a novel scheme for human-friendly and time-delays robust neuropredictive teleoperation

Journal of Intelligent and Robotic Systems 25: 311–340, 1999. © 1999 Kluwer Academic Publishers. Printed in the Netherlands. 311 A Novel Scheme for Human-Friendly and Time-Delays Robust Neuropredictive Teleoperation PLATON A. PROKOPIOU and SPYROS G. TZAFESTAS Intelligent Robotics and Automation Laboratory, Department of Electrical and Computer Engineering, National Technical University of Athens, Zografou, GR-15773, Athens, Greece; e-mail: tzafesta@softlab.ece.ntua.gr WILLIAM S. HARWIN The Human-Robot Interface Laboratory, Department of Cybernetics, University of Reading, Whiteknights, PO Box 225, Reading, RG6 6AY, UK (Received: 15 October 1998; in final form: 15 February 1999) Abstract. A novel Neuropredictive Teleoperation (NPT) Scheme is presented. The design results from two key ideas: the exploitation of the measured or estimated neural input to the human arm or its electromyograph (EMG) as the system input and the employment of a predictor of the arm movement, based on this neural signal and an arm model, to compensate for time delays in the system. Although a multitude of such models, as well as measuring devices for the neural signals and the EMG, have been proposed, current telemanipulator research has only been considering highly simplified arm models. In the present design, the bilateral constraint that the master and slave are simultaneously compliant to each other’s state (equal positions and forces) is abandoned, thus obtaining a “simple to analyze” succession of only locally controlled modules, and a robustness to time delays of up to 500 ms. The proposed designs were inspired by well established physiological evidence that the brain, rather than controlling the movement on-line, “programs” the arm with an action plan of a complete movement, which is then executed largely in open loop, regulated only by local reflex loops. As a model of the human arm the well-established Stark model is employed, whose mathematical representation is modified to make it suitable for an engineering application. The proposed scheme is however valid for any arm model. BIBO-stability and passivity results for a variety of local control laws are reported. Simulation results and comparisons with “traditional” designs also highlight the advantages of the proposed design. Key words: neuropredictive teleoperation, human arm model, time-delays compensation, hypothet- ical neural input/electromyograph prediction, enhanced Yokokohji–Yoshikawa scheme. 1. Introduction Although master-slave systems have been investigated since the first years of ro- botic research systems [1–21], there is still a considerable amount of interest in them. In addition to the traditional application fields (e.g., space and underwater exploration, hazardous industrial applications), new fruitful areas have emerged, including the entertainment [17] and health care industry [22]. The latter is of particular importance, because it can positively affect many persons and in some 312 P. A. PROKOPIOU ET AL. areas will constitute a reasonably big market that can afford expensive sophisti- cated systems. While telesurgery is a widely advertised and promising application, telemanipulators have already been employed in rehabilitation and secure a greater and sooner impact on the life of many impaired persons. However, an important drawback for such applications lays in that, although current telemanipulator de- signs succeed in performing their tasks, they are often accused of being tiring, slow, unable to cope effectively with large time delays and providing an obscure feeling of the telemanipulated objects [11, 14]. We believe this in part results from the fact that the modeling of the operator arm and the brain decision tactics, due to their complexity and the ambiguity of models previously available, are either neglected or overly simplified. Thus the main aim of our research was to investigate how to exploit a model of the human arm to enhance telemanipulation. Traditional con- cepts seem to have reached a maturity level, with little improvement possibilities, other than developing semi-autonomous and time-delays robust schemes. A key feature of our research is the use of the neural input to the human arm (NI hereafter) as an observable variable, either measured or estimated through an inverse model of the arm. We consider an “extended teleoperator system”, com- prising of the manipulators and the operator arm (reflex and local neural loops included), whose input is the NI rather than the “macroscopic” variables of position and force. Given the current level of research, which is outlined in the next Section, this is a realistic advance. Another basic feature of our designs is the prediction of the operator’s arm movement as a response to the NI, which enables the compensation of time de- lays in the communication channel. Three main groups of techniques robustifying against time delays have appeared up to date. The first is based on the use of predictive displays for the slave and the remote environment. The future slave state is calculated, so that the operator effectively interacts with an adaptable model of the remote site. Since this model has to be both dynamical and visual, complicated graphics and image processing, as well as dynamics identification techniques are necessary, so that relevant applications are computationally intensive and problem- atic for a totally unstructured and unknown environment. The second group entails the use of wave variables to form a passive communications channel [9, 15]. These provide stability but alter the force fed back. A final solution is to use supervisory control [13], which leads to a noncontinuous form of teleoperation, i.e., it changes the basic specification of manipulating as close to physical as possible. A thorough survey of all of the above can be found in [13, 23]. Recently, significant interest was focused on variable time delays, such as those arising from the Internet [10, 18, 19]. The scheme proposed in this paper, effectively uses a prediction of the master state only (which can be done with much more accuracy than predicting the slave and the remote environment), and incorporates this in a stable force-feedback scheme. It will also be shown that it cancels the computational burden of visually represent- ing the slave future state, needed in predictive displays: the slave-side cameras’ image is enough for optical feedback, since it turns out to be synchronized with the HUMAN-FRIENDLY TIME-DELAYS ROBUST NEUROPREDICTIVE TELEOPERATION 313 master. This is to the best of the authors’ knowledge an unexplored approach. In [13] two early efforts predicting the control input along with the rest of the system state are reported for nontelemanipulation tasks, but are there judged inadequate for telemanipulation. Attempting to exploit the two key ideas outlined above and the opportunities offered by the peculiarities of the arm physiology, a novel “from scratch” teleoper- ator design, called “neuropredictive teleoperation” (NPT) scheme, was developed. An effort to use a human arm model in a conservative way, by providing a classic scheme [7] with robustness to time delay was presented in [5] and named “the enhanced Yokokohji and Yoshikawa scheme”. This effort confirmed that the com- bination of employing a predictor module and the NI, can render classical designs obsolete: through a “traditional” design process, it was shown that the mechanical part of the teleoperator reduces under perfect prediction to an open-loop system. This convinced us to abandon the overly conservative bilateral constraint whereby the two robots are simultaneously compliant to each other’s state (equal positions and forces in both sides). As a result a succession of only locally controlled mod- ules and a much simpler control problem are obtained . The compensation of time delays of up to 500 ms was also made possible.  The paper is organized as follows: in Section 2 some human arm models are reviewed and explained how these are currently employed in the teleoperation area. Additional possibilities are also discussed. In Section 3 the proposed time delay compensation method is presented. The main topic of the paper, i.e., the Neuropredictive Teleoperation architecture is analyzed in Section 4. In Section 5 the enhanced Yokokohji and Yoshikawa scheme is presented, and simulation re- sults and comparisons are reported in Section 6. Some conclusions follow and an Appendix describes the Stark model of the human arm together with some modifications that allow robust inversion and simulation. 2. Employing a Human Arm Model in Telemanipulation 2.1. HUMAN ARM MODELS Preliminary attempts to model the human arm or muscles were made in the 1930’s [24], and since then models of human and animal limbs continue to appear with increasing level of sophistication. These can be divided in two broad categories: anthropometric (or biochemical) models [25, 26], which represent each part of the arm physiology separately, and neural networks or other “black-box” models [27 – 30] which fit to experimental data a global mathematical model, i.e., an equivalent but not identical, detailed system. The first type is useful to study the human phys- iology and invent cures or ways to imitate it, whereas the second type, through its simplicity, is suitable for control applications and can be adaptive towards injury,  This paper summarizes and extends previously presented work by the authors [4, 5, 11, 12]. 314 P. A. PROKOPIOU ET AL. illness or other natural variations. “Engineering” simplified models, for qualitative analysis, have also appeared [31]. Although models of arm impedance, based on only the measured “macroscopic” variables (arm force and position) do exist [32], it is the nervous input (NI) of the muscle that is used as input in most cases. The neural signal actually measured, termed electroneurograph (ENG), has to be rectified and further processed to be- come suitable for use within a model (see, e.g., [33] and references therein). In this paper, the acronym HNI (which stands for Hypothetical Neural Input) will be used to refer to this processed signal. In most bioengineering applications the signal measured from surface electrodes, the electromyograph (EMG), is preferred, since it does not require invasive methods. However, it is poorer in information. Measurement devices for EMG and ENG, e.g., surface and implanted electrodes, are available, although (especially for the latter) inneed of improvement. Implanted electrodes have indeed been used on humans for both stimulation and neural signal measurement successfully over a large period (e.g. [33]). The simplest arm model, is the easiest to exploit in a teleoperator implementa- tion, although not necessarily the best. Bobet et al. [29] proposed one of the most convenient models: a simple 2nd-order linear model with time varying parameters. Slightly more complex is the discrete-time model of Bernotas et al. [30]. In both papers, the parameters are bounded, with small variations, and the system poles are always stable. If these parameters are considered independent of the muscle input and output, BIBO-stability is guaranteed by applying theorems of “almost time-invariant systems” [11, 34]. If not, total stability theorems are expected to yield this result. More accurate is a Hammerstein model recently proposed by Efranian et al. [27]. This is also BIBO-stable, as verified with similar arguments. Gollee et al. [28] employed Local Model Networks, comprising of a set of linear submodels with regions of application determined by a Gaussian function. Since the linear submodels are stable and the Gaussian function is bounded, the network is BIBO stable. This model, or another neural network one, is ideal for engineering applications. In our work we used the Stark model of the human muscles, [26, 35] adapted for the arm with parameters from [31], because it is well-established and, being anthropometric, can be easily simulated and give valuable insight. The control schemes outlined in the sequence are however valid for any human arm model. The Stark model as well as some improvements we deemed necessary are outlined in the Appendix. A central point of interest for our research, is the way the human arm is con- trolled by the brain. To obtain insight on the task of telemanipulation, we can think of the human operator as comprising of two parts: the cognitive part,in- corporating the brain’s situation analysis and decision making abilities, and the “subordinate” neuromuscular and sensory systems, which respectively execute the brain commands and provide sensory feedback to it. Our designs were inspired by the well established physiological evidence that the brain, rather than controlling HUMAN-FRIENDLY TIME-DELAYS ROBUST NEUROPREDICTIVE TELEOPERATION 315 Figure 1. Superviso- ry control of the hu- man arm by the brain. The diagram is as in [23], except that the labels are changed. Figure 2. Teleoperation through time delay and predictor. exp(−sT p /2) denotes delay due to transmission through the communications channel. exp(+sT p ) denotes prediction. Other blocks are free of delay.  is the neural input. Hat  denotes estimate. X ss is only used to illustrate the signals’ timing. the movement on-line, “programs” the arm with an action plan of a complete move- ment, which is then executed largely in open loop, regulated only by local reflex loops [36]. Thus, to think in engineering terms, the brain acts as a “supervisory controller” of the “semiautonomous” arm, in exactly the sense proposed in [23] (Figure 1). This is yet another interpretation of the physiology with engineering tools [8, 37, 38] which, being more general than previous ones, enables the merging of any biological model and all brain functions into a standard control architecture. 2.2. HUMAN DYNAMICS IN CURRENT TELEOPERATOR DESIGNS In the field of artificial intelligence, extensive research has been done on the cog- nitive abilities of humans and how to mimic some of them in artificial systems, employing techniques such as neural networks, expert systems and fuzzy systems. Results of interest to teleoperation can be found in [23]. A literature survey of teleoperation-related research [11] has revealed that the human arm dynamics and the decision making process of the human brain are neglected or excessively simplified. Only three research groups [3, 8, 16] attempted to model the full process. However, both the human arm and brain were reduced to simple closed-loop linear models. These may only be valid for a narrow set of operating conditions, and do not comply with the semiautonomous physiology described above. Although they acknowledge that the system is in reality more complex, most papers adopt for simulations of the human and for stability analysis 316 P. A. PROKOPIOU ET AL. a simple, time invariant, spring-damper-mass model for the human arm (exceptions include [1]).  Stability and passivity results are based on a (usually underlying) assumption, which is best summarized in [7] by the statements: 1. The operator’s input τ op is independent of the state of the master–slave system. 2. In other words, the operator does not generate τ op that will cause the system to be unstable. Thus current designs ensure that the mechanical part of the system is stable or passive and rely on the operator to use it successfully, the same way he can use a “low-tech” passive tool such as a hammer. This is effectively the “Extended Physi- ological Proprioception” concept [39] found in rehabilitation literature. According to the discussion in Section 2.1, the first statement above only holds for the interval that the arm functions open-loop, but is obviously generally inaccurate. The second statement is a well-established empirical observation. The method of ensuring the stability/passivity of only the “rest of the system” (human arm and mechanical part), is actually the best we can do, as also argued in Section 2.3. Regardless of whether overall stability is claimed shown or not, the arm models used are too simple. This can be attributed to the complexity and inaccuracy of models previously available, the focusing of the papers on the tedious “mechanical” part design and the fact that experiments demonstrate stability. Since engineering is an applied science, practical success is sometimes enough to make a technique acceptable, despite existing gaps in the theoretical analysis, as can be seen in the ubiquitous applications of Neural Networks. The poor modeling of the human is, we believe, one of the causes of the “tiring” and “sluggish” performance often reported, the other major cause being the bilateral constraints discussed in the following paragraphs. Needless to say that excessive simplification can lead to am- biguous arguments. For example, in [3] it is shown that even a small time delay in the vision feedback path can cause instability. Although instability indeed incurs in experiments, one should not ignore that the brain performs some kind of predictive control [37]. On the other hand, much insight as well as significant performance improvements can result by exploiting recent physiological research. 2.3. POSSIBLE ARCHITECTURES A complete teleoperator analysis and design would take into account both the cog- nitive and musculosceletal parts of the human operator. However, an analysis of the cognitive part would not only be tedious but even possibly philosophically futile (can a human build a perfect, detailed model of itself? If not, are the unmodeled parts really irrelevant to the teleoperation?). We maintain that cognitive models are not needed for teleoperator design: if we could predict the brain’s functions, a (semi-) autonomous robot acting according to this brain model would be more suit-  Recently Peñ ´ ın et al. [10] summarized the relevant literature similarly. However, their perspec- tive totally differs from this paper’s, since instead of opting for more complicated arm models like this paper, they identify simple models of the brain. HUMAN-FRIENDLY TIME-DELAYS ROBUST NEUROPREDICTIVE TELEOPERATION 317 able. Thus, mental functions have to remain uncompensated. Of course, no univer- sal stability can be guaranteed with the unmodeled brain in the control loop, unless it is at least assumed passive or BIBO-stable. This is far from being disastrous, as the human is known to have a stabilizing role. It is desirable though to keep the unmodeled area as small as possible. By incorporating a human arm model in the design, the scheme reported in this paper reduces the brain’s contribution to a state- dependent input signal. All power-generating mechanisms (robots and muscles) are under the designer’s reach. Thus, the uncertainty which the presence of the human brings is “pushed” to a higher level, leaving less “uncharted” areas. According to the argument at the beginning of this section, this is the best one can get. A straightforward way to exploit a human arm model in telerobotics, is to use it in conjunction with existing architectures to produce more valid (in the sense that fewer assumptions are needed) stability results. No alteration to previous de- signs is made, other than replacing the usually simple model with a more accurate one and checking for stability. Such an approach is not considered in this paper. A more fruitful attempt would include the adaptation of current architectures them- selves. As a first step, we could retain the basic lines of traditional design, such as the master–slave–master loop (see Section 4.1), or even the schemes themselves, and just add the model in the part under designer’s reach, so that he can exploit it for control. This implies that the input to the “extended teleoperator system” would then be the neural command (NI) to the muscular system (reflex and lo- cal loops included), rather than the “macroscopic” variables of arm position and force. This is still a conservative approach, since the fundamental specifications remain unchanged, but it is based on well studied existing schemes. An example is the enhanced Yokokohji and Yoshikawa scheme discussed in the Introduction and outlined in Section 5. A yet more fruitful approach would be to re-examine the basic teleoperator design concepts and try to exploit the HNI or EMG input, the arm model and the relevant conclusions of biomedical research through novel schemes. The work reported in this paper follows this line of research. Previous research has regularly thought of the HNI (usually expressed as an “intentional” force acting on a Linear Time Invariant (LTI) arm model) as the starting point of the design but only for the stability analysis. However, once an arm model is known, the HNI signal can also be literally used as a control variable. It can be either directly measured or estimated from the arm position and force by an inverse (mathematical or iterative) model (where it exists). To accommodate for the inevitable errors, robust or adaptive control laws have to be used. The more the relevant bioengineering research matures, the milder the specifications imposed on the controller will be. For easier application, models utilizing the EMG as input can be used. In a Functional Electrical Stimulation (FES) context [27], it is argued that these models are more accurate, since they take into account more general factors (more muscles, other sources of simulations, disturbances, etc.). However, for a thorough teleoperator system analysis, the HNI is needed. Alternatively EMG can 318 P. A. PROKOPIOU ET AL. be used to estimate the HNI (with relevant models suggested in [27, 40]) instead of the possibly harder to get muscle inverse model. However, the achievable pre- diction horizon is then expected to be smaller. The gain of using the HNI or EMG rather than the arm position and force lays in the enhanced stability analysis and the short term predictability of them. The designs reported in the sequence will exploit the HNI. 3. Time Delay Compensation through Prediction A basic feature of our designs is the use of a module that predicts the operator’s arm movement as a response to the HNI. As detailed in [41], it is preferable to make a prediction of the HNI/EMG signal, rather than the model output,since it has a simple form and some well studied characteristics: it is a three pulses’ sequence, modeled by varying the amplitude or the period. It is a square or trian- gular waveform, either continuous or “spiky” modulated by a square or triangular function, whereas the rectified EMG is (roughly) sinusoidal [11]. Such a signal is indeed easily predictable. Depending on the model, this is accomplished either by simulation (“running the model forward”) or by modifying it to a predictive formulation (i.e., by analytical calculations). This predictor offers significant advantages over conventional architectures. A di- rect result is that it immunizes a conventional architecture to time delays up to half the prediction horizon (i.e., up to T p /2). The key idea is to command the slave robot to follow the predicted command, so that it is “ahead in time” from the master. As illustrated in Figure 2, after the two transmissions of signals through the communications channel, the reflected slave position/force has the same time index as the local master variables. This way the “feel” of teleoperation is natural and human-friendly, since the variables are simultaneous, and also not altered by the algorithm, as in existing approaches [9, 15]. This is exactly what happens when a human manipulates by his own hands, i.e., the scheme is transparent. If the time delay is smaller than T p /2, then additional artificial delay (buffering) has to be introduced. Thus variable time delays, arising, e.g., through the Internet, can also be accommodated. To also cover the time needed for measurement or estimation of the neural input and the computations, this limit has to be lowered. Even for small time delays, this prediction can be helpful in providing us time for control error corrections and compliance. Of course, the fidelity is affected by the human arm model prediction accuracy. A central problem is specifying an horizon T p for reliable prediction. A clear upper limit is posed by the frequency with which the brain changes its “program”, its predefined sequence of pulses. Such an upper limit is estimated as about 1 s, whereas a safe limit is 500 ms [35]. Within this limit a partially predictable set of three pulses can be expected. This would require good prediction of the duration of each pulse or robust control laws, with fast correction of prediction errors. Given the margin of 500 ms, it is reasonable to think that such control tactics are feasible. HUMAN-FRIENDLY TIME-DELAYS ROBUST NEUROPREDICTIVE TELEOPERATION 319 To be safer, we can resort to short term predictions of HNI/EMG and “invest” on the activation time of the muscle as a response to HNI. Since this is modeled as a 1st-order linear system with typical time constant 50 ms [26], its step response to the square pulse input used in [26] will have settled after around 200 ms. So, after measuring the HNI, one can predict the muscle response for ≈200 ms. Considering typical values [23], with 0.5 s prediction, we can “erase” time delay for many earth-to-earth applications (e.g., telesurgery through dedicated communi- cation lines or a fast Internet link), or underwater applications for depth of 200 m. With 1 s prediction even single-link earth-to-orbit applications are immunized. Employment of hints about the intended human movement or combination with existing techniques could increase the applicability of the method. For contact tasks, the prediction can be made with the nominal arm parameters, i.e., without load, and then rely on the slave controller to provide compliance. However, since the brain calculates the NI expecting a certain load, a model of the manipulated object was also included in the predictor. Alternatively, as mentioned in Section 4.2.4, the measured slave force can be used as a substitute of it. If the human arm model and predictor modules can be placed at the slave side controller (Figure 3), an object representation does not induce heavy calculations: due to the collocation and time-coincidence of the model output and the environment, mea- sured environment forces can replace predicted ones. This point is made clearer by considering the sequence of events: the measured HNI arrives to the slave delayed by T p /2 s. As soon as a safe prediction can be attempted, the HNI and the resulting slave state are calculated for the prediction horizon, requiring a local only environment model. Then the slave is servoed to the predicted state. After this “future master” state is reached (soon after), the HNI needs only to be predicted one-step-ahead from the previous prediction, so that the measured environment and slave states or an easy one-step-ahead prediction of them can be employed. When the input prediction is judged to be erroneous, based on information that keeps coming from the master, correcting calculations for the whole or less than the prediction horizon are made. This would entail a revision of the path already traveled, so that mostly lately measured environment data can be employed. In summary, a simple (e.g., linear) and local environment model is only needed from time to time. In addition, the plain camera image transmitted back to the master is all the visual information the operator will need, since it coincides in time with the force feedback (Figure 2). Special analysis, such as a world model or graphics to overlay the predicted slave position on the image, is not needed. This simplification is due to the fact that the slave leads in time the master in the real world, not as a computer model. In traditional schemes the master leads the slave, as illustrated in Figure 4. Moreover, choosing the slave side is conceptually equivalent to placing a virtual human arm there. Control reflexes of the arm can be built locally in the slave, to provide compliance. A hardware version of this concept would employ an anthro- pomorphic arm as slave [42]. 320 P. A. PROKOPIOU ET AL. Figure 3. Placement of the predictor: (a) in the master side; (b) in the slave side.  is the neural input. Figure 4. Comparison of time lead with NPT and traditional teleoperation schemes. 4. Neuropredictive Teleoperation (NPT) 4.1. CONCEPTUAL AND MATHEMATICAL DESCRIPTION By employing a model of the human arm, the forward signal path (brain to slave) can be decoupled from the backward one (slave to brain), so that a simpler, open loop, system is obtained. In classic schemes (Figure 5(a)) there are two basic control loops in series:thehardware loop (master–slave–master) and the propri- oceptive loop (brain-operator arm-brain), which are both nested inside the visual loop (slave-brain-operator arm-master–slave). However, since the two latter loops are usually considered to be open (they only close while new command programs are downloaded [36]), the hardware loop is the dominant one, from a designer’s point of view. To further complicate things, traditionally the bilateral constraint that the two robots should be simultaneously compliant to each other’s state (equal positions and forces) is imposed, so that the operator can correctly feel the slave state. This multiplicity of feedbacks, although dictated by the need to use measur- able macroscopic variables,clearly complicates the analysis. Of course, there are also local control loops inside the operator arm, the master and the slave. In NPT (Figure 5(b)), in contrast to classic architectures, we break up the control loop master–slave–master. We conceptually close the control loop in the human brain, mimicking natural human movements. The slave is then directed by the pre- [...]... T (A. 4a) (A. 4b) (A. 5a, b) (A. 6) (A. 7) The original Stark model comprises of Equations (A. 1a, b) below and Equations (A. 2), (A. 3), (A. 5) and (A. 7) 1.25HTL ˙ XL = HTL − Fsl , ˙ Bh + |XL | 1.25HTR ˙ XR = HTR − Fsr , ˙ Bh + |XR | (A. 1a, b) where: the subscripts 1 and r denote left and right muscle, θ position and v velocity of the arm; XL and XR internal model variables; Kp , Jp and Bp , passive arm parameters,... French) Lee, S and Lee, H S.: Modeling, design and evaluation of advanced teleoperator control systems with short time delay, IEEE Trans Robotics Automat 9(5) (1993), 607–623 Prokopiou, P A. , Harwin, W S., and Tzafestas, S G.: Exploiting a human arm model for fast, intuitive and time-delays- robust telemanipulation, in: S G Tzafestas (ed.), Advances in Manufacturing: Decision, Control and Information Technology,... HNI and output the position and force on the human arm, just as in direct manipulation of a physical object The impression to the human brain is more important than the tracking between slave and master, as also discussed in Section 4.1 4.2.1 Mechanical Part With all the control schemes of Section 3, the mechanical part is BIBO-stable for a wide range of parameter values [11] In this section, the final... an inverse human arm model 334 P A PROKOPIOU ET AL (ii) easier and more efficient manipulation Other advantages already mentioned include the expected more natural feeling (human friendly style) of manipulation and the fact that incorporating an operator arm model leads to more valid stability results Finally, the NPT is much simpler to analyze, so that stability can be easily derived and simpler, less... A. , Acaril, R., and Barientos, A. : Human behaviour modeling in Master–Slave teleoperation with kinesthetic feedback, in: Proc IEEE Internat Conf on Robot Automat., Leuven, Belgium, 1998, pp 2244–2249 Durlach, N.: The potential of teleoperation for entertainment and education, Presence 6(3) (1997), 350–351 Sano, A. , Fujimoto, H., and Tanaka, M.: Gain-scheduled compensation for time delay of bilateral teleoperation, ... G.: A computational model if the simplest motor program, J Motor Behavior 25(3) (1993), 153–161 Prokopiou, P A. , Harwin, W S., and Tzafestas, S G.: Variable -time-delays- robust telemanipulation through master state prediction, AIM ’99: 1999 IEEE/ASME Internat Conf on Advan Intel Mechatr., Atlanta, USA, September 19–22, 1999 Hannaford, B., Winters, J M., Chou, C P., and Marbot, P H.: The anthroform arm:... Sheridan, T B.: Space teleoperation through time dealy: Review and prognosis, IEEE Trans Robotics Automat 9(5) (1993), 592–606 Lawrence, D A. : Stability and transparency in bilateral teleoperation, IEEE Trans Robotics Automat 9(5) (1993), 624–637 Anderson, R J and Spong, M W.: Bilateral control if teleoperators with time delay, IEEE Trans Automat Control 34(5) (1989), 494–501 Peñín, L F., Caballero, A. ,... specified for a vector a, “current time” a( t) holds, and all vectors in the equation are undelayed HUMAN-FRIENDLY TIME-DELAYS ROBUST NEUROPREDICTIVE TELEOPERATION 323 The above general system set-up, will be adapted for our scheme by specifying suitable fddm and fdds With alternative choices, systems with varying properties suitable for different applications, result The manipulation will be as close... 44 45 P A PROKOPIOU ET AL Narendra, K S and Annaswamy, A M.: Stable Adaptive Systems, Prentice-Hall, Englewood Cliffs, NJ, 1989 Zangemeister, W H., Lehman, S., and Stark, L W.: Simulation of head movement trajectories: Model and fit to main sequence, Biological Cybernetics 41 (1981), 19–32 Houk, J C., Buckingham, J T., and Barto, A G.: Models of the cerebellum and motor learning, Behavioral Brain Sci... that γM and γiM disappear from the equations 4.3 PASSIVITY OF THE SYSTEM Passivity is a handy concept often used in teleoperation, as an alternative tool to guarantee system performance As summarized in [9], passivity is a sufficient condition for L2 -stability of a system when it is coupled to a passive environment and is receiving bounded input energy In addition, it is a necessary condition for stability . time index as the local master variables. This way the “feel” of teleoperation is natural and human-friendly, since the variables are simultaneous, and also not altered by the algorithm, as in existing. rehabilitation and secure a greater and sooner impact on the life of many impaired persons. However, an important drawback for such applications lays in that, although current telemanipulator. lead with NPT and traditional teleoperation schemes. 4. Neuropredictive Teleoperation (NPT) 4.1. CONCEPTUAL AND MATHEMATICAL DESCRIPTION By employing a model of the human arm, the forward signal