) cos2 (x8 ) has an upper bound of 16. Using
loop control inputs (ux0 ) for the maneuver. The first control condition (16) and recalling that j ≤ λ yield
input u1 was given in (11) as
C  ρ−1 2
x6 [j + 1] = x6 [j] − δg + δR33 u1 [j] ( δλρρ )2 16 λρ−1 i=0 ρ Ci
r1 x ≤ r1 2
2 2
x0,6
where, R33 = cx7 cx8 is the element in the third row and cos2 (x7 ) cos2 (x8 ) 0,6 δ λρ
third column in the rotation matrix R. Thus, the control input  2
ρ−1 ρ
u1x0 for all x0 ∈ X (or u1 for the sake of simplicity) can be i=0 Ci
≤ σ q6 2
x0,6
calculated as follows: δ 2 λ2
 ρ   ρ 
λ − (j + 1)ρ λ − jρ Moreover, the third and fourth terms in (23) can be bounded
x0,6 − x0,6
λρ λρ in the same manner as follows:
= −δg + δR33 u1 . (19)  2
  ρ−1 ρ
Using the binomial expansion, (19) reduces to −Cρ 2 2 1 i=0 Ci
 ρ−1  r2 x0,7 ≤ r2 2
x0,7
δ K¯φ λρ δ 2 K̄φ2 λ2
− i=0 ρ Ci ji  2
x0,6 + g = R33 u1 . ρ−1 ρ
δλρ i=0 Ci
ρ−1 ≤ σ q7 2
x0,7
Let Cρ = i=0 ρ Ci ji that yields, δ 2 λ2
 2
 2 ρ−1 ρ
C
C
g − δλρn x0,6 −Cρ i=0 i
u1 = . r3 2
x0,8 ≤ σ q8 2
x0,8
(20) ¯
δ Kθ λ ρ δ 2 λ2
cx7 cx8

Authorized licensed use limited to: Xian Univ of Architecture and Tech. Downloaded on September 10,2021 at 01:29:30 UTC from IEEE Xplore. Restrictions apply.
444 IEEE CONTROL SYSTEMS LETTERS, VOL. 6, 2022

As a result, the running costs (23) can be estimated by


 ρ 
λ − jρ 2  x0 − xf 2

(xux0 (j; x0 ), ux0 (j)) ≤ ρ
λ Q
⎛ ⎞2
ρ−1
 8
σ
+ 2 2⎝ ρ ⎠
Ci qi x0,i
δ λ
i=0 i=6
⎡ ⎛ ⎞2 ⎤
 ρ 2 ρ−1

⎢ λ −j
ρ σ ⎝ ρ ⎠ ⎥
 2
≤⎣ + Ci ⎦ x0 − xf
 (24)
λρ δ 2 λ2 Q
i=0

Fig. 3. (a) 3D trajectories of six simulation runs starting from various


Therefore, the bounding sequence cj in (15) can be attained
 initial positions, S1, S2, . . . , S6, and stabilizing at the final position (E).
as in (17). Finally, the growth bound γk := k−1j=0 cj , k ∈ N0 (b) Evolution of the value function.
can be obtained as given in Theorem 2 by
⎡ ⎛ ⎞2 ⎤

 ⎢ λ ρ − jρ
k−1  2 ρ−1

TABLE I
σ ρ ⎠ ⎥ C OMPARISON R ESULTS . E I S THE RMS E RROR IN MM . #I T. I S THE
γk = ⎣ ρ
+ 2 2⎝ Ci ⎦. (25) N UMBER OF I TERATIONS R EQUIRED TO S TABILIZE THE Q UADROTOR .
λ δ λ TA AND TM A RE THE AVERAGE AND M AXIMUM C OMPUTATION T IME OF
j=0 i=0
THE OCP AT E ACH I TERATION IN MS , R ESPECTIVELY

For the sake of simplicity and for better tracking, without


loss of generality, assume ρ = 1, that yields

1 ! 2 σ"
k−1 Iᴛ. Iᴛ.
γk = λ − 2λj + j 2
+
λ2 δ2
j=0
⎡ ⎤
1 ⎣ 2 
k−1 k−1
σ
= 2 j − 2λ2 j + (λ2 + 2 )k⎦
λ δ
j=0 j=0
k−1 k−1
while expanding j=0 j and j=0 j2 yields
The symbolic toolbox CasADi [4] is used to set up the
# $ OCP due to its high efficiency in implementing nonlin-
1 k3 k2 k σ
γk = 2 kλ2 − (k2 − k)λ + ( − + + 2 k) . ear optimizations. The direct multiple-shooting discretization
λ 3 2 6 δ method is employed to turn the OCP into a nonlinear program-
Feasible motion primitives of the quadrotor that guide the ming problem (NLP) and IPOPT is used as the solver. The
quadrotor from some initial state (position, velocity, and orien- dynamic model (11) is used for the state prediction required in
tation) to desired translational variables, e.g., position, velocity the solution of the OCP. The system parameters in system (10)
and acceleration, can be generated in the quadrotor’s jerk [24]. are given as; mass= 0.645 kg, τφ = 0.0914 s, Kφ = 0.7551,
Thus, in order to recover the thrust and attitude rates inputs τθ = 0.0984 s, and Kθ = 0.7226. The control inputs given
from this thrice differentiable trajectory, a prediction horizon in (12) are constrained by; u1min = 5 m/s2 , u1max = 15 m/s2 ,
of 4 is required, i.e., N ≥ 4. and u2max = u3max = 10◦ .
It can be noticed from (25) that the growth bound γk The simulation was conducted for six runs  to stabilize
T the
depends only on the sampling time δ, the exponent value ρ, quadrotor at the reference point xr = 0 0 1 01×5 from
the number of steps λ, and the weight ratio σ . Therefore, for various initial positions. The simulation terminates when the
a maneuver length λ ≥ N ≥ 4 and a weight ratio σ = δ 2 , the stopping criterion x(k) − xr ≤  is met for all k ∈ N, where
performance index given in (8) is αN ∈ [0, 1) for all δ > 0 and  > 0 is the error threshold and selected as  = 1 mm, oth-
the system is asymptotically stable. In addition, the NMPC erwise, it terminates after 30 s. The prediction horizon was
algorithm is recursive feasible since Assumption 1 and the selected as given in [1] as N = 20 and the sampling time was
conditions needed for the stability are met [25], [26]. selected in the range of δ ∈ [0.01, 1] s.
The trajectories of the six runs are shown in Fig. 3(a). The
results show that the quadrotor may go from the initial position
IV. R ESULTS to the final position in straight lines, curves, or lattice shape as
This section presents the MATLAB numerical validation proposed in the proof of Theorem 2. The results show that the
for several point stabilization (hovering) and trajectory track- proposed NMPC algorithm is able to asymptotically stabilize
ing problems, and the comparison with the traditional NMPC the quadrotor from arbitrary initial states. Fig. 3(b) shows that
scheme that incorporates terminal costs [1] and given by the evolution of the value function is monotonically decreasing
with time for all initial positions, which leads to the conclusion

N
JN (xk , u) = ( x − xr 2Q + u − ur 2Ru ) + xN − xr 2T that conditions of Theorem 1 are successfully met and the
system is asymptotically stable.
k=1
The comparison results are given in Table I. It shows that
where, T ∈ Rn×n is the terminal weight matrix and is the proposed NMPC has asymptotically stabilized the quadro-
computed by solving the Algebraic Riccati Equation. tor at the reference point while achieving stopping criteria in

Authorized licensed use limited to: Xian Univ of Architecture and Tech. Downloaded on September 10,2021 at 01:29:30 UTC from IEEE Xplore. Restrictions apply.
GOMAA et al.: NONLINEAR MPC WITHOUT TERMINAL COSTS OR CONSTRAINTS FOR MULTI-ROTOR AERIAL VEHICLES 445

operating system,” in Robot Operating System. Cham, Switzerland:


Springer, 2017, pp. 3–39.
[2] M. W. Mehrez, K. Worthmann, J. P. V. Cenerini, M. Osman,
W. W. Melek, and S. Jeon, “Model predictive control without terminal
constraints or costs for holonomic mobile robots,” Robot. Auton. Syst.,
vol. 127, May 2020, Art. no. 103468.
[3] I. Batkovic, U. Rosolia, M. Zanon, and P. Falcone, “A robust scenario
MPC approach for uncertain multi-modal obstacles,” IEEE Control Syst.
Lett., vol. 5, no. 3, pp. 947–952, Jul. 2021.
[4] J. A. E. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and M. Diehl,
“CasADi—A software framework for nonlinear optimization and
optimal control,” Math. Prog. Comput., vol. 11, pp. 1–36, Mar. 2019.
[5] M. A. Müller and K. Worthmann, “Quadratic costs do not always work
in MPC,” Automatica, vol. 82, pp. 269–277, Aug. 2017.
Fig. 4. Point stabilization performance of the proposed and traditional
[6] R. Bitmead, M. Gevers, and V. Wertz, Adaptive Optimal Control—The
NMPC at δ = 0.01 s.
Thinking Man’s GPC. London, U.K.: Prentice-Hall, Jan. 1990.
[7] D. Q. Mayne, “Model predictive control: Recent developments and
future promise,” Automatica, vol. 50, pp. 2967–2986, Dec. 2014.
[8] S. S. Keertht and E. G. Gilbert, “Optimal infinite-horizon feedback
laws for a general class of constrained discrete-time systems: Stability
and moving-horizon approximations,” J. Optim. Theory Appl., vol. 57,
pp. 265–293, May 1988.
[9] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert,
“Constrained model predictive control: Stability and optimality,”
Automatica, vol. 36, pp. 789–814, Jun. 2000.
[10] L. Grüne and J. Pannek, Nonlinear Model Predictive Control: Theory
and Algorithms. London, U.K.: Springer, Jan. 2011.
[11] V. Nevistić and J. A. Primbs, “Receding horizon quadratic optimal
control: Performance bounds for a finite horizon strategy,” in
Proc. European Control Conf. (ECC), Brussels, Belgium, 1997,
Fig. 5. Trajectory tracking performance of the proposed and traditional pp. 3584–3589.
NMPC schemes at (a) δ = 0.01 s. (b) δ = 0.1 s. [12] G. Grimm, M. J. Messina, S. E. Tuna, and A. R. Teel, “Model predictive
control: For want of a local control Lyapunov function, all is not lost,”
IEEE Trans. Autom. Control, vol. 50, no. 5, pp. 546–558, May 2005.
finite number of time steps at various sampling time δ with- [13] L. Grüne, “Analysis and design of unconstrained nonlinear MPC
out updating the running cost (14) weighting parameters. In schemes for finite and infinite dimensional systems,” SIAM J. Control
contrast, the traditional NMPC failed to stabilize the quadro- Optim., vol. 48, no. 2, pp. 1206–1228, 2009.
tor criteria or achieve the stopping criteria at small sampling [14] K. Worthmann, “Stability analysis of unconstrained receding hori-
zon control schemes,” Ph.D. dissertation, Dept. Mathematics, Bayreuth
time, and requires larger convergence time than the proposed Univ., Bayreuth, Germany, 2011.
NMPC. One sample of the comparison results is illustrated [15] K. Worthmann, “Estimates of the prediction horizon length in MPC: A
in Fig. 4. It shows that the proposed NMPC has successfully numerical case study,” IFAC Proc. Vol., vol. 45, no. 17, pp. 232–237,
2012.
stabilized the quadrotor with faster convergence rate and less [16] J.-M. Coron, L. Grüne, and K. Worthmann, “Model predictive control,
value function. Moreover, Table I and Fig. 5 show the trajec- cost controllability, and homogeneity,” SIAM J. Control Optim., vol. 58,
tory tracking performance of a circular trajectory. It can be no. 5, pp. 2979–2996, 2020.
observed that the proposed NMPC is robust against changing [17] S. E. Tuna, M. J. Messina, and A. R. Teel, “Shorter horizons
for model predictive control,” in Proc. Amer. Control Conf., 2006,
the controller parameters, e.g., sampling time and can track pp. 863–868.
fast trajectories with lower computation time. [18] K. Worthmann, M. W. Mehrez, M. Zanon, G. K. I. Mann, R. G. Gosine,
and M. Diehl, “Model predictive control of nonholonomic mobile robots
without stabilizing constraints and costs,” IEEE Trans. Control Syst.
V. C ONCLUSION Technol., vol. 24, no. 4, pp. 1394–1406, Jul. 2016.
[19] M. Faessler, “Quadrotor control for accurate agile flight,” Ph.D. disser-
From this study it can be concluded that implementing the tation, Dept. Informatics, Univ. Zurich, Zürich, Switzerland, 2018.
traditional NMPC for a highly nonlinear and unstable system, [20] V. Kumar and N. Michael, “Opportunities and challenges with
such as a quadrotor, using a finite prediction horizon cannot autonomous micro aerial vehicles,” in Robotics Research. Cham,
guarantee stability at certain system configurations, e.g., small Switzerland: Springer, 2017, pp. 41–58.
[21] N. T. Nguyen, I. Prodan, and L. Lefèvre, “Stability guarantees for trans-
sampling time. From the simulation results it is clear that the lational thrust-propelled vehicles dynamics through NMPC designs,”
traditional NMPC can be stabilized only if the prediction hori- IEEE Trans. Control Syst. Technol., vol. 29, no. 1, pp. 207–219,
zon is set to a larger value. However, for quadrotor vehicles Jan. 2021.
the sampling time interval is limited and requires faster execu- [22] N. Cao and A. F. Lynch, “Inner–outer loop control for quadrotor UAVs
with input and state constraints,” IEEE Trans. Contol Syst. Technol.,
tion with minimum computational cost. The proposed NMPC vol. 24, no. 5, pp. 1797–1804, Sep. 2016.
provides a guaranteed stability for a variety of common tra- [23] O. Andersson, O. Ljungqvist, M. Tiger, D. Axehill, and F. Heintz,
jectories, and most importantly it has the ability to stabilize “Receding-horizon lattice-based motion planning with dynamic obstacle
the system at small sampling time intervals. This work facil- avoidance,” in Proc. IEEE Conf. Decis. Control, Miami, FL, USA, 2018,
pp. 4467–4474.
itated the computation of a stabilizing prediction horizon and [24] M. W. Mueller, M. Hehn, and R. D’Andrea, “A computationally efficient
its corresponding performance index through a derived growth motion primitive for quadrocopter trajectory generation,” IEEE Trans.
bound for a tailored running cost. Robot., vol. 31, no. 6, pp. 1294–1310, Dec. 2015.
[25] L. Grüne, “NMPC without terminal constraints,” IFAC Proc. Vol.,
vol. 45, no. 17, pp. 1–13, 2012.
R EFERENCES [26] W. Esterhuizen, K. Worthmann, and S. Streif, “Recursive feasibil-
ity of continuous-time model predictive control without stabilising
[1] M. Kamel, T. Stastny, K. Alexis, and R. Siegwart, “Model predictive constraints,” IEEE Control Syst. Lett., vol. 5, no. 1, pp. 265–270,
control for trajectory tracking of unmanned aerial vehicles using robot Jan. 2021.

Authorized licensed use limited to: Xian Univ of Architecture and Tech. Downloaded on September 10,2021 at 01:29:30 UTC from IEEE Xplore. Restrictions apply.

You might also like