Yun-hui Liu. Department of Automation and Computer-aided Engineering. The Chinese University of Hong Kong. Shatin, N. T., Hong Kong email: {wtlo, yhliu} ...
Proceedings of the 2002 IEEE international Conference on Robotics 8. Automation Washington, DC May 2002
Improving Efficiency of Internet based Teleoperation using Network QoS* Wai-keung Fung Ning Xi Department of Electrical and Computer Engineering Michigan State University East Lansing, MI 48824, U S A . email: (fungwaik, xin} @egr.msu.edu
Abstract This paper presents a QoS based eficiency improving scheme for Intemet-based teleoperation. One of the widely used QoSparametersfor showing network status . is network delay. With the help of event-based control technique, the stability of the teleoperation systems is . guaranteed under random network delay. This paper ' imiestigates how to improve the eficiency of teleoperated tasks based on the network quality by introducing a Command Negotiator and a robot controller gain adjustment scheme using the measured QoS parameters into the proposed QoS based teleoperation systems. The presented methods improve the efJiciency and system responses of the teleoperation systems when the network quality is pool: Moreover, a teleoperation experiment is presented to demonstrate the effectiveness of the proposed controller gains adjustment scheme.
1 Introduction The concept of Quality of service (QoS) is first introduced in computer network area and is widely discussed in distributed multimedia applications. QoS is defined as a set of quality requirements on the performance of data transmission necessary to achieve the required functionality of an application. The quality of data transmission via the Internet is reflected by a set of QoS parameters. However, the quality of service provided by Internet based teleoperation cannot be guaranteed due to the indeterministicproperties of the network, like network delay. With the help of event-based control technique, the stability of the the teleoperation system is guaranteed regardless the network quality (in term of network time delay). Poor network quality, however, degrades significantlythe efficiency of teleoperated tasks. This paper investigates how the efficiency of teleoperated tasks is improved based on the real-time network quality. Few works have been conducted to investigate the QoS issues in teleoperation control. Nahrstedt and ~~
~~
Wang-tai Lo Yun-hui Liu Department of Automation and Computer-aided Engineering The Chinese University of Hong Kong Shatin, N. T., Hong Kong email: {wtlo, yhliu} @acae.cuhk.edu.& ~
Smith employed the QoS Broker for network resource management (e.g. bandwidth allocation) for a teleoperation system [ 11. Only heuristics were devised in the QoS Broker for network resource management and no mathematical analysis on the performance of the QoS Broker was included. Moreover, the stability issue of the telerobotic system, which is an important issue in control systems design, was also not studied. Wang et, al., moreover, proposed a fuzzy inference approach to generate a QoS index from QoS parameters for adjusting the frame rate of online video feedback in a teleoperation system [2]. The generated QoS index, however, is difficult to be incorporated in robot system models for analysis. Only heuristics can be devised for control strategy adjustment based on this QoS index. Although the stability of the teleoperation system is guaranteed in the presence of network delay, the efficiency of teleoperated tasks is still sensitive to network delay and other network properties, like delay jitter and so on. The effect of poor network quality on the efficiency of teleoperated tasks becomes significant on tasks that require dextrous manipulation of tools and high precision. This paper investigates methods to improve the efficiency of teleoperation systems based the network QoS. In this paper, we propose an adaptive teleoperation system architecture based on the communication quality of the Internet (in terms of QoS parameters) so as to maintain the system dynamic performance and improve system efficiency when the communication performance of the Internet is not good. We propose to improve the efficiency of the teleoperation system in the planning and control aspects. A Command Negotiator with an online operator command learner is introduced in the teleoperation system to improve the system efficiency in the planning process. A quadratic programming based robot controller gains adjustment .scheme for closed loop poles placement in efficient region using the measured QoS parameters is also developed to enhance the system efficiency improvement in terms of response time, overshoot, etc.
2 QoS based Teleoperation
~
*Research partially supported under NSF Grants 11s-9796300, 11s-9796287 and EIA-9911077. under DARPA Contract DABT6399- 1-0014.
0-7803-7272-71021$17.00 0 2002 IEEE
This section describes a QoS based multi-operators multi-robots teleoperation system architecture. Event
2707
Figure 1: QoS based Control in Teleoperation. based control technique is employed in system design so that robot system stability is guaranteed in spite of large network delay in teleoperation [3].
2.1 Event-based Control The concept of event based control, which was first proposed by Xi [4],is to use a non-time based motion reference s for robot system behavior description. Traditional systems employ time as motion reference, which makes the systems sensitive to the effect of random communication delay and thus result in instability and desynchronization among different components in the system even though the original systems without delay are stable. The non-time based motion reference, however, can solve the stability and synchronization problems because it is independent of time delay. The nontime based reference is usually related directly to real time sensor measurements and the task and thus time delay will not have effect on the stability of the teleoperation system. The stability of teleoperation systems with non-time based motion reference is guaranteed if their time-based counterpart are stable and the non-time . motion reference is a non-decreasing function of-time [3]. Unlike the classical approaches, no assumption is imposed on the Internet time delay for system design using the event-based control. Moreover, event based control can prevent “buffering” effect of the communication media (the Internet) in teleoperation so that the synchronization of robot actions and sensor data feedback is achieved [ 5 ] . By using the event-based control technique in teleoperation system, the stability is guaranteed in the presence of network time delay. However, the efficiency of teleoperated tasks is still sensitive to network delay and other network properties, like delay jitter and so on. The effect of poor network quality on the efficiency of teleoperated tasks becomes significant on tasks that require dextrous manipulation of tools and high precision. We propose to improve the efficiency of the teleoperation system with the use of a Command Negotiator and a robot controller gain adjustment scheme based on the measured QoS parameters.
2.2 QoS Parameters The quality of service of a teleoperation system is mainly reflected by a set of network QoS parameters measured that indicate the quality of data transmission via the Internet. The quality of data transmission via the Internet is measured by a set of QoS parameters,
Figure 2: The teleoperation system framework. including network delay, delay jitter, bandwidth allocated, packet loss rate, etc. Network delay is the most widely employed QoS parameters and this paper studies the role played by network delay in command negotiator and controller gains adjustment in the proposed teleoperation system. The measurement of QoS parameters for teleoperation systems may sometimes be difficult because realtime measurement is required for collecting the on-time and accurate picture of the status of the operating environment. For instance, small probe packets are required to be sent periodically from one node to other nodes in the network for the measurement of network delay. However, the accuracy provided by existing clock synchronization algorithms is not acceptable for network delay measurement. Therefore, round-trip delay is usually employed for network delay measurement because clock synchronization among nodes on the network is not required. QoS parameters, including Internet time delay, delay jitter, allocated bandwidth, packet loss rate and so on, measured in each control cycle of the teleoperation s stem are grouped inlo a vector representation, cj = 42,. * ,qm], where m is the number of QoS parameters measured. Since usually subset of QoS parameters are required for the proposed Command Negotiator and controller gains adjustment scheme, the QoS parameters vector is piutitioned into two subsets 41, Q E 4, such that @I U Q = 0,as shown in Figure 1.
it,
-
2.3 Teleoperation System Architecture Figure 2 depicts a multi-operators multi-robots teleoperation framework. In the proposed framework, each operator can be anywhere in the world that has Internet access and has a force feedback joystick setup in hisher side to control robot in remote site. Each operator is responsible for controlling a robot in remote site with the help of online video feedback and force feedback from the robot. Each robot is equipped with a local controller that accept operator’s command as input. Non-time based action reference is computed based on sensor data gathered from the robots and sent to local robot controllers. Moreover, a Command Negotiator (see Section 3) is introduced into the architecture to coordinate conflicting operators’ commands and fuse operator commands with predicted commands based on the QoS parameters measured in order to improve system efficiency when the network quality is poor. In additions, based on the QoS parameters (network delay) 2708
-\\
.
has an online command learner to capture the coordinated operator’s command. The online command learner is realized by the B-spline neural network [6]. Figure 3 depicts the architecture of the B-spline neural network. The widely adopted B-spline neural network is a kind of kemel-based neural network which has spline basis functions Ni(s), as its activation functions. The B-spline neural network is simple and flexible and provides more control for users in the learning process. The high generalization ability of the B-spline neural network allows it to be a good predictor of the operator’s command. The Command Negotiator takes part of the operators’ role in controlling the robots when the network quality is unsatisfactory. The second function of the Command Negotiator is then to fuse predicted commands with operator commands based on the measured QoS parameters so as to improve the efficiency of the teleoperation system (Figure 1). Firstl all QoS parameters are normalized into the interval fb, 11 and transformed to have decreasing influence on the confidence of operator command. For examples, network delay, delay jitter and packet lost rate have decreasing influences on the confidence of remote operators’ commands while bandwidth has increasing influence on the confidence of operators’ command. If only one QoS parameter Q is involved in command fusion, the fused command 6 is determined by
.
Figure 3: The architecture of B-spline Neural Network. measured, a controller gain adjustment scheme is developed (see Section 4) to place closed loop poles of the teleoperation system in a desired region so as to improve the response behaviors and efficiency of the teleoperation system in the presence of indeterministic Intemet time delay. Fast response and small overshoot are maintained in teleoperation system performance under poor network quality, like large Internet time delay, as reflected by QoS parameters.
6 = (1- @“)U0
+ Q’2)N
(1)
where vo and V N denote the velocity commands from operator and the prediction of online command learner in the Command Negotiator and cr > 1 is a smoothing parameter. When m QoS parameters ijj are involved in command fusion, a combined QoS parameter 4 is computed as
3 QoS based Command Negotiator The Command Negotiator is introduced into the proposed teleoperation system architecture. All operator commands are first sent to the Command Negotiator which serves two functions in teleoperation. The first function of the Negotiator is to coordinatejoystick command from each operator and motion of each robot when conflicting commands are given to the robots. Operators sometimes give conflicting joystick commands even though they know the task goals clearly Due to the limited bandwidth available in the Internet, each operator can only receive one or two video feedback with different views of the remote site where the robots are operating. Usually the image quality received is not good. The lack of complete views of the remote site in visual feedback and poor image quality often lead to wrong decision-making of operators, especially when the robots are operated for dextrous manipulation. The Command Negotiator is equipped with an online command learner to capture the intention of each operator based on the past commands given to the robots. When the network status is in poor quality (reflected by QoS parameters measured), the confidence of the Command Negotiator on received operators’ commands drops and the Command Negotiator tries to predict what the actual operator commands are. Each robot in the multi-operator multi-robot teleoperation system
and substitute it for Q in (1) to obtain the fused command. For instance, when only network delay T is employed for QoS parameter in command fusion, the fused command 6 is given by
’.
7
where 8 = - is defined as a phase lag of the sysT tem and T is the sampling period of the system. The Command Negotiator then outputs the fused velocity command 6 as reference velocity command to each local robot controller. The fusion of remote (operator commands) and local (Command Negotiator prediction) controls based on the real-time network quality (reflected by QoS parameters) allows teleoperation systems to drive the robots with confident velocity commands so that high task efficiency is maintained.
4 QoS based Controller Gain Adjustment The QoS parameter employed in robot controller gains adjustment in teleoperation is the time delay involved in
’We exclude the cases that some of the operators intentionally give commands that conflict with the task goals.
2709
follows [9],
information transmission via the Internet. Time delay is introduced into the model of the robot control system. Based on the theory of event-based control, the stability of a system with non-time based motion reference is guaranteed if its time-based counterpart is stable, regardless the indeterministic Internet time delay [3]. The introduction of the controller gains adjustment scheme is to improve the response behaviors and efficiency of the teleoperation system in the presence of Internet delay by placing the closed loop poles in the desired region. Fast response and small overshoot are maintained in teleoperation system performance under poor network quality, like large Internet time delay, as reflected by QoS parameters. In this paper, we take the control of a robot manipulator as an example to study the proposed controller gains adjustment scheme based on measured QoS parameters. The dynamic model of a robot manipulator with six degrees of freedom (DOF)is given by,
oi(~ T + T=) 90i(kT)+ro(Tk)2)(kT)+rl (Tk)W(kT-T) (6)
By introducing the PD feedback control law
wi(t+) = w t ( t ) - F&(t - Tk), i = 1 , . . ., 6 , where Fi = [fi,, f j J T and vi(t+) is piecewise continuous
+ T k ) . the closed
and only changes value at time (kT loop subsystem is given as,
&(kT + T)= A&i(kT)+ BWt(kT) where & is the augmented state vector defined as
where q is the joint angle vector, T is the joint torque vector, Y = [x,y, z, 0 ,A, TI is the robot position and orientation output, D(q) is the inertia matrix, C(q,q) is the centripetal and Coriolis terms, G(q)is the gravity term and q, 7,Y E R6. Nonlinear feedback control [7] technique is employed to linearize and decouple the dynamic model (4) and convert the nonlinear control problem to a linear control problem. After a series of state transformations, the linear control model is in the Brunowsky canonical form and can be decomposed into six linear decoupled subsystems
I
A
1
0
0
In order to study the efficient condition of the robot system for controller gain Fi adjustment scheme derivation, Jury’s test [101 is applied on the characteristic equation of (7). By considering the positivity of the entries in the first column of the odd rows if the Jury array, the desired region for the controller gain Fi of the subsystem i is described as follows,
B
yi=[1
O]Oi
. +
where Oi = [hi,L f h i I T , i = 1 , . . ,6. The model (5) represents the behavior of the system in the k-th time interval t E [kT Tk, ( k l)T Tk+1) in a teleoperation system that has non-deterministic time delay Tk inherited in information transmission via the Internet. Note that each identical subsystem has double poles at the origin and thus is not asymptoticallystable.
+
(7)
+
where
- 2Tk).fi],+ 4Tkf?,
a ( f i 1 , fiz)=-Tz(T
- 3 7 k ) , f i l f i z + 2(T + 2Tk)fil - 4 f i z b(fi,, fi,)=Ti(T2- 2TTk -k 2 T z ) f i + 4 ~ k ( 2 ~ kT ) f i +2Tk(3TTk - 47; - T 2 ) f i lfi, -2T(T 2 7 k ) f i 1 4 T f i 2 - 8 +2Tk(T
+
4.1 QoS based System Dynamic Behavior Analysis
+
It is worth to be noted that the conditions depend on
00s parameters (network delay Tk), sampling period T and controller gains (fi, ,fi,). If only positive gains (fi, ,fi, > 0) are considered, we have ‘2 T k f . < fiz < $ T k f i l + 1 c: a ( f i 1 : L f i z ) b ( f i , , f i , )> 0 (9) 0 < fi, < fi,Tk(T-Tklr)+2
Consider a system that has both the time-based and event-based dynamic models, if the nonlinear feedback control is applied to the two models and their linearized systems are assumed to have the same assigned poles, then the system receives an identical control command regardless which model is used. The designed controllers based on the time-based and event-based systems are, therefore, equivalent. The analysis of the event-based controller can, thus, be performed in the time-based system [8]. Assume that the time delay Tk is less than one sampling period T, Tk < T , V k . By sampling the system, the time-delayed discrete-time model is given as
{
21
T-2Tk
It can be easily showed that a(fil,fi2) = 0 and b(fi,, fi,) = 0 represent hyperbolic boundaries in the fi,-fi, gain plane [ll]. A quadratic programming problem is formulated to adjust the controller gains so 2710
E
gains pair in the fi, -fi, plane and lies withii the desired region defined by conditions C. The objective function of the optimization problem is thus constructed by the Euclidean distance between the nominal gains and any gains pair. The constraint set of the optimization problem, on the other hand, is constructed from the conditions C, which consists of linear constraints, except the second conditions a ( f i , ,f i , ) b ( f j , , fj,) > 0. As discussed in previous section, both a(fil, fi,) = 0 and b ( f i , , f i ) = 0 represent hyperbolic boundaries plane. Each hyperbolic boundary is apon the fj, proximated by three linear boundaries with two of them come from the asymptotes of the hyperbola and the remaining one comes from a line that is perpendicular to the major axis of the hyperbola and passes through the focus, as shown in Figure 4. After the approximation, all the conditions described in C are linear. Therefore, a quadratic program can be constructed to adjust controller gains for the robot system. The quadratic program is expressed as follows,
-A,
-10
-5
0 1.
5
10
Figure 4: Approximation of hyperbolic boundaries.
Robotics and Automation Lab. MSU Michigan, USA
Robot Cootrot Lab., CUHK
Hong Kong
Figure 5: Experimental setup.
as to place the closed loop poles of the teleoperation system within the desired region described by conditions C in the fi, -tiz gain plane.
st.
Fi > 0
where Ac and bc are coefficient matrices of the linear constraints derived from conditions C. Since quadratic programming problem is a well studied optimization problem, many efficient algorithms for solving quadratic programs can be employed for this controller gain adjustment problem [12].
5
4.2 Gain Adjustment Scheme
AcFi 5 bc,
Experimental Results
A teleoperation experiment has been conducted which connects a mobile manipulator (of Robotics and Automation Laboratory, Michigan State University) and a human operator (of Robot Control Laboratory, The Chinese University of Hong Kong) via the Internet (see Figure 5). In the experiment, the Hong Kong operator was asked to control the mobile manipulator via a joystick with force feedback to swap positions of two metal pieces on a small platform, with the help of video feedback from the remote robot side for operator guidance. The experimental setup of the series of pick-and-place tasks (swapping) is depicted in Figure 6. The metal pieces are named respectively A, B, C and D and the Hong Kong operator was asked to swap the positions of metal pieces B and D. This task was performed with and without the proposed QoS based controller gain adjustment scheme applied on robot controllers. Interested readers may refer to [5] for the implementation details and setup of the system employed in the experiment. The QoS parameter considered in this experiment is network delay, which is assumed to be half of the round-trip time (RTT) experienced between US and Hong Kong during the experiment. Average and standard deviation of RTT are 277.8ms and 37.6ms respectively when the experiment is conducted. The main difficulty of the teleoperated pick-andplace tasks is the correct positioning of the robot gripper for grabbing the metal pieces provided with single view of video feedback and indeterministic network quality.
The basic concept of controller gain adjustment scheme is described as follows. - Assume that a pair of nominal gains = (fi, ,fi,) is assigned to the robot subsystem i. The nominal gains are the controller gains designed for the system under scenarios that do not have time delays. The nominal gains are, at the same time, designed to satisfy certain dynamic performance requirements of the system, such as response time and overshoot. The current employed controller gains (fi, ,fi,) are examined to see whether the robot subsystem lies outside the desired region defined by conditions C. If the system lies within this region with the current controller gains, the current gains are kept using in the robot system, else the nominal gains are examined to see whether the system lies within this desired region with the nominal gains. If the system lies within the desired region with the nominal gains based on the measured QoS parameter (network delay ~ k ) ,the controller gains are then set to the nominal gains, else a controller gains , is selected for the system so that it pair Fi = is within the desired region and the distance between it and the nominal gains is minimized. A quadratic programming (QP) problem is formulated to approximate the controller gain Fi selection process. The quadratic program helps to determine a controller gains pair 4, that is nearest to the nominal
( A 6,)
2711
Figure 6: setup.
tem efficiency is not ensured in the presence of network time delay. In this paper, we proposed to improve the efficiency of the teleoperation system in the planning and control aspects. In the planning process, a Command Negotiator, that has an online operator command learner installed, is introduced in the teleoperation system to improve system efficiency. Moreover, a quadratic programming based robot controller gains adjustment scheme using the measured QoS parameters is developed to intensify the improvement of system efficiency. In additions, a teleoperation experiment is presented to demonstrate the effectiveness of the proposed controller gains adjustment scheme.
Experimental Figure 7: RTT trace during experiment.
References K. Nahrstedt and J. M. Snlith, “QOS Negotiation in a Robotics Environment,’’ in Workshop on Distributed Multimedia Applications and Quality of Service Verifcation, Canadian Institute for Telecommunication Research, Montreal, 1994, pp. 153-61. Q. P. Wang, D. L. Tan, N. Xi, and Y. C. Wang, “The Control-Oriented QoS: Analysis and Prediction,” in Proceedings of the 2001 IEEE International Conference on Robotics and Automation, ICRA’2001, 2001, vol. 2, pp. 1897-1902. N. Xi and T. J. Tam, “Stability Analysis of Nontime Referenced Intemet-based Telerobotic Systems,” Robotics and Autonomous System, vol. 32, pp. 173-8, 2000. N. Xi, Event-Based Planning and Ccontrolfor Robotic Systems, Doctoral Dissertation, Washington University, Dec. 1993. I. Elhajj, N. Xi, W. K. Fung, Y.H. Liu, W. J. Li, T. Kaga, and T. Fukuda, “Haptic Information in Internet-Based Teleoperation,” IEEWASME Transactions on Mechatmnics, vol. 6, no. 3, pp. 5-14, Sept. 2001. M. Brown and C. Harris, Neurofiuy Adaptive Modelling and Control, Prentice Hall, 1994. T. J. Tam, A. K. Bejczy, A. K. Isidori, and Y.L. Chen, “Nonlinear Feedback in Robot Arm Control,” in Proceedings of the 23rd IEEE Conference on Decision and Control, CDC’84, 1984. T. J. Tam, N. Xi, and A. K. Bejczy, “Path-based Approach to Integrated Planning and Control for Robotic Systems,” Automatica, vol. 32, no. 12, pp. 1675-87, 1996. K. J. Astrom and B. Wittenmark, Computer-Controlled Systems Theory and Design, Prentice Hall, 3rd edition, 1997. G. E Franklin, J. D.Powell, and M. L. Workman, Digital Control of Dynamic Systems, Addison-Wesley Publishing Company, 2nd edition, 1990. D. Zwillinger, Ed., CRC Standard Mathematical Tables and Formulas, CRC Press, 3rd edition, 1996. P. E. Gill, W. Murray, and M. H. Wright, Practical Optimzation, Academic Press Inc., 1981.
Figure 8: QoS based pro- Figure 9: QoS based difportional gain K p adapta- ferential gain K,,adaptation. tion. During the experiment, the Hong Kong operator can accomplish the requested swapping task successfully and swiftly with QoS based controller gain adjustment, in spite of large network delay. On the other hand, the Hong Kong operator, in general, took longer time to position the gripper for grabbing one of the metal pieces and achieve the swapping task without the proposed QoS based controller gain adjustment. Moreover, the remote operator controlled the gripper to traverse unnecessary paths in order to position it to the desired positions, which con&buted to the long task completion time. The average task completion times for the cases with and without the QoS based controller gain adjustment are 582.3s and 887.4s respectively, which demonstrate that the teleoperation efficiency is improved with the proposed gain adjustment scheme. In additions, the proposed controller gain adjustment scheme guarantee the robot system to have closed loop poles placed in an efficient region during the experiment. The remote operator did not lose control on the robot arm in achieving requested tasks. The mobile manipulator also exhibits motions with fast response to operator commands and small overshoot because the adjusted controller gains are selected to be as similar to the nominal gain as possible, which contributes to the good performance of the system without network delay.
6 Concluding Remarks In conclusions, this paper considers network delay as a representative QoS parameter for the study of the proposed QoS based teleoperation architecture. With the help of event-based control technique, the stability of the teleoperation system is guaranteed in spite of the indeterministic Internet time delay. However, sys2712
