A Psychophysically Motivated Compression Approach for 3D Haptic Data

Transcription

1 A Psychophysically Motivated Compression Approach for 3D Haptic Data Peter Hinterseer Eckehard Steinbach Institute of Communication Networks Fachgebiet Medientechnik Technische Universität München Munich, Germany ABSTRACT Haptic interaction with virtual as well as real environments is much more demanding to the underlying network than traditional audio/video communication. Because of very strict delay constraints imposed by the global control loop which is closed by the communication system very high packet rates are generated. Each packet typically carries a small amount of payload data leading to a significant packet header versus payload overhead. This paper presents a psychophysically motivated deadband transmission approach for multidimensional haptic data using the example of three dimensional haptic interaction. Packets are only generated and sent if the change in haptic variables exceeds the just noticeable difference of the human operator. The presented approach reduces packet rates in haptic systems by up to 9% without any perceivable reduction of immersiveness. To our knowledge the proposed approach is the first psychophysically motivated compression approach for multidimensional haptic data in literature. Keywords: haptics, compression, psychophysics, deadband, telepresence, teleaction 1 INTRODUCTION With the ongoing growth of popular Internet based audio/video (AV) applications like real time AV-streaming and online 3D gaming it is only a matter of time when new senses of the human being will be addressed in professional and entertainment multimedia applications. Using the possibilities of modern computer technology makes computer generated worlds more realistic than ever. To gain even more realism the sense of haptic perception is the next logical step towards a user which is fully immersed in a virtual (Virtual Reality, VR) or real remote (Telepresence and Teleaction, TPTA) environment. To achieve such an immersion it is imperative that the multimodal stimuli produced by the operator display are as realistic as possible. Unfortunately, the more realistic the presentation for the user, the higher the bitrate that is required to transport this information. This is one main reason why compression techniques in audio and video applications have gained such a high interest among researchers as well as in industry. Therefore a significant number of very efficient and high quality compression schemes exist for these two types of multimedia data. The transmission of the haptic modality which mainly consists of data like positions, angles, velocities, forces and torques has not found a very broad interest in the multimedia communication community yet. This is particularly surprising when considering the ph@tum.de eckehard.steinbach@tum.de high requirements of this modality with respect to delay time, delay jitter and packet rate. This high demand comes from the fact that a global control loop is closed over the communication system. The Operator (OP) commands certain positions/velocities using his Human-System-Interface (HSI) which are transferred over the network to the Teleoperator (TOP). The TOP tries to reach/match the commanded positions/velocities using a local control loop. If the TOP has contact with the environment during this action it sends back the occuring forces which are either measured or calculated. Numerous research activities report that already very small time delays [16, 15] in the communication (<1ms [17, 13]) result in instability of the system. This leads to user dissatisfaction in virtual environments and can even be dangerous in case of TPTA when robots and HSIs get out of control. This is the reason why control loops on both sides of the system are necessary. In order to make transmission of this very time critical data less demanding we propose a deadband transmission approach which uses psychophysically motivated threshold values to reduce the packet rate on the network. Packet rate reductions of up to 9% are reported in [9] for a 1-Degree-of-Freedom (DoF) TPTA system. In this paper we present a more comprehensive study using 3-DoF haptic interaction with a virtual 3D environment. The remainder of this paper is organized as follows. In Section 2 we first give some theoretical background of the psychophysical point of view of haptic TPTA. After that a detailed description of the proposed multidimensional deadband transmission approach is given (Section 3). Our setup for experimental verification is described in Section 4. After the presentation and discussion of the results in Section 5 a conclusion (Section 6) is given along with some acknowledgements in Section 7. 2 PSYCHOPHYSICAL BACKGROUND 2.1 Weber s Law Human perception has undergone thorough research during the last century. The respective perceptual threshold values for all kinds of stimuli put on the human body have been studied. Apart from very detailed information for every modality a human being can perceive, one major conclusion emerged from these studies: Human haptic perception often follows Weber s Law. Ernst Weber was an experimental psychologist who in 1834 first discovered the following implication I I = k or I = ki where I is the so called Difference Threshold or the Just Noticeable Difference (JND). It descibes the smallest amount of change of an (arbitrary) stimulus which can be detected just as often as it cannot be detected. I is the initial stimulus which is altered by the JND and the constant k describes the linear relationship between the JND and the initial stimulus.

2 2.2 Application to Haptic Perception Weber s Law can be generally applied to human perception. It is not always true but still gives a good baseline for stimulus relationships. Further studies [12] tell us that the human haptic perception system follows Weber s Law quite well. The factor k which signifies the magnitude of the change in a stimulus which can be perceived lies mostly between 5% and 15%. This depends both on the type of the stimulus and the limb or joint where it is applied [2]. 3 DEADBAND TRANSMISSION The main idea of the deadband transmission approach as initially proposed for 1-DoF haptic data in our previous work [9] is presented in the following. 3.1 Transmission of Haptic Data Data in a haptic interaction system is sampled at a sampling frequency of between 5Hz and 1Hz which has been found to be necessary to decently reproduce all kinds of haptic sensations a human being is able to sense. Because of the tight delay constraints in such a system, every single sample value has to reach the receiver side as quickly as possible. Every effort to collect sample values to improve header-to-payload ratio will result in additional delay in the system. Furthermore more than one sample value per packet is not necessary at all, because sample values serve as set values for the control loop on the receiver side. So old set values are simply useless because they represent a state of the system which is no longer the most accurate. 3.2 Deadband Principle The basic idea which stands behind the deadband transmission approach is that new sample values taken from sensors only have to be transmitted to the receiver side when the user is able to sense the change of the sampled variable in comparison to the previously transmitted sample value. If for example the user is presented with a force of 1N (this was also the value which was last received) and the deadband is given with p = 1% the next force sample value is only transmitted once it goes either below.9n or above 1.1N. Every force change in the interval from.9n to 1.1N is considered imperceptible by the human operator and therefore not necessary to be transmitted. The parameter k in Weber s Law (Section 2.1) is basically the upper bound of the deadband p. Once p is larger than k the subsequent changes in the signal become perceivable for the user and interaction starts to feel distorted and the quality of immersion is reduced. 3.3 Previous Work In [9] we have recently presented a preliminary study of haptic deadband transmission for a 1-DoF velocity-force architecture. In this architecture the velocity of the master device is measured and sent to the receiver side. There the velocity is matched using a velocity control algorithm with position error correction. If the slave device has contact with the environment (in the presented case a stiff wall) forces are created, measured, and sent back to the master device where they are displayed using a force control algorithm. The experimental setup is shown in Figure 1. In this very simple experiment the basic applicability of the deadband approach was investigated. The results showed successfully that the characteristics of the human haptic system can be exploited for data reduction. An example of the resulting signals with 25% deadband is shown in Figure 2. Note the increasing step size of the signal for increasing sample magnitude. The usefulness of a 1-DoF OP side PC force control Sensoray S626 IO DAC ADC counter channel with deadband control RT Linux velocity control Sensoray S626 IO DAC ADC counter environment strain gauge motor TOP side Figure 1: Experimental setup from our previous work. encoder device is very limited. Common haptic display devices have at least 3-DoFs to allow interaction in 3D space. Velocity [rad/s] Force [N] original signals signals after deadband Time [s] Figure 2: Velocity and force signals before and after applying the deadband transmission algorithm with p =.25 on the 1-DoF device in Figure Limitations of the 1-DoF Deadband Approach Using a 3D haptic device for interaction in either VR- or TPTAsystems is usually done by sending the three cartesian components (or another representation of 3D space) of the current position or velocity of the device to the TOP side. The TOP side then either calculates (VR) or measures (TPTA) the resulting forces and sends its three cartesian components back to the OP side. The simplest way of deadband transmission would now be to apply the 1-DoF deadband approach to every single component of the cartesian representation. While this is a straight forward extension it does not perform well in practice. The deadbands of the cartesian components are mostly uncorrelated. It is always possible that, for example, velocity in x-direction is very large, so the deadband on this axis is also quite large whereas velocity in y-direction is very small and therefore reacts very sensitively to change because of the small deadband. Now we have to transmit a packet every time when only one of the components exceeds its respective deadband. Although only the value for the surpassing component would have to be sent, a packet has to be generated nevertheless and so the goal of reducing packet rates is not very well served. If random movements with equally distributed directions and magnitudes of forces and velocities are examined, always the component with the lowest magnitude and therefore the smallest deadband is mostly responsible for packet generation. The probability of having a component with low magnitude therefore increases with the number of components used. It becomes obvious that the more DoFs a system has, the worse the 1-DoF deadband approach will work.

3 3.5 Extension to Multiple DoFs To overcome the aforementioned limitation we propose a multidimensional deadband approach in this paper. In the following we explain the extension of the one dimensional deadband (basically an numeric interval) to two dimensions (where the deadband becomes a circular area) respectively three dimensions, where a spherical volume element serves as deadzone (we will denote a multidimensional deadband as deadzone from now on). To be able to use a deadzone the respective sampled unit has to be represented as a vector. In the following we will explain the proposed approach in two dimensions for simplicity reasons. The extension to three dimensions is then similar. In the 1-DoF case the magnitude of the difference d between an initial value v i and a current value v c has to be evaluated. This is done by calculating the numeric difference between those two sample values and comparing it to a threshold value p (the deadband multiplied by the initial value v i ). d = v i v c d p v i = Do nothing d > p v i = Transmit new value In the two dimensional case the scalar variables v i and v c become the vectors v i and v c and the operators become their respective vectorial counterparts. d, the magnitude of the vectorial difference, and p, the deadband are still scalar. d = v i v c d p v i = Do nothing d > p v i = Transmit new value To better understand what that means geometrically, take a look at Figure 3. Figure 4: Criterion for transmission of new values in the 2-DoF case. The deadzone having the shape of a circle makes is computationally easy to calculate whether the deadband was violated or not. Basically the length of the vector v i v c has to be compared with p v i, the length of the initial vector multiplied by the deadband factor p and depending on the result the respective action is taken just like shown in the equations above. The size of the deadzone circle depends only on the length of vector v i whereas the maximum of the angle a depends only on the deadband factor p. It can be calculated as follows: a reaches its maximum, when and v c v i v c v i v c = p v i i.e. v c is tangential to the deadzone circle. So a max can be calculated as follows: In this case v camax would be sina max = v i v c = v i = p v i = v i = p a max = arcsin p Figure 3: Geometrical description of a 2-DoF deadzone. So v i v c is the difference vector of the initial and the current vector. The deadzone is depicted as a circle around the tip of vector v i with radius p v i. The angle between v i and v c is a. The decision whether to transmit a new value is made as shown in Figure 4. If the tip of vector v c lies within the deadzone circle, the deadband is not violated and thus no new value is transmitted. If the tip lies outside the deadzone circle, updated sample values are sent. cosa max = v c amax v i v camax = cosa max v i = cos(arcsin p) v i This means that no matter how large a sampled 2-DoF variable (velocity, force,... ) is, once it changes its direction by a max an updated value will be sent to the receiver. So the multidimensional deadband algorithm provides a constant threshold value and no magnitude dependant deadband as far as direction is concerned. The extension of this approach to 3D is straight forward. The vectors v i and v c become 3-dimensional, the circular deadzone becomes a spherical deadzone, in which the tip of v c has to lie to not

4 trigger an update value. The values of amax and vca max stay the same, because the vectors vi and vc in any case define a plane in which the above calculations are true. 4 E XPERIMENTAL V ERIFICATION In order to verify the presented 3-DoF deadzone approach, several experiments with a comimercially available 3-DoF haptic device are conducted. Figure 6: The graphical display of the OP side (HSI). The pose of the haptic display device is sampled with 1Hz. The graphical display is refreshed with the standard refresh rate of 6Hz Figure 5: The SensAble Phantom Omni device used for the experiments. 4.1 Experimental Setup The conducted experiment consists of a haptic interaction task with a remote virtual environment. The hardware and software setup is the following: On the OP side the haptic display device SensAble PHANTOM Omni (see Figure 5) serves as the HSI. Over a 1Mbit/s Ethernet LAN connection this OP side transmits current position and velocity samples to a simulated haptic environment on another machine in the same LAN OP Side Haptic Display Device The haptic device is capable of 6-DoF input and 3-DoF output. This means both the endeffector s position in space as well as its orientation can be read from the device drivers. In contrast to that it is only possible to output forces in 3-DoFs namely the three directions in space. The torques necessary for altering the endeffector s orientation cannot be produced. In our experiment, the additional 3-DoFs of endeffector orientation are only used to display the 3D-cursor of the graphical display correctly. They are neither sent to the TOP side nor do they have any other influence Graphical Display The graphical display consists of a simple OpenGL based 3D visualization of the workspace. Both the current cursor position and the position of the haptically manipulated object is displayed. See Figure 6 for an impression of the HSI graphical display. What one sees in this display is a big grey sphere in the middle of the workspace of the haptic display device along with the blue cursor which signifies the current position and orientation of the device. Deadzone Implementation The three components of the current device velocity are combined in a 3D vector. According to the formulas in Section 3 the size of the deadzone is calculated. The reference (or initial) vector is always the one which was last transmitted to the receiver. Now at every 1ms a new value of the 3D velocity vector is read from the device drivers and it is decided whether its tip lies in the deadzone or not. According to the result of the decision, a new vector is sent to the receiver or the new vector is discarded Position Update The sending of only velocity values in such a deadband system would result in a more or less severe degradation of position tracking. It is therefore necessary to send the actual position values along with the current velocity values so that the TOP can take care of eventual position errors. Since the packet sizes are very small anyway this additional amount of data is negligible in this case TOP Side Virtual Environment The virtual environment is implemented by a C++-Class which manages the positions and properties of virtual objects which are to be manipulated as well as the positions of one or more users interacting with the environment. It is capable of 3-DoF input and 3DoF output. This means it is fed with a 3-DoF velocity input (along with a 3-DoF position input for reasons of position tracking) and calculates the resulting forces for this position. The environment as well as the haptic display device at the OP side are refreshed at a rate of 1Hz. Unlike the system in [7] which also implements a virtual haptic environment it is intentional that all computations concerning the haptic feedback are done centralized on one machine like in [4] and [11]. This central approach is chosen in order to have a system which is as comparable as possible to a real TPTA system. Also other than in [6] where only one packet is in transit at all times for stability reasons, the presented system communicates in both directions at the same time. Stability problems were not observed.

5 4.3.2 Sphere Object The only object in the virtual environment in this experiment is a sphere in the middle of the virtual workspace. This sphere is of course registered with the sphere in the graphical display (see Section 4.2.2) so that contacts between the cursor and the sphere in the graphical display exactly correspond with contacts in the virtual environment. The virtual sphere is fixed at the center of the workspace and can be touched with the virtual cursor. The resulting force during the interaction is calculated by a simple Hooke s Law F = u d where u is the stiffness of the sphere and d the amount of penetration into the sphere body. F is the resulting force s magnitude. The direction of the force always points from the sphere s center to the actual cursor position. So in this case it can be calculated as F = x s (r x s ) u x s where the resulting force vector F is calculated from the current position of the user x, the sphere s position in space s, the sphere s radius r, and the sphere s stiffness u Deadzone Implementation The deadzone implementation is also straight forward in this case. The initial vector for the deadzone calculation is the force vector which was last sent to the OP side. Every time the virtual haptic model is updated it either sets the most current position and velocity values for the users position (in case an update packet has arrived since the last update) or calculates a new position from the last known user position, the last known user velocity, and the exact time since the last update. This updated position is then used to calculate an updated force which then is used as the current vector for the deadzone calculations. In case the deadzone is violated by the new vector, a new packet containing the updated force vector is sent and the sent vector is saved as the new initial vector. 4.4 Subjective Evaluation Due to the problem of measuring the results of perception based compression techniques objectively it is imperative to analyze the methods in a subjective evaluation. Standard objective measures like MSE or PSNR cannot be applied here. For the subjective evaluation presented in this paper a number of test subjects underwent the procedure described in the following to determine suitable values for the deadband parameters so that no decrease of immersiveness can be noticed Subjects All subjects were males in the age of 27 to 31, their sensory motor capabilities were normal Evaluation Procedure Three types of experiments were conducted with the goal to estimate a perception threshold value for the deadzone for the following three cases: 1. Deadzone is only applied on velocity values 2. Deadzone is only applied on force values 3. Deadzone is applied on both To achieve that, the subjects are first presented with a system completely without deadband to get used to handling the device and to learn to know what it feels like. Then a heavily distorted system is shown to the subjects in which they can clearly feel the kind of distortion which is introduced into the system by the deadzone transmission. This phase is called the familiarization phase. After the subjects feel familiar with the system and the kind of distortion they are presented with, three test runs are conducted each consisting of twelve 3-second intervals (36 intervals in 18 minutes total) in which the subjects were told to haptically explore the virtual environment and assess the quality of the haptic presentation. In the first run with twelve intervals the deadband is only used for the velocity values which are sent from the OP to the TOP. In the second run the deadband is only used on force values which are sent from TOP to OP. And finally in the third run the deadband is used in both directions with the same deadband value for force and velocity. During the tests, the subjects wore a headphone which played loud music so they had to concentrate on their haptic sensations. The 12 intervals of each run use a randomly chosen order of the following possible deadband values: %, 2.5%, 5%, 7.5%, 1%, 12.5%, 15%, 2%, 25%, 3%, 35%, and 4%. The subjects did neither know which value was currently used nor did they know in which direction(s) the deadband was applied. After every interval the subject was asked to give a rating for the just perceived interval. If it felt perfectly like the undistorted signal from the familiarization phase, they should give a rating of 1 points. If it feels just as bad as the heavily distorted signal from the familiarization phase, they should give a rating of 1 point. The ratings in between can be chosen according to the quality of the signal where of course higher ratings signify a better quality. 5 RESULTS 5.1 Ratings The average ratings for the three experimental runs can be seen in Figure 7. Average Rating Velocity only Force only Velocity and Froce Deadband value [%] Figure 7: The average rating of velocity, force, and combined force/velocity deadband values. The most interesting result is that a deadband usage on velocity values seems to be far less perceptible as on force values. One can see that the velocity deadband can be increased until up to 2% while still reaching an average rating of almost 7 points, which most subjects described as barely perceivable distortion (see Figure 7). In comparison, the force deadband should not be far above 5% for the average rating to stay also above 7 points.

6 This behavior has two different reasons. The first reason lies in the fact that the used TOP is only a VR environment and not a real one as in TPTA. The inevitable position error in such architectures can be very easily corrected by just setting the actually transmitted position value as the current position. In TPTA it is much more complicated to correct this position error because the endeffector has to be moved towards the correct position. This introduces additional distortion into the system whereas in (admittedly quite simple) VR environments the position can instantaneously be updated. The second reason can be explained by the deadband principle itself. As we mentioned in Section 3.2, a new value should only be transmitted if the user can just not sense the introduced change. This is of course true for the direction from the TOP to the OP. The transmitted forces are directly sensed by the human being. In contrast to this in the other direction no human sensory system is involved. So there is basically no reason to transmit new velocity values strictly according to human sensor capabilities. This velocity is only processed by a non-human system which uses it merely to generate new force values for the HSI. The bottom line is that it is possible (at least in the presented case, which stands somehow for most VR haptic environments) to use deadband transmission far beyond the human haptic sensory capabilities for the direction from OP to TOP. Another factor becomes obvious when analyzing Figure 7. The rating profile for deadband usage in both directions is almost the same as for the case that only the force deadband is applied. This once again shows, that the influence of the force deadband on system quality is much higher than of the velocity deadband. 5.2 Packet Rates The measured average packet rates as a function of the deadzone value can be seen in Figure 8. Packet rate [pkts/s] Velocity Force Deadband value [%] Figure 8: The average measured packet rates of velocity and force packets. The first observation is that in general more velocity packets than force packets are generated. This is easy to explain. Velocity packets have to be sent all the time for tracking the endeffector whereas force packets have only to be sent when contact with the environment takes place. Secondly force usually reaches higher magnitudes (and therefore higher deadbands) more quickly than velocities. In the case of this experiment the test subjects are in contact with the environment almost all the time, and so velocity is mostly small whereas force is quite high in most cases. The observation that with % deadband less than 1 packets per second are sent comes from the fact that even with % deadband a change in the measured variable must occur to trigger a new packet transmission. In the case of calculated forces of the VR environment, force is exactly zero while no contact to the environment is made. Therefore only in case of contact, packets are sent. Knowing that, we can conclude that during the experiments with % deadband the subjects had contact with the environment about 77% of the time. Because velocity values are smoothed by the device drivers of the PHANTOM Omni to reduce high frequency jitter, also velocity packets do not reach 1 packets per second at % deadband. Where normally the measument noise of the tracking sensors would trigger the deadband at very low velocities this does not occur with the mentioned smoothing applied. In comparison with our previous work in [9] we can conclude that the deadband usage in three dimensions leads to similar packet rate reductions as the 1-DoF approach. Packet rates for velocity packets are reduced by almost 75% when using barely perceivable 2% deadband. For force packets a reduction by almost 9% is possible by chosing the also barely perceivable 5% deadband. 6 CONCLUSIONS Concluding this paper we can state that in general the 3-DoF deadband approach leads to significant reductions in packet rates without impairing the immersiveness of the system. We have shown that the multidimensional deadzone approach works almost as well as the application of a 1-DoF deadband presented in previous work. For the usage in TPTA systems it has still to be proven that this approach, just like the 1-DoF approach, is passive and therefore stable. Research in this direction is soon to be made but the fact that no kind of instability occurred during the conducted experiments even with deadband values up to 5% is very promising. 7 ACKNOWLEDGMENTS We would like to thank the collaborative research center SFB 453 of the German Research Foundation (DFG) to make this work possible. REFERENCES [1] R. Anderson and M. Spong. Bilateral control of teleoperators with time delay. In IEEE Transactions on Automatic Control, volume 34, pages , [2] Grigore C. Burdea. Force and Touch Feedback for Virtual Reality. Wiley, [3] M. Buss and G. Schmidt. Control problems in multi-modal telepresence systems. In P. M. Frank, editor, Advances in Control: Highlights of the 5th European Control Conference ECC 99 in Karlsruhe, Germany, pages Springer, [4] H. R. Choi, B. H. Choi, and S. M. Ryew. Haptic display in the virtual collaborative workspace shared by multiple users. In Proceedings of the 6th IEEE International Workshop on Robot and Human Communication, pages , Sendai, Japan, September [5] N. Chopra, M. W. Spong, S. Hirche, and M. Buss. Bilateral teleoperation over internet: the time varying delay problem. In Proceedings of the American Control Conference, pages , Denver, Colorado, 23. [6] I. Elhajj, N. Xi, W. K. Fung, Y. H. Liu, W. J. Li, T. Kaga, and T. Fukuda. Haptic information in internet-based teleoperation. IEEE/ASME Transactions on Mechatronics, 6(3):295 34, September 21. [7] J. P. Hespanha, M. McLaughlin, G. S. Sukhatme, M. Akbarian, R. Garg, and W. Zhu. Haptic collaboration over the internet. In Touch in Virtual Environments: Haptics and the Design of Interactive Systems. Prentice Hall, 22.

7 [8] K. Hikichi, H. Morino, I. Fukuda, S. Matsumoto, Y. Yasuda, I. Arimoto, M. Iijima, and K. Sezaki. Architecture of haptics communication system for adaptation to network environments. pages , Tokyo, Japan, August 21. [9] P. Hinterseer, E. Steinbach, S. Hirche, and M. Buss. A novel, psychophysically motivated transmission approach for haptic data streams in telepresence and teleaction systems. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing, pages , Philadelphia, PA, USA, March 25. [1] S. Hirche and M. Buss. Packet loss effects in passive telepresence systems. In Proceedings of the IEEE Conference on Decision and Control, Atlantis, Bahamas, December 24. to appear. [11] Y. Ishibashi, T. Hasegawa, and S. Tasaka. Group synchronization control for haptic media in networked virtual environments. In Proceedings of the 12th International Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems, pages , Chicago, Illinois, USA, March 24. [12] L. A. Jones and I. W. Hunter. Human operator perception of mechanical variables and their effects on tracking performance. Advances in Robotics, 42:49 53, [13] A. Kron, G. Schmidt, B. Petzold, M. F. Zäh, P. Hinterseer, and E. Steinbach. Disposal of explosive ordnances by use of a bimanual haptic telepresence system. In Proceedings of the IEEE International Conference on Robotics and Automation, pages , New Orleans, USA, April 24. [14] P. G. Otanez, J. R. Moyne, and D. M. Tilbury. Using deadbands to reduce communication in networked control systems. In Proceedings of the American Control Conference, Anchorage, Alaska, May 22. [15] E. Ou and C. Basdogan. Network considerations for a dynamic shared haptic environment. In Proceedings of The National Conference on Undergraduate Research (NCUR), Whitewater, Wisconsin, USA, April 22. [16] K. S. Park and V. Kenyon. Effects of network characteristics on human performance in a collaborative virtual environment. In Proceedings of the IEEE Virtual Reality Conference, pages , Houston, TX, USA, August [17] R. T. Souayed, D. Gaiti, G. Pujolle, W. Yu, Q. Gu, and A. Marshall. Haptic virtual environment performance over ip networks: A case study. In IEEE Symposium on Distributed Simulation and Real-Time Applications, pages , Delft, Netherlands, October 23. [18] Kay M. Stanney. Handbook of Virtual Environments. Lawrence Erlbaum Associates, 22.