Scientia Iranica D (2021) 28(3), 1592{1605
Sharif University of Technology Scientia Iranica Transactions D: Computer Science & Engineering and Electrical Engineering http://scientiairanica.sharif.edu
Tele-operation of autonomous vehicles over additive white Gaussian noise channel
A. Parsa and A. Farhadi
Department of Electrical Engineering, Sharif University of Technology, Tehran, Iran.
Received 17 September 2019; received in revised form 21 January 2020; accepted 14 April 2020
KEYWORDS Abstract. This paper is concerned with the tele-operation of autonomous vehicles over analog Additive White Gaussian Noise (AWGN) channel, which is subject to transmission Networked control noise and power constraint. The nonlinear dynamic of autonomous vehicles is described system; by the unicycle model and is cascaded with a bandpass lter acting as encoder. Using the Tele-operation system; describing function method, the nonlinear dynamic of autonomous vehicles is represented The describing by an approximate linear system. Then, the available results for linear control over analog function; AWGN channel are extended to account for linear continuous time systems with non-real The unicycle model; valued and multiple real valued eigenvalues and for tracking a non-zero reference signal. AWGN channel. Subsequently, by applying the extended results on the describing function of autonomous vehicles, a mean square control technique including an encoder, a decoder, and a controller is presented for reference tracking of the tele-operation of autonomous vehicles over AWGN channel. The satisfactory performance of the proposed control technique is illustrated by computer simulations. © 2021 Sharif University of Technology. All rights reserved.
1. Introduction the tele-operation of system given in Figure 1. Very often, autonomous vehicles are battery powered and 1.1. Motivation and background hence, communication from these vehicles to the base Tele-operation of autonomous vehicles has become an station where the operator is located must be done with active research direction in recent years. In these minimum possible transmission power to increase the systems, the remote autonomous vehicle must track a on-board battery life time. Therefore, communication reference signal generated by a remote operator which from vehicle to remote controller is subject to noise is communicated to it via a wireless link. In this and power constraint. However, as the communication application, the measurements from on-board sensors from the base station to remote vehicle can be done are also communicated to remote operators to help op- with high transmission power, in the block diagram erator to generate a desired reference signal. One of the of Figure 1, the communication link from remote abstract models for wireless communication is Additive controller to the remote vehicle (dynamic system) can White Gaussian Noise (AWGN) channel. This channel be considered without imperfections and limitations. is an abstract model for satellite communication, deep The tele-operation of Figure 1 is an example space communication, and when the line of sight is of networked control systems. In networked control strong. Therefore, in these situations, we deal with systems, we deal with controlling dynamic systems over communication channels subject to imperfections and *. Corresponding author. limitations. Some results addressing basic problems in E-mail address: [email protected] (A. Farhadi) stability and/or state tracking of dynamic systems over communication channels subject to imperfections and doi: 10.24200/sci.2020.54447.3755 limitations can be found in [1{28]. Dynamic systems A. Parsa and A. Farhadi/Scientia Iranica, Transactions D: Computer Science & ... 28 (2021) 1592{1605 1593
back systems, in which the Multi-Input Multi-Output (MIMO) communication link is modeled as a parallel noisy Additive White Noise (AWN) channel. The idea of using the describing function for controlling nonlinear dynamic systems over AWGN channel for the rst time was presented in [29]. In addition to the AWGN channel, other communication channels, which have caught the attention of networked control systems research community, are the real and the packet erasure channels that are abstract models for communication Figure 1. A dynamic system over Multi-Input via the Internet and WiFi links. Other researchers [30] Multi-Output (MIMO) parallel Additive White Gaussian addressed the problems of optimal control and stability Noise (AWGN) channel. of LTI systems over the real erasure channel, which is subject to random packet dropout, where both remote can be viewed as continuous alphabet information and local controllers operate on the system. Moreover, sources with memory. Therefore, many works in the the problem of optimal control of LTI systems by a literature (e.g., [1,12,13,18,19,22{28]) are dedicated to remote controller over the real erasure channel was the question of state tracking and/or stability over addressed in [31]. AWGN channel, which itself is naturally a continuous The above literature review reveals that the alphabet channel. Researchers [12,13] addressed the available results on the stability and state tracking problem of mean square stability and state tracking of of dynamic systems over AWGN channel are mainly linear Gaussian dynamic systems over AWGN channel concerned with linear dynamic systems. To the best when noiseless feedback channel was available full time of our knowledge, control of nonlinear dynamic sys- and the communication of control signal from remote tems over AWGN channel is limited to [23] and [29], controller to system was perfect (see Figure 1). In [12], which are not concerned with the tele-operation of the authors presented an optimal control technique for autonomous vehicles subject to communication imper- asymptotic bounded mean square stability of partially fections. In the tele-operation of autonomous vehicles, observed discrete time linear Gaussian systems over we deal with state tracking as well as reference tracking AWGN channel. In [13], the authors addressed a of nonlinear systems. The dynamic of autonomous continuous time version of the problem addressed in vehicles (autonomous underwater, unmanned aerial, [12]. In [22], the researchers considered a framework and autonomous road vehicles) is described by the for discussing control over a communication channel six-degree-of-freedom model. However, the vehicle based on Signal-to-Noise Ratio (SNR) constraints and dynamic is handled by an on-board control loop that focused particularly on the feedback stabilization of an results in a kinematic unicycle model, which is non- open loop unstable plant via a channel with a SNR linear [16]. In addition, in deep space tele-operation constraint. By examining the simple case of a Linear systems or when the line of sight is strong, AWGN Time Invariant (LTI) plant and an AWGN channel, channel is a suitable model for wireless communication. the authors in [22] derived necessary and sucient These issues motivate research on tele-operation of conditions on the SNR for feedback stabilization with the unicycle model over AWGN channel, which is the an LTI controller. In [23], the authors addressed the subject of this paper. problem of state tracking of nonlinear systems over AWGN channel. In [25], the authors presented a sub- 1.2. Paper contributions optimal decentralized control technique for bounded Key contributions of this paper compared to the afore- mean square stability of a large-scale system with mentioned literature are summarized as follows: cascaded clusters of sub-systems. Each sub-system 1. In the eld of nonlinear dynamic systems, the is linear and time-invariant and both sub-system and describing function is used for building oscillators its measurement are subject to Gaussian noise. The [32]. However, in this paper, for the rst time, it control signals are exchanged between sub-systems has been used to stabilize and control a practical without any imperfections, but the measurements are nonlinear dynamic system. That is, the unicycle exchanged via an AWGN communication network. In model is an abstract model for describing the [27], the author presented a sub- optimal technique dynamics of autonomous underwater, unmanned for mean square stability of a distributed system aerial and autonomous road vehicles; with geographically separated Gaussian sub-systems interconnected by a AWGN communication network. 2. To the best of our knowledge, earlier works on In [28], the authors investigated stabilization and controlling dynamic systems over the AWGN chan- performance issues for MIMO LTI networked feed- nel are mainly concerned with the stability and 1594 A. Parsa and A. Farhadi/Scientia Iranica, Transactions D: Computer Science & ... 28 (2021) 1592{1605
state tracking of linear dynamic systems, e.g., where x(t) and y(t) are the position vectors, (t) [1,12,13,18,19,22,24,25,27,28]. Only the referenced is the heading angle, and the control inputs are the studies [24,29] addressed the problem of controlling vehicle forward velocity v(t) and the angular velocity nonlinear dynamic systems over the AWGN chan- u(t). Note that for the remote controller the initial nel. Nevertheless, this study [24] is just concerned conditions (x(0), y(0), (0)) are unknown and have a with the estimation problem; and [29] addressed the Gaussian distribution. problem of tracking and stability of those systems that have periodic outputs to sinusoidal inputs. Communication channel. Communication channel Therefore, another study [29] used a linear dynamic between system and controller is an MIMO AWGN system subject to saturated excitation for computer channel without interference (parallel) with n inputs simulation as this type of noncommon dynamic and n outputs. The output of the encoder (which systems was consistent with the theoretical devel- will be described shortly) is transmitted through the opment in [29]. However, by nding a describing MIMO channel and a white Gaussian noise vector function for the unicycle model, which also has peri- is added to it (as is shown in Figure 1), where 0 ~ odic outputs to periodic inputs, this paper presents N(t) = n1(t) : : : nn(t) i.i.d. N(0; R) a real application of the theoretical development in ~ (R = diagfr~1; :::; r~ng) is the MIMO channel noise and [29] by controlling a practical nonlinear dynamic ni(t) N(0; r~i) is the additive noise of the ith input system over the AWGN channel. - output path of the MIMO parallel AWGN channel. 1.3. Paper organization Also, the MIMO parallel AWGN channel is subject to The paper is organized as follows. In Section 2, the the channel input power constraints P oi, i = 1; 2; :::; n 2 problem formulation is presented. Section 3 describes as follows: E[fi (t)] P oi, i = 1; 2; :::; n, where the describing function method. Then, Section 4 de- fi(t) is the ith element of the encoder output vector 0 velops the theory of mean square reference tracking of F (t) = f1(t) : : : fn(t) , which is the input of linear dynamic systems with multiple real and non-real the channel. That is, Y (t) = F (t) + N(t), where 0 valued eigenvalues over MIMO AWGN channel. Sec- Y (t) = y1(t) : : : yn(t) is the channel output. tion 5 is devoted to the tele-operation of autonomous vehicles. Section 6 gives simulation results for the Encoder. The encoder is a bandpass lter cascaded unicycle model and Section 7 concludes this paper with a matrix gain. This bandpass lter saves only by summarizing the contributions of the paper and the fundamental frequency of the system outputs and directions for future research. omits the other harmonics which have less information to be sent. For this purpose, a high pass lter with a 2. Problem formulation relatively low cut-o frequency (e.g., 0:1 Hz) is used to omit the DC part. This lter is cascaded with a low Throughout, certain conventions are used: E[] denotes pass lter with a cut-o frequency of ! 2 0:1 the expected value, var[:] the variance, and j j the rad/s. For constructing such a lter, we can use the absolute value and V 0 the transpose of vector/matrix following transfer functions: 1 V . A denotes the inverse of a square matrix A and H (s) = H (s)H (s); (2) N(m; n) is the Gaussian distribution with mean m and bp hp lp covariance n. R denotes the set of real numbers and In s2 the identity matrix with dimension n by n. trac(A), Hhp(s) = ;!h = 2 0:1 rad/s; s2 + !h s + !2 denotes the trace of a square matrix A, diagf:g denotes Q h the diagonal matrix, [A]ij denotes the i; jth element of Q = damping factor = 0:707; (3) the matrix A and 0 denotes the zero vector/matrix. This paper is concerned with asymptotic mean !2 H (s) = : (4) square stability and reference tracking of autonomous lp s2 + ! s + !2 vehicles over AWGN communication channel, as is Q shown in the block diagram of Figure 1. The building It will be shown in the next section that the nonlinear blocks of Figure 1 are described below. dynamic system (1) together with the bandpass lter has an approximate linear dynamic system with n = 7 0 Dynamic system: The dynamic system is the follow- states X(t) = x1(t) : : : xn(t) , in which these ing time-invariant unicycle system [16]: states and the matrix gain 8 2 3 c11(t) c12(t) ::: c1n(t) <>x_(t) = v(t) cos((t)) 6 7 6 c21(t) c22(t) ::: c2n(t) 7 y_(t) = v(t) sin((t)) (1) C(t) = ; > 4 ::: 5 : _ (t) = u(t) cn1(t) cn2(t) ::: cnn(t) A. Parsa and A. Farhadi/Scientia Iranica, Transactions D: Computer Science & ... 28 (2021) 1592{1605 1595 form the encoder output as F (t) = C(t)(X(t) X^(t)), of !, and all other harmonics are eliminated. In other ^ 0 where X(t) = x^1(t) ::: x^n(t) is the estima- words, the output of the bandpass lter is given in the tion of states. following: y (t) = a sin(!t) + b cos(!t): (6) Decoder: Decoder is the minimum mean square f 1 1 estimator or the Kalman lter, also known as Linear Having that, we call the nonlinear dynamic system Quadratic Estimator (LQE). At each time instant t, that is cascaded with the bandpass lter with the the Kalman lter generates an estimation X^(t) using high cut-o frequency of !, a quasi-linear system [32], the channel output Y (t). because we can nd a linear dynamic system with the input u(t) = cos(!t) and the output yf (t) with the Controller. Controller is a certainty equivalent con- following transfer function: troller of the following form: u(t) = l(t)X^(t) + (t). The objective of this paper is to nd the matrix H(j!) = jH(j!)j\H(j!); (7) gain C(t) and u(t) to force the positions x(t) and y(t) p a2 + b2 and the heading angle (t) follows the desired paths. jH(j!)j = 1 1 ; (8)
3. Implementation of the describing function a1 method \H(j!) = arctan rad: (9) b1 In this section, we rst present the idea of the describ- This means that the nonlinear dynamic system can ing function to obtain the approximate linear dynamic be represented by a linear dynamic system with the system from a nonlinear time-invariant dynamic sys- above transfer function called the describing function tem. Then, we obtain the describing function for the of the nonlinear dynamic system. Note that the nonlinear dynamic system of (1). describing function represents the transfer function 3.1. The Describing Function of the approximate linear dynamic system and the This subsection is borrowed from [29]. For all nonlinear transfer function represents the response of the time - invariant dynamic systems that respond period- dynamic system to the input signal when the initial ically to sinusoidal inputs, we can nd an approximate conditions are set to be zero. Hence, the describing linear dynamic system, as de ned below [32]. function is obtained for zero initial conditions. Consider an Single-Input Single-Output (SISO) 3.2. The approximate linear system for the nonlinear time-invariant dynamic system with periodic unicycle model outputs in response to periodic inputs. Suppose For a given xed forward velocity v(t), the nonlinear that this nonlinear dynamic system is excited by the dynamic system (1) has periodic outputs to periodic following input: u(t) = cos(!t), where > 0 is large inputs u(t) = cos(!t). Hence, in the block diagram of enough to excite all modes of the nonlinear system. Figure 1, as this dynamic is cascaded with the bandpass Then, as the output is a periodic signal, it has a Fourier lter, it has an approximate linear system description. series representation that includes all harmonics of the To obtain this describing function, we rst assume that input with frequency of !. That is: the input is u(t) = cos(!t); andx _(t) andy _(t) are the X1 outputs of System (1), hence the inputs of the bandpass y(t) = y + (a sin(i!t) + b cos(i!t)); f1 d i i lter. Therefore, we obtain Hx_ (s) = 2 and f2s +f3s+f4 i=1 c1 Hy_ (s) = 2 transfer functions as describing c2s +c32s+c4 Z 2 functions from the input to the outputsx _(t) andy _(t), ! ! respectively (f s and c s are real coecients). Then, yd = y(t)dt; i i 4 2 ! obviously for obtaining the describing functions from input u(t) to outputs x(t) and y(t), we must multiply Z 2 ! ! 1 s to these transfer functions. Note that as is clear ai = (y(t) sin(i!t))dt; 2 2 from Figure 1, the inputs to the bandpass lter are ! x(t) and y(t). However, we can assume that these Z 2 ! ! inputs are obtained by integration ofx _(t) andy _(t); bi = (y(t) cos(i!t))dt: (5) thus, for the simplicity of obtaining the describing 2 2 ! functions, by moving this integration operator after Now, if this nonlinear system is cascaded with a the transfer function of the bandpass lter, we can bandpass lter with a high cut-o frequency of !, then assume that the outputs of the nonlinear system and we have a periodic output at the end of the lter which the inputs to the lter arex _(t) andy _(t). Also, note consists of only the rst harmonic with the frequency that from Eq. (1), it follows that the transfer function 1596 A. Parsa and A. Farhadi/Scientia Iranica, Transactions D: Computer Science & ... 28 (2021) 1592{1605
1 between input u(t) and output (t) is H(s) = s . value of is unknown and hence N(X0;Q0) That is, the relation between u(t) and (t) is linear. (Q0 is diagonal) is treated as the Gaussian random The equivalent state space representation of the variable with known mean and variance. Note that 0 ~ ~ approximate linear system has seven states: X(t) = N(t) = n1(t) : : : nn(t) i:i:d: N(0; R)(R = 0 [x1(t) x2(t); : : : ; x7(t)] which correspond to d1x(t) + diagfr~1; :::; r~ng) is the additive noise of the MIMO d2x_(t) + x(t), d3x(t) + d4x_(t), d5x(t) + d6x_(t), d7y(t) + channel and (ni(t) N(0; r~i) is the additive noise of d8y_(t) + y(t), d9y(t) + d10y_(t), d11y(t) + d12y_(t) and the ith path of the MIMO parallel AWGN channel, (t), respectively (dis are real coecients). The input which can be treated as the measurement noise pro- of this representation is u(t) and the output vector is vided the channel input power constraint is met. The [x(t) y(t) (t)]0 with the system matrices: objective here is to achieve mean square asymptotic ref- 2 3 erence tracking for the linear system (10) with n states Ax 0 0 4 5 0 and n outputs over the MIMO parallel AWGN channel. A = 0 Ay 0 ;B = [Bx By B] 0 0 A Throughout, it is assumed that the system matrix A has real eigenvalues real multiple eigenvalues and 2 3 distinct complex conjugate eigenvalues as the system Cx 0 0 4 5 matrix A that corresponds to the describing function C = 0 Cy 0 ; of the nonlinear system (1) including these types 0 0 C of eigenvalues. According to the previous section, where: 2 3 2 3 for the unicycle model, the system matrix A of the 0 0 0 0 0 0 approximate linear system is in the real Jordan form. 4 5 4 5 Ax = 0 e1 e2 ;Ay = 0 e5 e6 ; Therefore, throughout this section, it is assumed that 0 e2 e1 0 e6 e5 System (10) can be decomposed to several decoupled sub-systems; hence, for each sub-system, an encoder A =0;Bx = [e9 e10 e11];By = [e12 e13 e14]; and a decoder are designed separately. Note that the Jordan block associated with a real eigenvalue i(A) B = 1;Cx = [e15 e16 0];Cy = [e17 e18 0]; with multiplicity 2 is the following matrix: (A) 1 C = 1: i ; 0 (A) Hence, we need a MIMO parallel AWGN channel with i and the Jordan block associated with the complex 7 inputs and 7 outputs for transmitting x1(t) to x7(t). p Note that the equivalent state space representation of conjugate pair of eigenvalues i(A) = 1w w the approximate linear system is in the real Jordan (w 6= 0) is . form. w 4.1. Mean square asymptotic state tracking 4. Reference tracking of linear systems with In this section, it is assumed that each of the Jordan multiple real and non-real valued block is at most a 2 by 2 matrix. Then, for all three eigenvalues over AWGN channel possible cases: w Now, we extend the results of [13] to account for a) A = , ; w 2 R; w 6= 0, reference tracking and hence, the stability of linear w dynamic systems with multiple real and non-real a 1 b) A = , a 2 R, valued eigenvalues over MIMO parallel AWGN 0 a channel. In the next section, by applying these a 0 extended results on the describing functions associated c) A = 1 , a ; a 2 R; a 6= a , 0 a 1 2 1 2 with the nonlinear system (1), we address the reference 2 tracking problem of tele-operation of autonomous we nd the matrix gain C(t) for mean square asymp- vehicles, as is shown in the block diagram of Figure 1. totic state tracking of system states at the decoder. Suppose that the dynamic system in Figure 1 is 4.1.1. Sub-systems with complex conjugate eigenvalues linear with n states. Hence, the system that is seen by Suppose that the system matrix A in the linear system the remote controller is as follows: ( of (10) has the following form: _ X(t) = AX(t) + Bu(t);X(0) = ; w (10) A = ; ; w 2 R; w 6= 0: (11) Y (t) = C(t)(X(t) X^(t)) + N(t) w where X(0) is the initial state that is known for Then, we have the following proposition for mean the encoder, but is unknown for the remote con- square asymptotic state tracking of system states at troller, that is, for the remote controller the exact the decoder. A. Parsa and A. Farhadi/Scientia Iranica, Transactions D: Computer Science & ... 28 (2021) 1592{1605 1597