Inter national J our nal of Robotics and A utomation (IJRA) V ol. 10, No. 3, September 2021, pp. 235 260 ISSN: 2089-4856, DOI: 10.11591/ijra.v10i3.pp235-260 235 Contr ol of teleoperation systems in the pr esence of cyber attacks: A sur v ey Mutaz M. Hamdan, Magdi S. Mahmoud Systems Engineering Department, King F ahd Uni v ersity of Petroleum & Minerals, Saudi Arabia Article Inf o Article history: Recei v ed May 28, 2021 Re vised Jul 9, 2021 Accepted Jul 16, 2021 K eyw ords: Bilateral teleoperation system Cyber attack Denial of service (DoS) attack P assi vity based control T elesur gical robot systems W a v e v ariable-based control ABSTRA CT The teleoperation system is often composed of a human operator , a local master ma- nipulator , and a remote sla v e manipulator that are connected by a communication net- w ork. This paper proposes a surv e y on feedback control design for the bilateral tele- operation systems (BTSs) in nominal situa tions and in the presence of c yber -attacks. The main i dea of the presented me thods is to achie v e the stability of a delaye d bilateral teleoperation system in the presence of se v eral kinds of c yber attacks. In this paper , a comprehensi v e surv e y on control systems for BTSs under c yber -attacks is discussed. Finally , we discuss the current and future problems in this eld. This is an open access article under the CC BY -SA license . Corresponding A uthor: 1. INTR ODUCTION A teleoperation system is referred to a plant that is controlled remotely . Man y teleoperation system s ha v e been designed and used in the recent decades to help people in performing tasks remotely especially in hazardous en vironments. These tasks include w orking in a toxic or harmful en vironment, w orking remotely lik e telesur gery and e xplorations of space, and carrying out high preci sion tasks such as chemical and nuclear reac- tors. In general, teleoperation systems consist of a human operator , a master and sla v e manipulators connected through a communi cation netw ork and the en vironment. The schematic diagram of a typical teleoperation system is sho wn in Figure 1. T eleoperation systems allo w operator to e xchange information of posi tion, v elocity , and/or force re- motely to perform the desired m otion, sensing, and ph ysical manipulation. The importance and wide applica- tions of teleoperation systems attract researchers in both control systems and robotics mainly focusing on the stability and the tele presence [1]-[3]. A lot of result are presented to handle practical control problems using se v eral approaches such as anti-wind control [4], [5], fuzzy control [6], [7], adapti v e control [8]-[13], passi vity- based control [2], [14], [15], and slide-mode control [16], [17]. W e will discuss these and other approaches in details in section 3. There are tw o main types of teleoperation systems: unilateral and bilateral teleoperation system s. When the transmission of motion goes from master to sla v e it is called a unilateral teleoperation system (UTS). On the other hand, a bilateral teleoperation system (BTS) includes transmissions in both the forw ard and backw ard directions between the master and sla v e [18]. Generally speaking, BTSs ha v e more interest among researchers since the y are more used in places that are dif cult for humans such as underw ater v ehicles, airspace J ournal homepage: http://ijr a.iaescor e .com Magdi S. Mahmoud Systems Engineering Department King F ahd Uni v ersity of Petroleum Minerals P . O. Box 5067, Dhahran 31261, Saudi Arabia Email: msmahmoud@kfupm.edu.sa Evaluation Warning : The document was created with Spire.PDF for Python.
236 ISSN: 2089-4856 applications, mining, and remote medical sur geries. Comprehensi v e historical surv e ys on BTSs are presented in [1], [19]. In 2011, Nu ˜ no, Ba sa ˜ nez and Orte g a [2], a tutorial on passi vity based control on BTSs is pro vided. A surv e y on the en vironment and task-based controller system is sho wn in [20]. Also, a re vie w on bilateral teleoperation with force, position, po wer , and impedance scaling is detailed in [21]. In addition to the basic requirements of stability , a major aspect to consider in BTSs is the trans- parenc y . Actually , there is a tradeof f between these tw o concerns mainly as a result of the time delays caused by the communication channel [22]. The delay-induced ins tability of BTSs is one of the main challenges for researchers since time delays e xist in most communication channels. Some major studies based on time- delayed bilateral teleoperation ha v e been done in [12], [23]-[25]. The time delays could be symmetric [26] as well as asymmetric [27]. Some recent studies on the stability analysis of asymmetric time-v arying delay can be found in [28]-[30]. Most of the recent w ork on stability analysis with a symmetric time-v arying delays emplo yed L yapuno v–Kraso vskii functional to formulate the relations among the controller parameters, and the upper bound of time-v arying delays [31], [32]. Linear matrix inequalities (LMIs) were applied to present the stability criteria. So, it can be solv ed to obtain the allo wed maximum v alues of delays. Figure 1. Block diagram of the typical teleoperation system Recently , man y re vie ws on the BTSs were published. In 2018, Gha vifekr , Ghiasi and Badamchizadeh [33], discrete-time control methods of BTSs wer e presented with concern on problems of passi vity , stability , transparenc y , and time delays. The recent control approaches for teleoperation systems while considering internet-based communication, unkno wn time-v arying delay , and model uncertainty were re vie wed [34]. More concern w as gi v en to control algorithms that were applied to nonlinear uncertain systems. The application of predicti v e control schemes in BTSs are discussed including qualitati v e and quantitati v e comparisons among these methods re g arding rob ustness, transparenc y , and stability [35]. In this paper we address the control problem of bilateral teleoperation systems (BTSs) in nominal situations and in the presence of c yber -attacks. The main contrib utions of this paper are summarized as follo ws: A discussion on the methods of modeling BTSs will be presented. A comprehensi v e surv e y on control schemes for BTSs under c yber -attacks is discussed. The literature that considers c yber -attacks in the design of the controller for BTSs is re vie wed. The current and future problems in this eld are discussed. The remaining of this paper is or g anized as follo ws: Modeling of BTSs are discussed in section 2. In section 3, control approaches are presented. Then, the c yber -attacks in BTSs are described in section 4. Finally , Current and future research are listed in section 5. are deri v ed by applying the Lagrangian systems with the actuated re v olute joints such that sho wn in (1): M m ( q m ) ¨ q m + C m ( q m , ˙ q m ) ˙ q m + g m ( q m ) = J T m ( q m ) F m + τ m M s ( q s ) ¨ q s + C s ( q s , ˙ q s ) ˙ q s + g s ( q s ) = J T s ( q s ) F s + τ s (1) Int J Rob & Autom, V ol. 10, No. 3, September 2021 : 235 260 2. MODELING OF BTSS In secti on 2, we will discuss the modeling of BTSs. The dynamics of both the master and sla v e robots Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Rob & Autom ISSN: 2089-4856 237 with the subscripts { m } and { s } denote the master robot and the sla v e robot, respecti v ely . q m ( t ) n and q s ( t ) m are the generalized coordinates, M m ( q m ) n × n and M s ( q s ) m × m are the inertia matrices. C m ( q m , ˙ q m ) n × n and C s ( q s , ˙ q s ) m × m are damping matrices. g m ( q m ) n and g s ( q s ) m are the gra vity forces. τ m ( t ) n and τ s ( t ) m are the applied torques which is deifned sometimes as the control signals. J m ( q m ) n × n and J s ( q s ) n × m are the Jacobian matrices, and F m ( t ) , F s ( t ) n × 1 are the forces e x erted by the human operator and the en vironment. See for e xample [10], [36], [37]. Let us assume that the sla v e robot is a redundant manipulator to obtain semi-autonomous teleoperation and the master robot is non-redundant manipulator for simplicity . Let x ( t ) dened as the state v ector and equal to [ x m ( t ) ˙ x m ( t ) x s ( t ) ˙ x s ( t )] T with x ( t ) and ˙ x are denoting the position and speed of the end ef fector , respecti v ely . And by considering the e xternal forces as disturbances, the general state-space model for system (1) is represented by (2): ˙ x ( t ) = Ax ( t ) + B u ( t ) + B w w ( t ) (2) with A = 0 1 0 0 0 B m M m 0 0 0 0 0 1 0 0 0 B s M s , B = 0 0 1 M m 0 0 0 0 1 M s = B 1 B 2 B 1 = 0 1 M m 0 0 , B 2 = 0 0 0 1 M s , B w = 0 0 1 M m 0 0 0 0 1 M s u ( t ) = F m ( t ) F s ( t ) T , w ( t ) = F h ( t ) F e ( t ) T (3) In equation (2), the state v ariables are x m ( t ) , ˙ x m ( t ) , x s ( t ) , and ˙ x s ( t ) . F or simpli city , the states x m ( t ) and ˙ x m ( t ) are called local state v ariables and x s ( t ) and ˙ x s ( t ) are called remote state v ariables to the master , and con v ersely with respect to the sla v e. The follo wing state-fee d ba ck controllers are designed to obtain the stability for the master and sla v e manipulators as (4): F m ( t ) = K m x m ( t ) ˙ x m ( t ) x s ( t d 2 ( t )) ˙ x s ( t d 2 ( t )) T F s ( t ) = K s x m ( t d 1 ( t )) ˙ x m ( t d 1 ( t )) x s ( t ) ˙ x s ( t ) T (4) where d 1 ( t ) and d 2 ( t ) are the forw ard time delay (the communication path from the master to the sla v e) and the backw ard time delay (the communication path from the sla v e to the master), respecti v ely . Remark 1. It is mor e pr actical to have upper limit on the time delays suc h as positive scalar s ¯ d 1 and ¯ d 2 . So, d 1 ( t ) and d 2 ( t ) ar e bounded as, 0 d 1 ( t ) ¯ d 1 , 0 d 2 ( t ) ¯ d 2 [38]. The controllers F m ( t ) and F s ( t ) is formulated in the follo wing compact system: F m ( t ) = K m x ( t d 2 ( t )) F s ( t ) = K s x ( t d 1 ( t )) (5) By substituting the controllers (5) in the BTS (3), one will obtain the follo wing o v erall cl osed-loop system sho wn in (6): ˙ x ( t ) = Ax ( t ) + B 1 K m x ( t d 2 ( t )) + B 2 K s x ( t d 1 ( t )) + B w w ( t ) (6) Contr ol of teleoper ation systems in the pr esence of cyber attac ks: A surve y (Mutaz M. Hamdan) Evaluation Warning : The document was created with Spire.PDF for Python.
238 ISSN: 2089-4856 3. CONTR OL METHODS FOR BTSS Man y literature discuss the challenges of controlling teleoperation systems [39]. The focus of these surv e ys directed to adapti v e control, w a v e v ariable, and predicti v e control approaches [35], [40], [41]. V arious methodologies were proposed to handle control issues in se v eral applications of teleoperation systems. Some e xamples are: nonlinear trilateral teleoperation systems [42], PD-lik e controller [43], four -channel structure [44], operator dynamics considerations [45], output feedback with force estimation [46], a s table and trans- parent microscale teleoperation [47], e xternal force estimation [48], position synchronization [49], and ne w stability criteria using h ybrid strate gies of position and impedance reection [50]. Figure 2. Control approaches applied to BTSs 3.1. P assi vity based contr ol appr oach P assi v e systems are referred to dynamical systems that ha v e positi v e consumption of total ener gy . P assi vity based control is a control approach that aim s to maintain the passi vity beha vior of the closed-loop system. So, in this approach, the controller is designed and formulated to k eep the balance of the po wer o w and ener gy consumption of the BTS [51]. One of the challenges in BTSs is the occurrence of time delays in the communication channel. In addition to its ef fect on the signal, the time delay could also increase the ener gy in the communication channel [52]. This could af fect the passi vity of the BTS and may lead to instability . T o handle this issue, se v eral control approaches based on passi vity ha v e been proposed. The passi vity-based con- trol approach is suitable and applicable to nonlinear systems, distinctly to mechanical/Euler -Lagrange systems [53]. Let us consider the BTS (1), the input po wer of this BTS in terms of the passi vity is gi v en by [54]: P in ( t ) = x T y (7) where x is the input and y is the output of the system. The input po wer o w in the BTS (1) is formulated as (8): P in ( t ) = ˙ q T m f T e f h ˙ q s . (8) The passi vity of system (1) is assured if and only if a state function E ( q i , ˙ q i ) 0 , ( q i , ˙ q i ) e xists and satises the follo wing inequality [54] as (9): d dt E ( q i , ˙ q i ) P in . (9) Int J Rob & Autom, V ol. 10, No. 3, September 2021 : 235 260 As m entioned in the Introduction, some recent literature discusses the control of BTSs . In 2018, Gha vifekr et al . [33], discrete-time control methods of BTSs were presented with concern on pr o bl ems of passi vity , stabi lity , transparenc y , and time delays. The recent control approaches for teleoperation systems while considering internet-based communication, unkno wn time-v arying delay , and model uncertainty were re vie wed [34]. More concern w as gi v en to control algorithms that are applicable to nonlinear uncertain systems. The application of predicti v e control schemes in BTSs are discussed including qualitati v e and quantitati v e comparisons among these methods re g arding rob ustness, transparenc y , and stability [35]. In section 3, we will discuss se v eral kinds of control approaches applied to BTSs and the feature of each method will be e xplained. These approaches are summarized in Figure 2. Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Rob & Autom ISSN: 2089-4856 239 Man y literature refers to the ener gy storage function by E ( q i , ˙ q i ) and the dissipati v e rate by d dt E ( q i , ˙ q i ) [51], [53]. The pre vious inequality (9) is represented by the follo wing (10): E ( q i , ˙ q i ) q i ˙ q i + E ( q i , ˙ q i ) ˙ q i ¨ q i ˙ q T m f T e f h ˙ q s (10) No w , consider system (1) and se v eral manipulations to describe the torque control signals τ i as (11): E ˙ q i M 1 i τ i ˙ q T m f T e f h ˙ q s E q i ˙ q i + E ˙ q i M 1 i C i q i + g i f j (11) where i { m, s } and j { h, e } . Finally , the passi vity-based control approach is summarized as follo ws: Design a suitable τ i such that (11) holds for all q i and ˙ q i . Man y w orks of literature in v estig ate this control approach. In E. Kamrani [55], a multi v ariable control method with w a v e-prediction and passi vity technique is proposed for real-time systems in the presence of internet delay dynamics. A synchronizati on scheme for nonlinear dynamical netw ork ed systems including delays and uncertainty is designed using passi vity analysis [13]. A tw o-layer control scheme has been discussed for maintaining the stability of BTSs [56]. This scheme consists of one layer to assure the transparenc y and a second one for handling the passi vity of the system. In 2010, K. Hertk orn et al . [57], a general time-domain passi vity control approach is applied for a haptic system with multi-de gree of freedom. Other e xamples on this approach including stabilizing the ener gy or position drift compensation can be found in [58]-[63]. Some adv antages of this approach in the literature include: applying switching method for ener gy dissipation [64] or independent of time-v arying delay [65], the combination of passi vity and transparenc y [56], applying it to non-ideal system [61], applying both po wer -based and ener gy-based approaches [60], and the elimination of system uncertainties [66]. On the other hand, some dra wbacks of this approach include: weak transparenc y [64], the transparenc y is not guaranteed in the passi vity layer [56], limited application and no consideration of delay [61], applied to a system with constant delay [58], and limited random delays [60]. More e xamples on the passi vity-based control approach are [2], [14], [15], [67]-[71]. 3.2. W a v e v ariable-based contr ol The rst introduction of the w a v e v ariable-based control approach w as in 1991 by Nieme yer and Slotine [72]. The follo wing tr ansformation is applied to the reference signals in the BTS, i.e. joints v elocity and e xternal forces ˙ q i and f j , i { m, s } and j { e, h } to obtain a w a v e formulation sho wn in (12) and (13). u m = 1 2 b f h ( t ) + b ˙ q m ( t ) (12) u s = 1 2 b f e ( t ) b ˙ q s ( t ) (13) No w , the ne w signals u m and u s are sent by the communication channel. Additionally , the desired signals for tracking will be f h and ˙ q s , and the communication channel has a characteristic impedance b . So, the o w of po wer in the communication channel between the master and sla v e robots are gi v en by (14): P in ( t ) = f T h ( t ) ˙ q m ( t ) f T e ( t ) ˙ q s ( t ) (14) = 1 2 u T m ( t ) u m ( t ) v T m ( t ) v m ( t ) + 1 2 u T s ( t ) u s ( t ) v T s ( t ) v s ( t ) where v m ( t ) , v s ( t ) are the recei v ed signals and u m ( t ) , u s ( t ) are the reference signals o n the master and sla v e sides of a BTS satisfying by (15) and (16): f h ( t ) = b ˙ q m ( t ) + 2 bv m ( t ) (15) ˙ q s ( t ) = 1 2 2 bv s ( t ) f e ( t ) . (16) Contr ol of teleoper ation systems in the pr esence of cyber attac ks: A surve y (Mutaz M. Hamdan) Evaluation Warning : The document was created with Spire.PDF for Python.
240 ISSN: 2089-4856 The researchers ha v e de v eloped and designed controllers for BTSs in the w a v e domain due to the f act that intrinsic passi vity is preserv ed in the w a v e formulation [73]-[79]. Ho we v er , the passi vity could be unsa v ed in the po wer signals. According to this feature, it is important to b uild and place a passi v e lter at the master robot for estimating the model of the sla v e robot and enhance the force reection in haptic applications. A similar proposal w as pro vided for constant kno wn delays [78], [79]. An incident named w a v e reection occurs when the characteristic impedance of the netw ork channel is not similar to that of the ports of the master and sla v e. Since it has a high impact on the BTS performance, w a v e reection must be handled cautiously in the system design. Normally , the impedance matching element b is used for tuning this feature at both sides of the communication channel. But, the performance of the BTS, mainly the position tracking could be af fected by this parameter . T o o v ercome this scenario and decrease position drift, parameter b is not considered on the master side [79]. T o notice the dif ference between these tw o w ays, the position tracking drifts for the master side when b is remo v ed is gi v en by (17) and (18): q m ( t d 1 ( t )) q s ( t ) = 1 2 b Z t t 2 d 1 ( t ) f e ( η ) (17) and when t is not remo v ed q m ( t d 1 ( t )) q s ( t ) = 1 b Z t 0 f e ( η ) + q s ( t ) . (18) It is noted here that as b increases, the position drift decreases. Ne v ertheless, since more damping is introduced, the BTS will ha v e a poor performance. Some adv antages of this approach in the literature include: the impro v ement of position tracking [73], optimization of w a v e transformation [77], the impro v ement of tracking and po wer optimizati o n [80], dealing with passi v e and non passi v e human model [76], and guaranteed passi vity [75]. The dra wbacks of this approach include: the e xistence of estimation error [73], no consideration of time delay [77], [80], reducing the transparenc y [76], and considering constant time delay only [75]. 3.3. Adapti v e contr ol appr oach The structure of BTSs includes se v eral important parameters and model uncertainties to be concerned by designers. These parameters are re v ealed in all of the elements of the BTS sho wn in Figure 1, i.e. the master and sla v e robots, the communication netw ork, the human operator , and the en vironment. The adapti v e control approach is widely applied by researchers to solv e the problem of the e xistence of the uncertainties in BTSs. Moreo v er , a lot of w orks and ef forts were in v ested in this approach due to the linearity of parameters of BTSs as noted in (1). T o understand this method, the general design of the adapti v e controllers for the BTSs is gi v en by (19) and (21). The human operator and the en vironment are pres ented as a general nonpassi v e nonhomogeneous force as follo ws (19): ˆ f j ( t ) = ˆ M j ( t ) ¨ x j ( t ) + ˆ D j ( t ) ˙ x j ( t ) + ˆ K j ( t ) x j ( t ) + ˆ f j 0 ( t ) ˆ M i ( q i ) ¨ q i + ˆ C i ( q i , ˙ q i ) ˙ q i + ˆ g i ( q i ) + δ ( t ) = τ i + f j (19) where ˆ M j , ˆ D j , ˆ K j , and ˆ f j 0 are the estimation parameters of the mass, damping, stif fness, and nonhomogeneous parameters of the human operator , i { m, s } represents master and sla v e systems, and j { h, e } represents the human and en vironment, respecti v ely . Remark 2. The same philosophy could be applied for e xternal disturbance r ejection inserting a bounded term r epr esenting the e xternal disturbance δ ( t ) , as (20): δ ( t ) < (20) In this scenario, an adaptive par ameter ˆ δ ( t ) is needed to be added to the contr ol signal in the BTS (1) suc h that (21) and (22): τ = ˆ f j ( t ) + ˆ M ( t ) a + ˆ C ( q , ˙ q ) ˙ q + ˆ g ( q ) + ˆ δ ( t ) (21) Int J Rob & Autom, V ol. 10, No. 3, September 2021 : 235 260 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Rob & Autom ISSN: 2089-4856 241 in addition to the term of the PD contr oller a = ¨ q des K d ˙ e K p e (22) wher e K d , K p > 0 ar e symmetric matrices. Then, after some manipulation and linear parameterizing, the dynamics of the BTS are formulated as (23): ˜ M ( ¨ e + K d ˙ e + K p e ) = Ψ ( q , ˙ q , ¨ q ) ˜ θ ( t ) + ˜ δ ( t ) + ˜ f j ( t ) (23) with the estimation error of the corresponding parameter is indi cated by the sign . So, these estimated parameters ha v e to be handled in the design of the controller to achie v e or maintain the stability and the task performance of the BTS in the presence of uncertainties. The adapti v e la w for ˜ θ ( t ) and ˜ δ ( t ) is obtained by applying the s tability analysis such as the L yapuno v theorem such that the con v er gence of the error signals to zero is achie v ed and the estimation errors are bounded [34]. Additionally , one of the general methods in adapti v e control des ign is to utilize the linearity in the parameters of the BTS (1). Let the combined error signals of both position and v elocity of the master and sla v e be dened as (24) and (25): ϵ m = ˙ e m + P m e m (24) and ϵ s = ˙ e s + P s e s (25) where P m , P s > 0 , so, the control signals are (26): u i = u id ˆ M i ( q i ) P i ˙ q i ˆ C i ( q i , ˙ q i ) P i q i = u id Ψ i ˆ θ i , i { m, s } (26) and the closed loop BTS will be (27): M i ˙ e i + C i e i = Ψ i ˜ θ i + u id f j , i { m, s } , j { h, e } (27) will guarantee both of the errors’ asymptotic con v er gence and the parameter estimation errors ˜ θ i to be bounded. Also, the L yapuno v analysis or an y other stability analysis could be used to deri v e the adaptation la ws ˙ ˆ θ i . Remark 3. The adaptive contr ol appr oac h could be combined with intellig ent and some nonlinear algorithms lik e fuzzy contr ol and neur al network (NN). This featur e allows r esear c her s to design ne w algorithms to enhance and impr o ve the performance of the BTSs whic h mak es it one of the most powerful contr ol appr oac hes. The designed control system using the combination of algorithms leads to a more rob ust and ef cient BTS. Moreo v er , it is not necessary to ha v e the e xact model and precise information about the BTS. Also, these algorithms allo w o v ercoming other restrictions in the BTSs, for e xample, long time-v arying delays, input saturation, pack et loss, force reection error , and position drift. An adapti v e switched control considering passi v e and non-passi v e e xternal forces, actuator saturation, and unkno wn dynam ics is designed for BTSs including asymmetric time-v arying delay [81]. The obtained scheme has the capability of adapti v e systems whi ch is indicated by the achie v ed bounded position tracking error of the BTS. An adapti v e fuzzy control system w as de v eloped to attain the state-independent input-to- output stability for a multilateral asymmet ric teleoperation system [82]. Also, similar w orks are found such as [83]-[85]. Chen et al . [86], an adapti v e nite-time control method is presented using subsystem decomposition. The stability is guaranteed by applying the L yapuno v-Kraso vskii analysis and bounded tracking error is ob- tained. Rob ust adapti v e techniques were proposed for considering unkno wn parameters the uncertainties of the BTS [87]. More e xamples on this approach are [8]-[13], [41]. Some adv antages of the adapti v e control approach in the literature include: considering non-homogenous and state independent input/output stability [81], [82] non-passi v e input forces [24], considering drift dif fusion impro v ed haptic [88], and multi robot sla v e and input saturation [89]. The dra wbacks of this approach include: chattering of torque [81], [82], [89], does not consider time delay [88], constant time delay [90], and non assymptotic stability [24], [87]. Contr ol of teleoper ation systems in the pr esence of cyber attac ks: A surve y (Mutaz M. Hamdan) Evaluation Warning : The document was created with Spire.PDF for Python.
242 ISSN: 2089-4856 3.4. Rob ust contr ol appr oach The main objecti v e when applying the rob ust control approach is to consider the w orst conditions which ha v e ne g ati v e ef fects on the stability and the performance of the BTSs. So, the control scheme is called rob ust when it maintains the stability and the performance of the BTSs af fected by disturbing f actors. In addition to the elements af fecting normal systems such as the uncertainties in the model, ne glected dynamics, and e xternal disturbances, the netw ork-induced uncertainty is the most tragic element tha t af fects the stability and performance in BTSs. In mathmatic w ords, the rob ust control problem is minmax problem such that it is a minimization of the fraction y δ = y / δ while the term y x = y / x is maximized, with y δ refers to the disturbances’ contrib ution δ in the output of the system y , with the desired input x . . is an y Euclidean norm function. Ideal v alue for y δ is 0 and for y x is 1. The sliding mode control (SMC) is the most f amiliar rob ust control approach. In this technique, the controller is designed to achie v e the con v er gence of the error signal to a predetermined sliding surf ace. This surf ace is a dif ferential equation for the error which has a solution lying on a con v er gent set or point. One w ay for SMC is to dene the error between the motion of the master and sla v e robots as e = x s K x m . So, the sliding surf ace is gi v en by (28): s = ˙ e + λe (28) in which λ > 0 is a constant used for determining the features of the surf ace and the error’ s rate of con v er gence, and K > 0 is a coef cient matrix for scaling. Remark 4. A dissipation condition must be met to guar antee the r ob ust stability as the following foe e xample: s ˙ s γ s < 0 wher e γ > 0 is a constant. Either lar g e gain or c hattering contr ol signals will be g ener ated to obtain this featur e . A chattering-free SMC approach is presented to achie v e a rob ust performance of a BTS in unce rtain conditions [91]. The number of the needed sensors is reduced by applying a pseudo-sensorless approach which also decreased the uncertainties af fecting the BTS. A PD controller is proposed for specic delayed BTSs using the linear matrix inequalities (LMI) technique [92]. The BTS has been stabilized re g ardless of the occurrences of delays in the communication channel. But, the presented PD controller w as applied only on the sla v e robot and designed for the constant time delay . The rob ust control approach is more suitable for linear systems because of the norm-based a nalysis and calculations. So, Some techniques of the rob ust control lik e H 2 , H , and µ -synthesis are applied for linearized time delay BTSs. These techniques are normally used in the frequenc y domain to handle the w orst case scenario, uncertainties, and the upper limit of the delay of the linearized BTSs. Some adv antages of the adapti v e control approach in the literature include: reducing time sensi ti vity and con v e x optimization using H approach [93], achie ving optimal transparenc y using H 2 H approach [94], maintain passi vity [86], considering time v arying uncertainties [95], and simultaneous stabilit y and trans- parenc y [96]. The dra wbacks of this approach include: considering linear model with constant delay [93], [94], [97], linear model without time delay [95], occurrences of chattering torque [86], [96], constant time delay [98]-[100]. 3.5. Neural netw ork-based appr oach One of the w orst features af fecting the stability and the performance of BTSs is the e xistence of time- v arying delays. Some literature considering this situation by considering the input saturation with time-v arying delay using adapti v e fuzzy control method [101] or applying tw o controllers, a PD-lik e and P-lik e controllers to handle constant delay and time-v arying delay , respecti v ely [102]. The neural netw ork (NN) based approach for controlling BTSs has a great v alue due to its high poten- tial to learn and parallel adapti v e processing. These features mak e the NN more practical for comple x nonlinear BTSs and suitable to deal with the time delays and uncertainties e xisting in the BTSs. The basic idea of NNs comes from the neuron system in the human body . So, similar to the cell model, the y ha v e dendrite weights w j , a nonlinear function σ ( . ) which is normally called the acti v ation function of the cell as sho wn in (29), and Int J Rob & Autom, V ol. 10, No. 3, September 2021 : 235 260 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Rob & Autom ISSN: 2089-4856 243 a ring threshold w 0 . Let the output signal of the n -dimensional input signal x ( t ) R n to be y ( t ) . So, the neuron is represented by the follo wing (29): y ( t ) = σ n X j =1 w j x j ( t ) + w 0 . (29) The insight of the neuron is e xci tatory synapses when w j > 0 and it is inhibitory synapses when w j < 0 [103]. The acti v ation function σ ( . ) is introduced separately and relat ed to the requirements of the current application. It also sho ws the beha vior of the cell in the case that y ( t ) is in its range. Also, since the deri v ati v e is required in learning algorithms for man y cases, σ ( . ) has to be dif ferentiable. Both the input v ector and weights v ector are assigned to deri v e the output signal. So, we will get the follo wing (30): y ( t ) = σ [ w 0 , w 1 , . . . , w n ] [1 , x 1 , . . . , x n ] T = σ w T x ( t ) + w 0 . (30) Remark 5. F or the case of the multicell model of a neur on consists of r cells, with an input signal x ( t ) and an output y m ( t ) , the system is r epr esented by the following (31): y ( t ) = [ y 1 , y 2 , . . . , y m ] T = σ W T x ( t ) + w 0 (31) in whic h the weight matrix is given by (32): W = [ w ij ] , i = 1 . . . r , j = 1 . . . n. (32) The features of the NNs allo w researchers to apply them in a wide range of applications to solv e the nonlinear and uncertainty problem for both control and estimation problems especially in robotics and BTSs [66], [89], [104], [105]. A comprehensi v e discussion on the application of NNs for robotics control is presented in [103]. NNs were used in BTSs for approximating and dealing with delays and uncertaint ies of the com- munication channel. So, NNs are applied for modeling the delay in the system d f ( t ) and d b ( t ) , or the total transmission between the master and the sla v e. Let us assume that the position signal of the master robot x m ( k ) has to be transmit ted to the sla v e side at the sampling time k . One can de v elop the NNs for modeling the netw ork to estimate this process as follo ws (33): ˆ x m ( k ) = T NN ( x m ( i )) , i [1 , 2 , . . . , k 1 , k ] (33) with T NN ( . ) represents the NN model obtained by (34): T NN ( x m ( i )) = σ W T x m ( i ) + w 0 . (34) The local controllers on the master and sla v e sides are designed to pro vide a prediction on the trans- mitted signals using the estimator model of the communication channel. So, the control signal in this system will be more ef cient in comparison to the normal case where the signals recei v ed by the controllers could be af fecting by the delay or the distortion in the system [34]. A prediction algorithm using adapti v e linear NN w as de v eloped for the application of an internet time delay system [106]. Auld et al. [107], a Bayesian NN w as implemented for classifying the internet traf c without information about the IP . The proposed method pro vided 95% accurac y after testing it on a real training NN for 8 months. A recurrent NN w as utilized for modeling and predicting the Internet end-to-end delay [108]. The recurrent NN w as trained and v alidated using the discrete-time data obtained by measuring the delay between tw o dif ferent nodes. NN approach w as also implemented for modeling traf c o w in self- similar computer netw orks which is statically do not depend on time [109]-[112]. Other e xamples including further details could be found in [113]-[115]. In addition to the aforementioned applications, NNs were implemented in the local cont rollers on both sides of the BTSs. A better performance w as obtained by combining the NN approach with other control methods, mainly the adapti v e control method [66], [89], [104]. Another application of the NN in the BTSs is to Contr ol of teleoper ation systems in the pr esence of cyber attac ks: A surve y (Mutaz M. Hamdan) Evaluation Warning : The document was created with Spire.PDF for Python.
244 ISSN: 2089-4856 utilize the NN in designing the control system for estimating the delayed or disturbed signals and for protecting the system simultaneous ly when it is subjected to uncertainties and latenc y [34]. Ho we v er , it is w orth kno wing that the parameters of the NN structure such as the number of layers, the number of neurons, etc. af fect the accurac y of the estimation process. So, the application of the NN control approach handle the uncertainties and delay indirectly by gi ving the best possible result within the unkno wn circumstances of the system. A radial basis function (RBF) NN-based procedure w as proposed for handling v ariable delays in teleoperation system [36]. The model contains linear viscous and Coulomb v elocity-dependent frictions. But, it is required to measure the acceleration or to estimate it, and thi s is not practical in man y cases. The follo wing RBF Gaussian functions were applied for approximating the dynamics of the BTS (1): Φ n ( x ) = e xp 1 2 H 2 n x C n 2 (35) where H n is the width and C n is the center for the n th neuron. Dyn i ( x i ) = W T i Φ i ( x i ) + ξ i ( x i ) (36) where x i = [ ¨ q T i , ˙ q T i , q T i , ¨ x T i , ˙ x T i , x T i ] T . Therefore, the follo wing approximated dynamic are implemented in the control system to o v ercome both of the unkno wn bounded terms and the in v erse dynamics by (37): Dyn i ( x i ) = f b i ( q i , ˙ q i ) + g i ( q i ) + C i ( q i , ˙ q i ) ˙ q i + M j ¨ x i + B j ˙ x i + K j x i + D i ( t ) . (37) Thus, the control signal τ i is calculated for the follo wing system sho wn in (38): M i ( q i ) ¨ q i + C i ( q i , ˙ q i ) ˙ q i + g i ( q i ) + f b i ( q i , ˙ q i ) + D i ( t ) = τ i + M j ¨ x i + B j ˙ x i + K j x i + f j 0 + Dyn i ( x i ) = Ψ i ( x i i . (38) The desired control signal for system (38) when the function Dyn i is ideally approximated, is obtained using (39): τ i = M i ( q i ) ¨ q i K D i ( ˙ q i ˙ q i ) K P i ( q i q i ) (39) with positi v e-denite g ain matrices K D i and K P i that by substituting in (38) leads to a Hurwitz f u nc tion sho w in (40): ¨ e i + K D i ˙ e i + K P i e i = 0 . (40) The NN-based approximations are inherently intelligent enough to use the pre vious e xperiences of the BTS for learning, although the y ha v e an error in real applications. The accurac y of the approximation could be increased by selecting lar ge quantities for NN nodes n in (35) for function Dyn ( x ) o v er a compact set of the input x x R x . Other methods of control were combined with NN to decrease the una v oidable estimation error . A neural adapti v e control w as proposed for a single-master multi-sla v e BTS [89]. The use of the NNvto model of the unkno wn nonlinear plants helped to handle t he uncertainties in the dynamics of and the input of the BTS. But, se v eral assumptions were applied in this scheme such as the singularity-free motion of the sla v e robot, the rigidity of the object, and rigid attachment between sla v es’ end-ef fector and the object. Consider the signals of the combined error to be (41): r m = ˙ x m + Γ m e m w , ˙ x mr = Γ m e m r s = ˙ x s + Γ s e s , ˙ x sr = Γ s e s (41) with error deniti o ns e m ( t ) = x m ( t ) x s ( t d b ( t )) and e s ( t ) = x s ( t ) x m ( t d f ( t )) , and matrices Γ m , Γ s > 0 . Then, the adapti v e control signal is selected as (42): u m = β m ˆ Θ T m Φ m ( Z m ) sgn ( r m ) u s = β s ˆ Θ T s Φ s ( Z s ) sgn ( r s ) (42) Int J Rob & Autom, V ol. 10, No. 3, September 2021 : 235 260 Evaluation Warning : The document was created with Spire.PDF for Python.