Professional Documents
Culture Documents
Manuscript
Manuscript
LICENSE
CC BY 4.0
19-07-2022 / 22-07-2022
CITATION
Khan, Subhan; Guivant, Jose (2022): Design and implementation of proximal planning and control of an
unmanned ground vehicle in the dynamic environment. TechRxiv. Preprint.
https://doi.org/10.36227/techrxiv.20336373.v1
DOI
10.36227/techrxiv.20336373.v1
Design and implementation of proximal planning
and control of an unmanned ground vehicle in the
dynamic environment
Subhan Khan, Jose Guivant, Member, IEEE,
Abstract—This paper presents a novel proximal planning and area. Therefore, motion planning and control are critical part 18
control (PPC) formulation for an unmanned ground vehicle of the ground vehicles. 19
(UGV) affected by skidding and slip disturbances. The control Proximal algorithms are advantageous for resolving com- 20
approach also considers the presence of moving and static
obstacles in the context of operation. The PPC technique is plex optimisation problems, particularly those with non- 21
divided into three parts; first, a nonlinear model predictive smooth or composite objective functions. A proximal algo- 22
control (NMPC) based path-planner is designed to periodically rithm is one whose fundamental iterations involve the proximal 23
generate an updated feasible trajectory to prevent the vehicle operator of some function and whose evaluation involves 24
from colliding with other objects from start to the goal pose. solving a specialised optimisation issue that is often simpler 25
In particular, a proximal averaged Newton-type method for
optimal control (PANOC) is used to design NMPC. Second, than the original problem [5]. In optimization theory, numer- 26
the velocity commands are produced via evolutionary program- ous techniques have been used with MPC, For instance, in 27
ming (EP) based kinematic control (KC). Third, a dynamic [6], a decentralised control strategy based on solving two 28
control mechanism with an extended state observer (ESO) is convex quadratic programming (QP) problems is described to 29
introduced to calculate the impacts of unidentified bounded determine the order in which agents traverse the intersection. 30
perturbations. Finally, simulations are performed in the presence
of linear and nonlinear trajectories of moving obstacles (MO) and In [7], an MPC-based approach addressed a high-level time 31
static obstacles (SO) to verify the performance of the proposed slot allocation optimum control problem (OCP), with agent 32
scheme. Additionally, we have investigated and confirmed that controls dictated by a layered low-level OCP. In essence, 33
the proposed PPC could run in real-time on a CPU with limited each agent optimises a quadratic programme (QP) and two 34
resources. linear programmes (LP) and sends the results to the central 35
Index Terms—Evolutionary programming, Extended state ob- coordinator, who solves the high-level nonlinear problem 36
server, Nonlinear model predictive control, Proximal planning (NLP). Rear-end collision prevention is added in [8], which 37
and control, unmanned ground vehicle improves the platform’s performance. A similar approach is 38
3
U NMANNED Ground Vehicles (UGVs) have gained pop-
ularity over the past few decades in civilian, commercial,
and defense applications. Driving a UGV on rough terrain is
consensus-based control system selects vehicle controls at
the lower level by solving a convex optimisation problem
42
43
S. Khan (e-mail: Subhan.Khan@sydney.edu.au) and J. Guivant (e-mail: ing the last few decades. Problems with posture stabilisation 58
J.guivant@unsw.edu.au) are with the School of Mechanical and Manufac- and trajectory tracking were researched concerning vehicle 59
turing Engineering, University of New South Wales (UNSW) Sydney, NSW, control (such as [16–20]). The co-design problem of targeting 60
Australia. In addition, Subhan Khan is also with the school of Electrical and
Information Engineering at the University of Sydney. posture and tracking control of UGVs was also studied [21– 61
Corresponding Author: S. Khan 23]. MPC has also been used to handle physical constraints on 62
2
63 state and input [24–26]. Some control techniques have consid- this purpose, a robust model predictive control is proposed 121
64 ered the control at a dynamic level while not needing UGVs to suppress the effects of uncertainties in [37], a disturbance 122
65 velocities [27, 28]. Most recently, some of the existing works observers (DO) are used to estimate the external disturbances 123
66 on autonomous bulldozers are based on high-level planning. and uncertainties [38, 39]. A robust tracking scheme DO that 124
67 A high-level planner is compared with manual operation in handles slip-skid is presented in [40]. The drawbacks of DO 125
68 [29], where the automatic operation can outperform the manual can be compensated using Nonlinear DO (NDO) [41]. ESO- 126
69 one. An efficient path-planner for autonomous bulldozers in based observations are used to estimate both internal and 127
70 [30, 31] proved to be essential for industries. external disturbances [42]. To estimate lumped disturbances, a 128
71 The above-mentioned existing techniques are considered generalized ESO (GESO)-based tracking at the dynamic level 129
72 without skid-slip. For low-and medium-level control problems, is presented in [40]. In [43], neural networks were used to 130
73 this pure rolling assumption may not hold for bulldozing ap- perform robust tracking under the presence of ski-slip in a 131
74 plications where the environment is uncontrolled. In addition, dynamic environment. Finally, the tracking control problem 132
75 safety hazards might occur in a real application. A solid is solved based on hybrid NDO and extended Kalman filter 133
76 investigation in controlling an UGV is required to mitigate any (EKF) presented in [44]. The proposed technique solves the 134
77 safety hazard. The ability of MPC to handle constraints and low-level control. However, it does not consider the skid- 135
78 predict the platform’s future response makes it an ideal choice slip presence model, which may still affect the closed-loop 136
80 have used MPC in the presence of wheel slip. For instance, In terms of motion planning of robotic systems, sampling 138
81 a min-max MPC for agriculture application [32], a traction based algorithms have significantly advanced the state of the 139
82 control problem on stabilizing wheeled mobile robot (WMR) art in robotic motion planning. For instance, When deter- 140
83 [33], a detailed analysis on controlling off-road vehicles [34], mining solution pathways, the planner takes uncertainty into 141
84 and a voltage control strategy based on MPC and nonlinear account, resulting in mobility that is safe in the presence of 142
85 disturbance observer (NDO) is implemented in [35]. For world model imperfections [45]. The planner directs sensing 143
86 industrial-grade applications, it is important to work without throughout the planning process in order to develop the world 144
87 colliding with a moving or static object. For this purpose, model. The experimental findings reveal that the suggested 145
88 bulldozing applications must be solved by considering the planning strategy enhances planning resilience in the presence 146
89 platform’s dynamics and kinematics, including skid-slip and of uncertainty greatly. In addition, in settings with dynamic 147
90 external disturbances. impediments, an integrated method combining the fast rapidly 148
91 A simulated technique using casadi-based ([15]) platform exploring random tree (RRT) algorithm and the configuration- 149
92 proposed in [20] deals with both skid-slip and obstacle avoid- time space may be employed to increase the quality of the 150
93 ance. They have also designed an NMPC-based planner, a planned trajectory and replanning [46]. An obstacle-cluttered 151
94 virtual kinematic controller, a saturated torque-based dynamic environment requires a search-based planning approach to 152
95 controller, and an extended state-observer. Although the results calculate dynamically possible routes for an autonomous robot. 153
96 are promising in this work, some issues require more investiga- This method explores the map using a collection of short- 154
97 tion. First, the prediction horizon of MPC used in [20] is large duration motion primitives to find smooth, minimum-time tra- 155
98 due to selecting the single-shooting method. In practice, the jectories [47].Furthermore, there are some numerical optimal 156
99 large prediction horizon could cause a slow processing time control based planners such as the multi-vehicle lane change 157
100 for the onboard processor in the industrial-grade application, motion planning job is a centralised optimum control issue, 158
101 which could be resolved with the multiple-shooting method. which is advantageous in that it is general and comprehensive 159
102 However, in real applications, it is important to investigate the [48], and a numerical technique for multiple-vehicle motion 160
103 processing time under different prediction horizons. Another planning that explicitly considers vehicle dynamics, temporal 161
104 solution to solve this problem is the proximal algorithms and geographical constraints, and energy needs [49]. 162
109 addition, actual velocities applied to the platform are not destination position safely in the context of moving and static 165
110 presented or compared with the reference ones, which is obstacles, skid-slip, and external disturbances in an isolated 166
111 important to investigate to confirm that the predictive planner’s environment. The main contributions of this research work 167
112 constraints meet the actual ones. In [12, 36], obstacles and are: 168
113 physical limitations can be handled; the fast NMPC in the 1) Implementation of a proximal algorithm-based NMPC 169
114 presence of PANOC is good enough for solving the posture path-planner that incorporates environmental impedi- 170
115 control problem. However, they have not tested PANOC for ments and generates workable reference trajectories. 171
116 bulldozing applications by considering the dynamics of the 2) A dynamic controller connects to a kinematic controller 172
117 platform. Additionally, external disturbances and skid-slip are tuned by evolutionary programming to produce velocity 173
118 not considered in the previous work on PANOC. vector, which controls the platform using torque. 174
119 In tracking control of UGVs, the platforms are always 3) An observer from the ESO to calculate the lumped 175
120 subject to different internal and external disturbances. For disturbances (skid-slip and external torque disturbances). 176
3
193 C. Outline where X consists the states of the platform, Ω is the velocity 204
Figure 1 shows the configuration of UGV, which consists Assumption 1: [40, 53]: M , H1 (X, Ẋ), H2 (X), and H3 (X) 217
of positions (x, y) and heading (θ) as the states, while are bounded but unknown. In addition, the nominal value M0 218
[v, ω]T represents the longitudinal velocity and yaw rate of the is unknown. 219
platform. In this work, we have used the following kinematics Assumption 2: [43, 52, 54]: The skid-slip parameters and
and the dynamic models provided in [51, 52]: torque disturbances are unknown but bounded. Additionally
the platform is subject to the following constraints:
Ẋ = S(X) · (Ω − η) + ρ(X, µ) (1)
M · (Ω̇ − η̇) + H1 (X, Ẋ) · (Ω − η) + H2 (X) · ρ̇(X, µ) Ωrmin ≤ Ωr (k) ≤ Ωrmax ,
(2) (3)
+ H3 (X) · ρ(X, µ) = B · (Sat(τ ) − τd ) armin ≤ ar (k) ≤ armax
where Ωr is the reference velocity vector and ar is the
with,
reference acceleration vector of the platform. The UGV has to
X = [x, y, θ]T ∈ R3 , Ω = [v, ω]T , η = [η1 , η2 ]T ∈ R2 , avoid dynamic and static obstacles to reach the goal location
r for the posture stabilisation control. Therefore, the following
τd = [τd1 , τd2 ]T ∈ R2 , η1 = (ηr + ηl ),
2 obstacle avoidance constraints can be formulated as:
r 0 −mdθ̇ q
η2 = (ηr − ηl ), H1 (X, ẋ) = , Doh (X) = (x − xho )2 + (y − yoh )2 , (4)
2b mdθ̇ 0
grh (X) = Doh (X) − (r + roh ) ≥ 0 (5)
m + 2J 2J
H2 (X) = r 2 cos θ (m + r 2 ) sin θ 0
2 , where h = 1, 2, 3, . . . is the number of obstacles in the
−md sin θ md cos θ J + 2br2J environment, Radii of the obstacles is ro . The obstacles have
−2J θ̇ sin θ 2J θ̇ cos θ xand y coordinates as xo and yo . The motion of UGV is
r2 r2 −mdθ̇
H3 (X) = ,
0 0 0 restricted to work in a domain defined as
T
T −µ sin θ X∈Ψ (6)
cos θ sin θ 0
S(X) = , ρ(X, µ) = µ cos θ ,
0 0 1 Thus, we can summarize the objective of this work is as 220
0
follows: Design a kinematic control law for Eq. (1) and 221
m 0 saturated torque control for Eq. (2) such that X can reach
M= , τ = [τ1 , τ2 ]T , ∈ R2 . 222
0 J − md2 the desired destination by respecting the constraints imposed 223
4
224 on platform (saturated torque, longitudinal velocity, acceler- Now, we are in a position to define the cost function of our
225 ation and yaw rate limitations given in Eq.(3) and collision point stabilisation problem as follows:
226 avoidance mentioned in Eq. (4) and Eq. (5).
arg min J(ar ) =
ar
227 III. PPC METHOD FOR UGV N
X −1
LN (XN ) + Lk (Xr (k + i), Ωr (k + i), ar (k + i))
228 In this section, first, we will discuss the design NMPC- i=0
229 based path-planner, which uses PANOC. Second, velocity- (9)
230 based control commands are generated by EP-based KC.
231 Finally, the dynamic controller, ESO, and stability analysis subject to following constraints
232 of the proposed scheme are dicussed.
Xr (k + i) = f1 (Xr (k + i − 1), Ωr (k + i − 1)), (10a)
Ωr (k + i) = f2 (Ωr (k + i − 1), ar (k + i − 1)), (10b)
233 A. NMPC-based path-planner
ak (k + i − 1) ∈ Uk , i ∈ N[0, N −1] , (10c)
Consider the platform operating in a context, with the
presence of dynamic and static objects. The UGV’s initial pose grh (Xk ) ≤ 0, i ∈ N[0, N ] (10d)
is X(0) and goal pose is Xg under physical limitations Eq. (4)
and Eq. (5). The reference posture and velocity can be denoted The stage cost: Lk : Rnp × Rnu → R could be taken as a 237
as Xr = [xr , yr , θr ]T and ωr = [vr , ωr ]T , respectively. We quadratic function in the form of L(Xr (k), Ωr (k), ar (k)) = 238
are considering the reference robot is pure rolling (excluding (Xr − Xg )T · Qk · (Xr − Xg ) + ΩTr · Q̄k · Ωr + aTr · R · ar . In 239
disturbances). So, our kinematics and dynamic models for the addition, the terminal cost LN : Rnx → R can be defined as 240
reference robot described in [53] LN (XN , ΩN ) = (XN −Xg )T ·QN ·(XN −Xg )+ΩTN Q̄N ΩN . 241
where ar = [v̇ , ω̇]T . Up until this point, we have assumed that gradient of the function L in Eq. 9 as shown in the Algorithm 246
the UGV operates in continuous time. However, a computer 1 obtained from [55], it allows the projected gradient iteration. 247
must produce discrete signals in order to function. Therefore, Now, we can use the adjoint method and convert it to
we may approximate our nonlinear continuous-time system produce forward-backward splitting (FBS) scheme:
into a discrete-time system as follows by applying Euler’s
approximation with a sampling time Ts : aν+1 = Tγ (aνr ) := proxγ (aνr − γ∇J(aνr )), (11)
r
1
Xr (k + 1) = Xr (k) + Ts · Ẋr (k) proxγ (ar ) = arg min kar − ξk2 . (12)
(8) ξ 2γ
Ωr (k + 1) = Ωr (k) + Ts · ar (k)
where aν+1
r consists of all accumulation points, γ-critical 248
234 where Xr (0) = X(0) and Ωr = Ωr (0) are taken as the
points always satisfy γ > 0, and ξ = γ∇J(aνr ). The 249
235 beginning conditions. The discrete indexing k ∈ Λ is a
fixed points Tγ could have a relationship with the following 250
collection of positive natural integers.
accumulation points as a∗r = Tγ (a∗r ). 251
(FBE) which is described in [13, 57] for Eq. 9 which can be Only the fitness function of a single population is necessary
defined by quadratic upper bound as follows: to understand evolutionary programming. Suppose P be the
population of all the parameters in Eq. (1), compensated state
Qγ J (ξ; ar ) = proxγ (ar ) + J(ar ) + ∇J(ar )T (ar − ξ) (15) gains are referred as k ∗ = [kˆx kˆy kˆθ ] ∈ R3 . The main goal of
ϕγ (ar ) = inf Qγ J (ξ, ar ), this exercise is to identify scaling factors kˆx , kˆy and kˆθ that
ξ∈U
γ are not ideal.
ϕγ (ar ) = L(ar ) − k∇J(ar )k2 (16)
vr cos(e3 ) + kˆx e1
2 v̂
ûc = c = (21)
+ dist2U (ar − γ∇J(ar )) ω̂c ωr + kˆy vr e2 + kˆθ vr sin(e3 )
253 The above expressions show the similar complexity with FBE The following formula can also be used to calculate the error
254 and forward-backward step. To solve the problem of Eq. (9), function for velocities ue = [ve , ωe ]T ∈ R2 :
255 we have employed PANOC, a fast nonlinear MPC introduced ve = −xe sin θ + ωye cos θ + v − vr cos θe (22)
256 in [13]. The advantage of PANOC over interior point method
257 or quadratic programming, is in terms of the processing time ωe = ω − ωr (23)
258 required at each iteration of the MPC process. Finding sub-optimal scaling factors for kˆx , kˆy and kˆθ is
The time derivative of tracking error of Eq. (18) is as follows of platform’s states and performance index of integral square 263
e˙1 −1 e2 vr cos(e3 ) ISE, and E is the shared policy for each process state, can be 265
e˙2 = v 0 + ω −e1 + vr sin(e3 ) (19) used to find these sub-optimal scaling factors. 266
e˙3 0 −1 ωr
Therefore, a back-stepping technique is employed to obtain D. Design of a dynamic controller and an ESO 267
control signals that do not include skid-slip [? ]. The following equation can be constructed from the dy-
namic equation in Eq. (2):
vc vr · cos(e3 ) + kx · e1
uc = = (20)
ωc ωr + ky · vr · e2 + kθ · vr sin(e3 ) M̄ · û˙ c + F = B̄ · τ (24)
6
with t ∈ [0, ∞) such that the first derivative holds the condition 285
x sin θ − ye cos θ 1
limt→∞ α˙1 (t) = 0. 286
M̄ = T T · M · T, B̄ = T T B, T = e , Theorem 1: Consider the model of UGV Eq. (1) and Eq.
1 0
(2), the posture control error Eq. (18) and Eq. (20), the ESO
F = T T [M · Ṫ (ue + Π) + M · T Π̇ − M · η̇ + H1 (v − η)] Eq. (28)- . (33), and controller Eq. (33). Then a compensated
+T T [H2 · ρ̇ + H3 · ρ] + B̄ · τd , error-inequality as follows
Π = [vr · cos θe + ωr · (xe · sin θ − ye · cos θ), ωr ]T Cmax
kec 2 k ≥ (Kcmin > kH¯m k) (34)
268 As per Assumption 2 we consider that the parameter M is Kcmin − kH¯m k
269 unknown, while nominal M0 is known, and error occurred on
the above inequality holds by adjusting the parameter, which 287
270 velocities could be defined as δ = u − ûc , we now have the
can get uniform aysmptotic stability of both e and ec 288
271 following expression
while,
−1
δ̇ = M̄0 B̄ · sat(τ ) + C · B̄ · sat(τ ) − M̄ −1 · F − û˙ c (25)
H̄m = H1 (X, Ẋ) · (Ω − η) + H2 (X) · ρ̇(X, µ)+
−1 −1 (35)
where C = M̄ − M̄0 . To address the saturation limi- H3 (X) · ρ(X, µ)
tations, we may also use ∆τ = sat(τ ) − τ . Additionally, it
will suggest that û˙ c is connected to an unidentified amount of Proof 1: . Consider the following Lyapunov function
skid-slip. Consequently, the above phrase takes the following V1 = k1 (ex + l(1 − cos(eθ ))2 + k1 (ey − l sin(eθ ))2
form:
+2k3 vr (1 − cos(eθ )) + V2
−1
δ̇ = M̄0 B̄ · τ + ∆ · τ + C · B̄ · sat(τ ) − M̄ −1 · F (26) (36)
−1
− û˙ c − (M̄0 B̄ − I)∆τ eT M̄ e
where V2 = c 2 c . Additionally, only when e = 0 and ec = 0 289
To accomplish the ESO, we first need to lumped the distur- is it known that the Lyapunov function V1 ≥ 0 and V1 = 0. 290
−1
zD1 = M̄ −1 · F − ˆ˙uc + (M̄0 B̄ − I) · ∆τ, (27)
V˙1 = 2k1 (ex + l(1 − cos(eθ ))(ėx + l sin(eθ )ėθ )
zD2 = C · B̄ · sat(τ ), (28)
+ 2k1 (ey − l sin(eθ )(ėθ − l cos(eθ )+ (37)
zD = zD1 + zD2 . (29)
2ke vr sin(eθ )ėθ + V̇2
272 Assumption 3: [59]: The first-order derivative of zD exists
273 and it is bounded. Let the function of Eq. (26) has an error eF = F − F̂ , and
The expression in Eq. (26) can now be rewritten as controller of Eq. (33), we can differentiate V2 and get
−1
δ̇ = M̄0 B̄ · τ + ∆τ + zD (30) 1
V˙2 = eTc (M̄˙ − 2H¯m ) · ec + eTc · (−Kc · ec + eF ) (38)
2
For the above system, the ESO design is as follows
It is crucial to note that the first term is zero due to the skew
ˆ −1
symmetry characteristic in this case, allowing us to recast the
δ̇ = M̄0 · B · τ + ∆τ + ẑD + LA · (δ − δ̂) (31)
V˙2 as
żˆD = LB · (δ − δ̂) (32)
V˙2 = eTc (−Kc ec + eF ) ≤ (−Kcmin kec 2 k
where δ̂ and ẑD are the estimations of δ and zD , respectively. (39)
LA and LB are observer gains. Thus, the dynamic control + kHm kkec 2 k + kec kCmax )
input τ can be defined as follows where Kcmin is the minimum value of Kc and Cmax is the
−1 β δ 2 maximum ESO estimate error. Now that the inequality and
τ = M̄0 · B̄ · (−α · δ + β · Kc −− ẑD ) (33) derivatives of the velocity error have been replaced in Eq.
2
(22) and Eq. (23),, we obtain
274 where Kc is the input saturation’s compensation gain, β is the
275 controller gain, and α is a diagonal positive definite matrix. V˙1 ≤ 2k1 (ex + l(1 − cos(eθ ))(vr cos(eθ ) − vc )+
276 The observer’s stability analysis is carried out in accordance (2k1 vr (ey − l sin(eθ )) + 2k3 vr · (ωr − ωc ) sin(e3 )−
277 with [20]. As a result, the following subsection will cover the
278 closed loop stability analysis of the controller design. Kc kec k2 + kH¯m kkec k2 + kec Cmax k
min
(40)
279 E. Stability analysis Now, we will introduce EPKC from Eq. (22), defining kˆ2 =
k1
280 We shall specify a key assumption before providing the kˆ v
and substitute in Eq. (38).
3 r
281 stability analysis of the closed-loop system in Figure (2):
282 Assumption 4: [60]: Consider a first order continuous dif- V˙1 ≤ −2k̂12 (ex + l(1 − cos(eθ ))2 − 2k̂32 vr2 sin(eθ )2 + kec k2
283 ferential variable with a limit of α1 (t) ∈ [0, ∞). The second (−Kcmin + kHm k) + kec Cmax k
284 derivative of α(t) is therefore present and constrained for every (41)
7
It demonstrates that the first two components of the equation A. Dynamic collision avoidance 314
mentioned earlier are negative, guaranteeing that V˙1 is negative The intuition behind this simulation is to reach a goal 315
as long as the conditions are met. location while avoiding multiple moving obstacles having a 316
Consequently, we can describe the difference in velocity with the information of their location, speed, and radius. In 320
between EPKC and the estimated velocity as this simulation, the velocity and acceleration constraints are 321
as follows: 322
v̄ (vr cos(eθ ) + k1 (ex + l(1 − cos(eθ ))) − v̂ −0.45m/s ≤ v ≤ 0.45m/s, −45deg/s ≤ ω ≤ 45deg/s,
ec = e =
ω̄e (ωr + k2 vr (ey − l sin(eθ ) + k3 )vr sin(eθ ) − ω̂ −3m/s2 ≤ a ≤ 3m/s2 , −45deg/s2 ≤ ar2 ≤ 45deg/s2
r1
(43)
1) Obstacles with linear trajectories: Figure 3a shows the 323
Bothkek and kėk are bounded since the error function es = collision avoidance of the UGV with the moving obstacle 324
[e ec ]T is bounded. Additionally, the Assumption 4 leads to having radius 0.2 at a starting position of (x, y) = (3.5, 1.5), 325
0
limt→∞ V˙˙1 = 0. The posture and velocity errors of the UGV driving at a speed of 0.40m/s with the heading of θ = 180 . 326
are therefore said to be asymptotically stable. Since we are using an NMPC-based planner that can predict 327
the obstacle’s future location and the vehicle. Therefore, when 328
lim ea = 0, lim eb = 0 (44) it predicts the incoming of the obstacle, the UGV turns left 329
t→∞ t→∞ to avoid the collision. The vehicle slows down to avoid the 330
Now, substituting equality Eq. (43) into equality Eq. (42) and incoming obstacle having a speed of 0.35m/s0 approaching 331
inequality Eq. (40), we get the following from (x, y) = (4.0, 0) with a heading θ = 135 as shown in 332
k1 (ex + l(1 − cos(eθ )) + k3 vr sin(eθ ) = 0 (45) The reverse action of UGV is of importance as in the actual 334
case; a driver can stop the vehicle and give way to the obstacle. 335
which follows the It can be improved by increasing the prediction horizon length 336
lim ex = 0, lim eθ = 0 (46) longitudinal velocity and yaw rate. Similarly, the UGV has 338
t→∞ t→∞
avoided the collision with another obstacle coming with a 339
Similarly, by the above expressions, there exists speed of 0.1m/s, as shown in Figure 3c, the vehicle’s turning 340
lim ey = 0 (47) Figure 3d illustrates, the UGV has safely reached the goal 342
t→∞ location and stopped (parked). Figure 4a shows that the 343
292 Thus, the position and velocity error of the vehicle are saturated input torque has reasonably respected the saturated 344
293 uniformly asymptotically stable. constraints to comment on the proposed control structure’s 345
294 IV. R ESULTS AND DISCUSSIONS UGV has avoided the collision, the input torque converges to 348
295 The simulation tests to verify the resilience of the suggested produced by the NMPC and the velocity applied by the EPKC. 350
296 strategy are covered in this part. The following physical For this purpose, Figure 4b illustrates, the EPKC has followed 351
297 parameters are established: b = 0.76 m, r = 0.29 m, the proposed velocities with little chattering. The reason for 352
298 d = 0.59 m, m = 49 kg, mw = 1.9kg, Jc = 14 kgm2 , linear velocities to suddenly slow down 7.5s is due to obstacle. 353
299 J = 0.008 kgm2 , Jm = 0.0034 kgm2 , .The skid-slip are set as The proposed accelerations generated by the NMPC are shown 354
300 [η1 , η2 , µ] = 0.03 · [sin(0.2t), 1, 1] and external disturbance in Figure 4c, which shows the smoothness in the control action 355
301 td = [0.02 · cos(0.4t)0.02 · sin(0.4t)]T . For the dynamic in the linear acceleration. The angular acceleration has some 356
302 collision avoidance, the skid-slip parameters are employed as a fluctuations till 15s due to a change in the heading at specific 357
303 low-frequency piece-wise linear (PWL) function. The side slip points in the path. 358
304 angle β of the front wheel is calculated as β = arc tan( ẋẏ ). In Figure 4d shows the tracking errors of the UGV from the 359
305 terms of controller gain tuning, we have selected the weighting proposed NMPC-based trajectory and the actual one. The 360
306 matrices Q = diag[6, 11, 0.02], Q̄ = diag[0.02, 0.02], and Median of states are close to 0 while the standard deviation µ 361
307 R = diag[0.04, 0.04]. The prediction horizon was set to for x, y, and θ are 1cm, 1.3cm, and 0.009rad/s, respectively. 362
308 N = 14 for the experiment related to static obstacles, and a Additionally, certain outliers are also highlighted, which are 363
309 larger one of N = 19, for the context which include multiple respectable for the proposed simulation. It is important to 364
310 dynamic objects. Additionally, Ts = 0.01 s is chosen for the mention here that most tracking errors are within (−2 − 0)cm 365
311 sample time. The input torque is considered to be saturated and for the position and (−0.06 − 0)rad/s for the heading. 366
312 to have values of τmax = 10 Nm and τmin = −10 Nm. The 2) Obstacles with nonlinear trajectories: In a real envi- 367
313 subsections that follow address the input and state limitations. ronment, it is not necessary for an obstacle to have a linear 368
8
5 5
UGV
4 Prediction 4
Trajectory
Goal
3 1.8 MO 3
TMO
1.6
y (m)
y (m)
2 2
1.4
1.2
1 1 UGV
Prediction
1
Trajectory
1.6 1.8
0 0 Goal
MO
TMO
-1 -1
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
x (m) x (m)
(a) (b)
5 5
4 4
3 3
y (m)
y (m)
2 2
UGV
1 UGV 1 Prediction
Prediction Trajectory
Trajectory Goal
0 Goal 0 MO
MO TMO
TMO
-1 -1
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
x (m) x (m)
(c) (d)
Fig. 3: (a) UGV slowing down to avoid collision from the dynamic obstacle approaching from (x = 3.5 , y = 1.5) with a heading of 180o .
(b) The vehicle is reversing and moving forward to prevent collision from an obstacle approaching from (x = 4.0 , y = 0) with a heading
of 135o . (c) UGV changing direction to avoid collision from the obstacle approaching from (x = 3.0, y = 4.0) with a heading of 255o . (d)
UGV stabilizes at the goal after avoiding dynamic obstacles.
15 0.5
(m/s)
10
0
1
Acutal longitudinal velocity
5 2
Reference longitudinal velocity
-0.5
10 Time(s) 15
(Nm)
0 5 20 25
0
50
Acutal angular velocity
-5
Reference angular velocity
(deg/s)
0
-10
-15 -50
0 2 4 6 8 10 12 14 0 5 10 15 20 25
Time (seconds) Time(s)
(a) (b)
1 0.08
Median
0.5
a r1 (m/s2 )
0.06
25%-75%
0 9%-91%
State Tracking errors
0.04 outliers
-0.5
-1 0.02
0 5 10 15 20 25 30
Time (s) 0
50
-0.02
a r2 (rad/s 2 )
-0.04
0
-0.06
-50
0 5 10 15 20 25 30 -0.08
Time (s) xe(m) ye(m) (rad)
e
(c) (d)
Fig. 4: (a) Saturated torque control applied to the platform in the presence of dynamic obstacles. (b)Actual and reference velocity commends
applied to the platform in the presence of dynamic obstacles. (c) Longitudinal and angular acceleration applied to the platform in the presence
of dynamic obstacles. (d) Platform’s state tracking errors, which shows the accuracy of the proposed scheme in the presence of dynamic
obstacles
9
2 0.4
0.2
1.5
(m/s)
0
-0.2
1
-0.4
0 1 2 3 4 5 6 7 8 9
y (m)
0.5 Time(seconds)
Reverse
40
0 UGV
Prediction 20
(deg/s)
Trajectory
-0.5 Goal 0
MO
TMO -20
-1 0 1 2 3 4 5 6 7 8 9
0 0.5 1 1.5 2 2.5 3 3.5
Time(seconds)
x (m)
(a) (b)
Fig. 5: (a) The UGV avoiding an incoming nonlinear trajectory of MO (TMO). (b) The UGV approaching to goal location after avoiding
the collision with MOs. (c) Longitudinal velocity and yaw rate applied to the platform under the presence of nonlinear TMO.
300
100 250
198.5
200 199
200
y[m]
300
y[m]
199.5
150
400 200
Goal 100
Start
500 Trajectory
MO 200.5
50
UGV
600 SO
Prediction 201
0
200 400 600 800 1000 1200 739 740 741 742 743 744 745 746
x[m] x[m]
Longitudinal velocity[m/s]
1.5 0.03
Median
1
0.02
0.5
0.01
0
-0.5 0
0 100 200 300 400 500 600
Time[s] -0.01
40
Yaw Rate [deg/s]
-0.02
20
-0.03
0
-20 -0.04
-40
0 100 200 300 400 500 600 -0.05
xe ye
Time[s] e
Fig. 6: (a) Bird’s eye view of path-tracking of a large map (b) MO avoidance in the large map (c) Applied longitudinal velocity and yaw
rate to the platform (d) Box plot of path-tracking errors of individual states of the platform
369 trajectory. Consider a nonlinear trajectory case of the obstacles nature of trajectory of MO can suddenly approach the vehicle. 386
370 as shown in Figure 5a and 5b. The UGV has first completely Thus, it is crucial for the UGV to instantly make a decision 387
371 to stop to prevent a collision from both of the vehicles (also to prevent the collision. 388
376 Additionally, two situations could occur: 1) Either prediction entries. Each subset is made of objects whose current costs fall 391
377 horizon is large enough to allow the UGV to break and stop. 2) within a specified cost interval. The queue simply preserves the 392
378 The cost function must be modified as per the given situation, order of its items. In this tracking experiment, the University 393
379 i.e. the penalization must be refined on the internal input states. of New South Wales Wale’s evacuation map is considered, 394
380 In contrast, if these things are neglected, the vehicle might take and the tracking method is combined with a 2 DoF optimum 395
381 a reverse and then goes forward, which could cause a collision planner (PPQ-Dijkstra). The campus’s buildings are shown on 396
382 due to the nonlinear nature of the obstacles.The longitudinal the 1350 x 600 m map. The experiment’s goal is to steer clear 397
383 velocity and yaw rate applied to the UGV can be seen in of MOs and SOs on a big map. As shown in Figure 8a, three 398
384 the Figure 5c, which clearly indicates the reverse action taken destinations are designated for this purpose, with the UGV 399
385 by the UGV from (1-2)seconds. This is because the nonlinear able to select the quickest route from the PPQ-Dijkstra (which 400
can deal with multiple destinations). Figure 6a’s coloured lines 401
10
Prediction Horizon
Prediction Horizon
50 SS[20] 2
MS [20]
15 PPC
0
0 20 40 60 80 100 120 N = 15 N = 50 N = 100 N = 150
Average NMPC processing time (ms) Prediction Horizon
(a) (b)
9
Multi-Core CPU
8 Single Core CPU
7
Average PTI (ms)
6
0
N = 15 N = 50 N = 100 N = 150
Prediction Horizon
(c)
Fig. 7: a) Comparing average PTI of the proposed PPC with single and multiple shootings method of IP at different prediction horizons. b)
Comparing PPC-based Euclidean distance tracking errors with single and multiple shootings method of IP at different prediction horizons.
c) PTI comparison of PPC, while running on a single-core CPU and multi-core CPU at different prediction horizons
402 depict the nominal cost of getting there (which implicitly prediction horizon, the error would be lower in PPC. Addition- 430
403 means approaching one of the valid destinations). For instance, ally, a detailed comparison of the optimization techniques is 431
404 red has a maximum value of 300, indicating a higher cost, provided in Figure 7b. The benefit of using prediction horizons 432
405 thus in order to follow the lowest cost, it is necessary to N > 100 is marginal in term of accuracy, as it can be 433
406 follow the automobile to blue (value 0). The vehicle’s non- appreciated in the table showing the errors. For computational 434
407 holonomic restrictions and collision avoidance are under the purposes, we have provided this comparison to save the PTI 435
408 control of the NMPC. Figure 6b shows the UGV avoiding of the onboard processor. 436
409 the MO as it approaches while also foreseeing the vehicle’s 2) Processing in single and multi-core CPUs: In terms 437
410 state. The velocity applied to the actual platform in terms of running the PPC in the single and multi-core CPUs, 438
411 of the effectiveness of the recommended system is shown in we have executed the PPC method with different prediction 439
412 Figure 6c. Chattering in the yaw rate and longitudinal velocity horizons. An increase in the prediction horizon could increase 440
413 greatly demonstrates environment collision avoidance. Figure the PTI due to an increase in the number of mathematical 441
414 6d shows that the actual states of the UGV are correct in terms computations. For instance, The Figure 7c shows that the PPC 442
415 of tracking errors. has a slight change in the PTI in both single-core CPU and 443
416 C. Performance analysis of the proposed scheme on a CPU with limited resources. Furthermore, it is evident 445
423 shooting method. In both the dynamic and static obstacles, the In this paper, a comprehensive novel PPC approach is 451
424 larger prediction horizon will have a more extensive processing proposed for designing the proximal formulation of an UGV. 452
425 cost, which is valid for practical cases. It is essential to The PPC consists of an NMPC-based path-planner, EP-based 453
426 mention that a prediction horizon of N = 100 has an almost KC, dynamic control law, and an ESO to handle the lumped 454
427 identical Euclidean distance error in PPC and IP methods disturbances. The proposed PPC was implemented on single 455
428 (SS and MS), with the former having less processing cost. and multi-core CPUs, which verifies that it has the ability to 456
429 Therefore, in regards to [20], using the same parameters and run on a CPU with limited resources despite an increase in 457
11
TABLE I: Average P T /I for PPC running with PPQ-Dijkstra using single, multi-core CPUs, and a comparison of Euclidean distance
accuracy
N Single Core this paper Single Core [20] Multi-Core this paper Multi-Core [20] This paper Accuracy Accuracy [20]
15 1.2 ms 10.3 ms 1.1 ms 10 ms -2.0 cm–2.0 cm -10.0 cm–10.0 cm
50 6.12 ms 21.47 ms 5.35 ms 21 ms -0.45 cm–0.45 cm -2.5 cm–2.5 cm
100 7.80 ms 38.61 ms 7.17 ms 38 ms -0.04 cm–0.04 cm -0.25 cm–0.25 cm
150 8.22 ms 51.82 ms 8.79 ms 51 ms -0.01 cm–0.01 cm -0.01 cm–0.01 cm
458 the prediction horizon. The velocity commands and dynamic American Control Conference (ACC). IEEE, 2018, pp. 5994– 512
459 control signals have shown less chattering, which guarantees 6001. 513
460 the small Euclidean distance error. Simulation experiments [11] A. Katriniok, P. Kleibaum, and M. Joševski, “Distributed model 514
463 in lowering the state-tracking errors in the presence of skid- [12] A. Sathya, P. Sopasakis, R. Van Parys, A. Themelis, 518
468 proper penalization on the internal inputs, and the prediction and efficient algorithm for nonlinear model predictive control,” 524
469 horizon’s length should be increased. The processing time of in 2017 IEEE 56th Annual Conference on Decision and Control 525
473 future work, we expect to validate our proposed scheme by [15] J. A. E. Andersson, J. Gillis, G. Horn, J. B. Rawlings, and 530
474 implementing it on an actual robot. M. Diehl, “CasADi – A software framework for nonlinear 531
476 [1] C. Samson, P. Morin, and R. Lenain, “Modeling and control IEEE transactions on Automatic Control, vol. 40, no. 1, pp. 536
477 of wheeled mobile robots,” in Springer handbook of robotics. 64–77, 1995. 537
478 Springer, 2016, pp. 1235–1266. [17] S.-O. Lee, Y.-J. Cho, M. Hwang-Bo, B.-J. You, and S.-R. Oh, 538
479 [2] S. G. Olsen and G. M. Bone, “Modelling of robotic bulldozing “A stable target-tracking control for unicycle mobile robots,” in 539
480 operations for autonomous control,” in 2011 24th Canadian Proceedings. 2000 IEEE/RSJ International Conference on Intel- 540
481 Conference on Electrical and Computer Engineering (CCECE). ligent Robots and Systems (IROS 2000)(Cat. No. 00CH37113), 541
482 IEEE, 2011, pp. 001 188–001 193. vol. 3. IEEE, 2000, pp. 1822–1827. 542
483 [3] Y.-S. Lee, S.-H. Kim, J. Seo, J. Han, and C.-S. Han, “Blade [18] J. Velagic, N. Osmic, and B. Lacevic, “Neural network con- 543
484 control in cartesian space for leveling work by bulldozer,” troller for mobile robot motion control,” World Academy of 544
485 Automation in Construction, vol. 118, p. 103264, 2020. Science, Engineering and Technology, vol. 47, pp. 193–198, 545
487 and V. G. Junior, “Robust path-following control for articulated [19] X. Zhang, Y. Fang, and N. Sun, “Visual servoing of mobile 547
488 heavy-duty vehicles,” Control Engineering Practice, vol. 85, pp. robots for posture stabilization: from theory to experiments,” 548
489 246–256, 2019. International journal of robust and nonlinear control, vol. 25, 549
490 [5] M. Natole, Y. Ying, and S. Lyu, “Stochastic proximal algorithms no. 1, pp. 1–15, 2015. 550
491 for auc maximization,” in International Conference on Machine [20] Q. Lu, D. Zhang, W. Ye, J. Fan, S. Liu, and C.-Y. Su, “Targeting 551
492 Learning. PMLR, 2018, pp. 3710–3719. posture control with dynamic obstacle avoidance of constrained 552
493 [6] G. R. Campos, P. Falcone, H. Wymeersch, R. Hult, and uncertain wheeled mobile robots including unknown skidding 553
494 J. Sjöberg, “Cooperative receding horizon conflict resolution at and slipping,” IEEE Transactions on Systems, Man, and Cyber- 554
495 traffic intersections,” in 53rd IEEE Conference on Decision and netics: Systems, 2020. 555
496 Control. IEEE, 2014, pp. 2932–2937. [21] Y. Wang, Z. Miao, H. Zhong, and Q. Pan, “Simultaneous 556
497 [7] J. Rios-Torres and A. A. Malikopoulos, “A survey on the co- stabilization and tracking of nonholonomic mobile robots: A 557
498 ordination of connected and automated vehicles at intersections lyapunov-based approach,” IEEE Transactions on Control Sys- 558
499 and merging at highway on-ramps,” IEEE Transactions on tems Technology, vol. 23, no. 4, pp. 1440–1450, 2015. 559
500 Intelligent Transportation Systems, vol. 18, no. 5, pp. 1066– [22] B. Li, Y. Fang, G. Hu, and X. Zhang, “Model-free unified 560
501 1077, 2016. tracking and regulation visual servoing of wheeled mobile 561
502 [8] J. Shi, Y. Zheng, Y. Jiang, M. Zanon, R. Hult, and B. Houskal, robots,” IEEE Transactions on Control Systems Technology, 562
503 “Distributed control algorithm for vehicle coordination at traffic vol. 24, no. 4, pp. 1328–1339, 2015. 563
504 intersections,” in 2018 European Control Conference (ECC). [23] Q. Lu, L. Yu, D. Zhang, and X. Zhang, “Simultaneous track- 564
505 IEEE, 2018, pp. 1166–1171. ing and regulation visual servoing of wheeled mobile robots 565
506 [9] M. Kneissl, A. Molin, H. Esen, and S. Hirche, “A feasible with uncalibrated extrinsic parameters,” International Journal 566
507 mpc-based negotiation algorithm for automated intersection of Systems Science, vol. 49, no. 1, pp. 217–229, 2018. 567
508 crossing,” in 2018 European Control Conference (ECC). IEEE, [24] Z. Li, J. Deng, R. Lu, Y. Xu, J. Bai, and C.-Y. Su, “Trajectory- 568
509 2018, pp. 1282–1288. tracking control of mobile robot systems incorporating neural- 569
510 [10] F. Molinari and J. Raisch, “Automation of road intersections dynamic optimized model predictive approach,” IEEE Transac- 570
511 using consensus-based auction algorithms,” in 2018 Annual tions on Systems, Man, and Cybernetics: Systems, vol. 46, no. 6, 571
12
572 pp. 740–749, 2015. to actuator saturation: An extended state observer approach,” 641
573 [25] G. Klančar and I. Škrjanc, “Tracking-error model-based pre- Automatica, vol. 107, pp. 353–361, 2019. 642
574 dictive control for mobile robots in real time,” Robotics and [43] S. J. Yoo, “Adaptive neural tracking and obstacle avoidance of 643
575 autonomous systems, vol. 55, no. 6, pp. 460–469, 2007. uncertain mobile robots with unknown skidding and slipping,” 644
576 [26] E. Kayacan, H. Ramon, and W. Saeys, “Robust trajectory Information Sciences, vol. 238, pp. 176–189, 2013. 645
577 tracking error model-based predictive control for unmanned [44] L. Li, T. Wang, Y. Xia, and N. Zhou, “Trajectory tracking con- 646
578 ground vehicles,” IEEE/ASME Transactions on Mechatronics, trol for wheeled mobile robots based on nonlinear disturbance 647
579 vol. 21, no. 2, pp. 806–814, 2015. observer with extended kalman filter,” Journal of the Franklin 648
580 [27] W. He, S. S. Ge, Y. Li, E. Chew, and Y. S. Ng, “Neural network Institute, vol. 357, no. 13, pp. 8491–8507, 2020. 649
581 control of a rehabilitation robot by state and output feedback,” [45] B. Burns and O. Brock, “Sampling-based motion planning with 650
582 Journal of Intelligent & Robotic Systems, vol. 80, no. 1, pp. sensing uncertainty,” in Proceedings 2007 IEEE International 651
583 15–31, 2015. Conference on Robotics and Automation. IEEE, 2007, pp. 652
584 [28] J. Huang, C. Wen, W. Wang, and Z.-P. Jiang, “Adaptive output 3313–3318. 653
585 feedback tracking control of a nonholonomic mobile robot,” [46] L. Ma, J. Xue, K. Kawabata, J. Zhu, C. Ma, and N. Zheng, “Ef- 654
586 Automatica, vol. 50, no. 3, pp. 821–831, 2014. ficient sampling-based motion planning for on-road autonomous 655
587 [29] M. Hirayama, J. Guivant, J. Katupitiya, and M. Whitty, “Artif- driving,” IEEE Transactions on Intelligent Transportation Sys- 656
588 cial interlligence in path planning for autonomous bulldozers: tems, vol. 16, no. 4, pp. 1961–1976, 2015. 657
589 comparison with manual operation,” International Journal of [47] S. Liu, N. Atanasov, K. Mohta, and V. Kumar, “Search-based 658
590 Innovative Computing, Information and Control, vol. 15, no. 3, motion planning for quadrotors using linear quadratic minimum 659
591 pp. 825–844, 2019. time control,” in 2017 IEEE/RSJ international conference on 660
592 [30] S. Khan and J. Guivant, “Nonlinear model predictive path- intelligent robots and systems (IROS). IEEE, 2017, pp. 2872– 661
594 for accurate placement of materials and debris of masonry [48] B. Li, Y. Zhang, Y. Ge, Z. Shao, and P. Li, “Optimal control- 663
595 in construction contexts,” IEEE Access, vol. 9, pp. 102 069– based online motion planning for cooperative lane changes of 664
596 102 080, 2021. connected and automated vehicles,” in 2017 IEEE/RSJ Inter- 665
597 [31] M. Hirayama, J. Guivant, J. Katupitiya, and M. Whitty, “System national Conference on Intelligent Robots and Systems (IROS). 666
598 and method for planning travel path for work machine, and work IEEE, 2017, pp. 3689–3694. 667
599 machine,” Jan. 30 2020, aU Patent AU2018334390B2. [49] A. J. Häusler, A. Saccon, A. P. Aguiar, J. Hauser, and A. M. 668
600 [32] X. Wang, J. Taghia, and J. Katupitiya, “Adaptive min–max Pascoal, “Energy-optimal motion planning for multiple robotic 669
601 model predictive control for field vehicle guidance in the pres- vehicles with collision avoidance,” IEEE Transactions on Con- 670
602 ence of wheel slip,” Robotics and Mechatronics for Agriculture, trol Systems Technology, vol. 24, no. 3, pp. 867–883, 2015. 671
603 pp. 157–184, 2017. [50] A. Robledo, J. Guivant, and S. Cossell, “Pseudo priority 672
604 [33] S. Wei, K. Uthaichana, M. Žefran, and R. DeCarlo, “Hybrid queues for real-time performance on dynamic programming 673
605 model predictive control for the stabilization of wheeled mobile processes applied to path planning,” in Australasian Conference 674
606 robots subject to wheel slippage,” IEEE transactions on control on Robotics and Automation, 2010. 675
607 systems technology, vol. 21, no. 6, pp. 2181–2193, 2013. [51] D. Wang and C. B. Low, “Modeling and analysis of skidding 676
608 [34] J. Taghia and J. Katupitiya, Applied Guidance Methodologies and slipping in wheeled mobile robots: Control design perspec- 677
609 for Off-road Vehicles. Springer, 2020, vol. 138. tive,” IEEE Transactions on Robotics, vol. 24, no. 3, pp. 676– 678
610 [35] S. Khan and J. Guivant, “Fast nonlinear model predictive 687, 2008. 679
611 planner and control for an unmanned ground vehicle in the [52] M. Chen, “Disturbance attenuation tracking control for wheeled 680
612 presence of disturbances and dynamic obstacles,” Scientific mobile robots with skidding and slipping,” IEEE Transactions 681
613 Reports, vol. 12, p. 12135, 2022. on Industrial Electronics, vol. 64, no. 4, pp. 3359–3368, 2016. 682
614 [36] M. W. M. Said, “Optimization based solutions for control [53] R. Fierro and F. L. Lewis, “Control of a nonholonomic mobile 683
615 and state estimation in non-holonomic mobile robots: stability, robot using neural networks,” IEEE transactions on neural 684
616 distributed control, and relative localization,” arXiv preprint networks, vol. 9, no. 4, pp. 589–600, 1998. 685
617 arXiv:1803.06928, 2018. [54] J. Huang, C. Wen, W. Wang, and Z.-P. Jiang, “Adaptive 686
618 [37] S. Khan, J. Guivant, and X. Li, “Design and experimental stabilization and tracking control of a nonholonomic mobile 687
619 validation of a robust model predictive control for the optimal robot with input saturation and disturbance,” Systems & Control 688
620 trajectory tracking of a small-scale autonomous bulldozer,” Letters, vol. 62, no. 3, pp. 234–241, 2013. 689
621 Robotics and Autonomous Systems, vol. 147, p. 103903, 2022. [55] J. C. Dunn and D. P. Bertsekas, “Efficient dynamic program- 690
622 [38] K.-S. Kim and K.-H. Rew, “Reduced order disturbance observer ming implementations of newton’s method for unconstrained 691
623 for discrete-time linear systems,” Automatica, vol. 49, no. 4, pp. optimal control problems,” Journal of Optimization Theory and 692
624 968–975, 2013. Applications, vol. 63, no. 1, pp. 23–38, 1989. 693
625 [39] R. ul Amin, I. Inayat, and L. A. Jun, “Finite time position [56] J. Nocedal and S. Wright, Numerical optimization. Springer 694
626 and heading tracking control of coaxial octorotor based on Science & Business Media, 2006. 695
627 extended inverse multi-quadratic radial basis function network [57] P. Patrinos and A. Bemporad, “Proximal newton methods for 696
628 and external disturbance observer,” Journal of the Franklin convex composite optimization,” in 52nd IEEE Conference on 697
629 Institute, vol. 356, no. 8, pp. 4240–4269, 2019. Decision and Control. IEEE, 2013, pp. 2358–2363. 698
630 [40] H.-S. Kang, Y.-T. Kim, C.-H. Hyun, and M. Park, “Generalized [58] Y. Kanayama, Y. Kimura, F. Miyazaki, and T. Noguchi, “A sta- 699
631 extended state observer approach to robust tracking control for ble tracking control method for an autonomous mobile robot,” in 700
632 wheeled mobile robot with skidding and slipping,” International Proceedings., IEEE International Conference on Robotics and 701
633 Journal of Advanced Robotic Systems, vol. 10, no. 3, p. 155, Automation. IEEE, 1990, pp. 384–389. 702
634 2013. [59] H. Yang, X. Fan, P. Shi, and C. Hua, “Nonlinear control for 703
635 [41] A. Mohammadi, M. Tavakoli, H. J. Marquez, and tracking and obstacle avoidance of a wheeled mobile robot 704
636 F. Hashemzadeh, “Nonlinear disturbance observer design with nonholonomic constraint,” IEEE Transactions on Control 705
637 for robotic manipulators,” Control Engineering Practice, Systems Technology, vol. 24, no. 2, pp. 741–746, 2015. 706
638 vol. 21, no. 3, pp. 253–267, 2013. [60] J.-J. E. Slotine and W. Li, Applied nonlinear control. Prentice 707
639 [42] Y. Yuan, Z. Wang, Y. Yu, L. Guo, and H. Yang, “Active distur- hall Englewood Cliffs, NJ, 1991, vol. 199, no. 1. 708
728