Web Telerobotic Conclusion

In this thesis it has been argued that the principle obstacle to the implementation of many plans for autonomous robots has been the complexity and subsequent cost, required for such autonomy. I have therefore suggested a somewhat less ambitious solution which involves reducing the sophistication of tasks that the robot is asked to perform and using human intelligence to supervise task performance. The use of human intelligence not only reduces application development cost but also obviates the need for developing artificial intelligence which, to date, has not provided an adequate solution for the building of clever robots.

Teleoperation provides a mechanism for a human to supervise and interact with a robot. Previously, and prior to the World Wide Web, telerobotics was restricted to hazardous environments where cost was a secondary consideration to human safety. However, using the Web as a platform for teleoperation removes the cost associated with operator workstations and makes teleoperable devices available to a mass market. Application areas include the remote operation of equipment such as underground mining vehicles, the sharing of expensive equipment such as machine tools, education and entertainment.

This thesis has reported on a project in which several telerobots were built to allow operators to manipulate wooden blocks and build structures. The implementation of this facility has been through a series of developments which have included three different robots and three different computer operating systems. The development process and the telerobots built are described in this thesis. Because developments were immediately available on-line and therefore accessible to the large numbers of operators that the Internet attracts, a substantial population was available to test reactions to modifications in design as they occurred.

The goal of successful interface design for a web telerobot is to provide maximum understanding with a minimum of transmitted data. Observation of actions by the large population of operators was used to guide interface development and modification. It was found that the multiple camera views with diffuse lighting provided the best visual feedback, while a wire frame model showing six views was not popular with operators. Very few operators were willing to make requests by clicking on a location in a wire frame view. Nor were many operators interested in specifying locations in three dimensional space by clicking in two camera views taken from different angles. However, the introduction of the concept of spin and tilt, to replace roll pitch and yaw, lowered the number of degrees of freedom from six to five and operators appeared to find this easier to understand. The control of the gripper orientation has always been carried out in absolute coordinates, but, where operators had a choice between specifying absolute positions or positions relative to the current gripper position, there was a strong preference for absolute positions. For a period, an operator was required to reset the robot after a collision, but operators found this too hard to understand so the reset function was subsequently automated. A major limitation of the human-computer interface design was the substantial bandwidth required to transmit camera images. Allowing operators to change image quality, image size or switch cameras off provided a method of determining operator preference for imaging alternatives, however, most (92% in one sample) were willing to accept the default settings.

The interface was constructed from a template with tags for variables that are replaced by the software. To seek further interface ideas a facility was provided to allow operators up load and use their own template designs. Few operators were willing to do this, even with the incentive of a competition and small prize, although some interesting ideas were generated. One technique that can be adopted on encountering a new interface is "task-free exploration" where an operator attempts to explore the capabilities of the available controls. The alternative is "task-oriented exploration" where an operator directs their efforts towards accomplishing a task. Operators using the telerobots were found to engage predominantly in ‘task-oriented exploration’. This result was similar to a study by Rieman (Rieman 1994:43) who found task oriented exploration to be the usual method of exploring a new interface.

Most operators, of which there were more than 5000 per month during 1998, were content just to see the telerobot move or perhaps pick up a block, but a small number spent many hours building complex structures that demonstrated considerable ingenuity. Operator reaction was found to be overwhelmingly favourable to the telerobots developed in the course of the project. Of the 1,157 comments recorded on the comments page by August 1998, 77% were congratulatory and 24% offered suggestions on how the system could be improved. Since telerobotic devices available to large numbers of people did not exist before this research, there was no information on how people would interact with a web telerobot and what could be learned from these interactions. Therefore, operator actions were recorded by a variety of techniques during the development. The most obvious records to use are those available from the web server logs and some information, for example, on how people find the telerobot is only available from this source. However, these records were found to be inadequate for providing other, more detailed, information so the software was altered to record the details of each request from an operator as well as the state of the robot after executing each request. To gain additional information, an operator registration facility was added which required the registrant to answer a series of questions. This provided demographic information and allowed operators to be tracked by ‘log-on’ name. Tracking operators with the 'log-on' facility enabled repeat visits to be measured and showed that tracking by internet address was inadequate for measuring repeat visits.

Statistics show that males represent 68.7 percent of internet users (Read et al. 1998) but 95% of operators registered to use the telerobot were male. This demonstrates a greater interest in this technology by males. It was also demonstrated that the population of registered operators was more skewed towards youth than the general internet population. An issue that has importance for potential commercial services is how users find a site. Most traffic originates from referrals. Nielsen (Nielsen 1997b) has suggested that distribution of referrals can be described by Zipf's law. However, fitting the Zipf distribution to the data was complicated by having to define what represents a distinct referrer. It was found that Zipf's law did not apply to the Pittsburgh telerobot, where most operators were referred from the Perth telerobot site and that Zipf's law approximated referrals to the Perth telerobot but did not completely explain the distribution of referrals.

After visiting the telerobot once, an operator more often than not will return for another session. From counting the number of sessions by registered operators, it was observed 29% of operators will operate the telerobot for three or more sessions. It was also found that those interested in operating a web telerobot would rarely wait more than three minutes to operate it if it were unavailable.

Some consideration has been given to funding sources for web telerobotics. Web robotics competes for funding in traditional application areas with other telerobotic systems but has some additional advantages based on availability to a mass market. When revenue is proportional to usage, attracting traffic becomes important. Since traffic on web sites has been found to follow a Zipf distribution, the most popular web site will have twice the traffic of the second most popular and therefore twice the revenue. Thus, the most popular service will generate disproportionately higher revenue than its less popular competitors. Banner advertising is frequently suggested as a method of funding web telerobotics, particularly a service aimed at entertainment. However, it has been shown in this thesis that such advertising would only generate minimal income with a web telerobot. Since banner advertising rates are low, significant revenue will only be generated when traffic volumes are high. So for a device that can be operated by, at most, a small number of simultaneous operators, volumes can never be high enough to be profitable. According to Zipf's law, only the top few web sites will ever generate significant revenue from banner advertising and telerobots will never be in that category.

The time interval between requests to the telerobot is the sum of an operator's decision making time and the time taken to complete the request and advise the operator of the results. An analysis showed that the average time taken per request drops sharply in the first few requests. After eight or so requests, the mean time per request drops only marginally from about one minute to a minimum of 50 seconds for long sessions. This suggests that on average, an operator has developed sufficient understanding to interpret and respond to the interface in close to the minimum time after eight requests. The amount of time per request also becomes less variable as the number of requests made in a session increases.

A new method for assessing operator preference telerobots has been proposed in this thesis. It is based on the finding that the number of operator requests made to the telerobot in one session follows a Weibull distribution. This was discovered by comparing the various statistics gathered with a range of standard statistical distributions. A Weibull distribution is most often used in reliability engineering to characterise machine failure. It is a two parameter distribution with a parameter which defines the shape of the distribution and another which defines the scaling. Previous research has documented a number of social phenomena that could be fitted by a Weibull distribution. These include length of courtship before marriage, length of marriage before divorce and length of conflicts, such as industrial strikes or wars. The conditions that give rise to a Weibull distribution also apply to telerobot operators. That is, that having begun to issue requests to the telerobot operators may stop for any one of many reasons and each of these have a probability of occurring at any particular time.

The number of requests per session for three quite different telerobots and interfaces were fitted by a Weibull distribution with a Chi squared level of significance greater than 0.05. This fit was, however, made after excluding the first few requests, where it was hypothesised that a mixture of event distributions was occurring due to some operators accessing the telerobot when it was faulty and therefore these operators did not make many requests. In reliability engineering, the shape parameter relates to the instantaneous failure rate. A machine that fails due to wear of components has a higher likelihood of failing at any particular instant as its life is extended and its shape parameter will be greater than one. Electronic components have a higher failure rate during their burn in period so the instantaneous failure rate declines with increasing machine life and the shape component in this case is less than one. A shape parameter of one means the instantaneous failure rate is unrelated to machine life and the distribution collapses to the simpler exponential distribution.

The shape parameter for the distribution of requests to the telerobots varied between 0.42 and 0.52, which means the likelihood of making additional requests to any of the telerobots increased substantially as more requests were made. For the ABB1400 telerobot in Perth only 57% of those who gained control made a further request, but 93% of those who had made seven requests will make at least one more. Thus, given that operator behaviour conforms to the Weibull distribution telerobots can be characterised and compared with only two numbers. For example, the distribution of session lengths for each of the telerobots looks quite different (see Figure 91) but the shape parameters are similar so that by rescaling, one distribution can be mapped onto another with only small differences (see Figure 93). While it is unclear why the shape parameter is so insensitive to major system changes, this property can be used to reduce the comparison between telerobots to a single number. If an assumption is made that humans will make more requests to a telerobot which they prefer, this number can be used to quantify operators’ preference for one telerobot over another. The scaling factor can be determined without ever fitting a Weibull curve to the data sets by finding a value for the scaling factor that will minimise the least squares difference between the two data sets after rescaling one of the data sets according to where q is the scaling factor and s is the number of requests. Having estimated a scaling factor between data sets, it is desirable to determine the quality of the estimate. This is affected by the size of the data sets, the degree to which they match the Weibull distribution and the extent to which they satisfy the assumption that both have the same shape factor. A good measure of estimate quality is the correlation coefficient between the data sets after rescaling because the measure of the quality of the estimate of the scaling factor is largely independent of the value estimated. This scale factor was used to compare the three different telerobot configurations. The scale factor between the first IRb6/L2-6 Perth telerobot and the IRb6/L2 Carnegie telerobot was 6.23 suggesting a much higher level of satisfaction with the IRb6/L2 Carnegie telerobot. The correlation coefficient was 0.94. The scale factor between the IRb6/L2 Carnegie telerobot and the ABB1400 Perth telerobot was 1.03 suggesting a marginally higher satisfaction with the Perth ABB1400 telerobot.

The telerobots built have been excellent test beds for telerobotic investigation and there is considerable potential to continue with them to investigate interface improvements by implementing ideas and monitoring operator reaction to them. However there is always the risk, particularly as more sophisticated ideas are tried, that the techniques developed will be limited in their application to the test task. Therefore, I would prefer to see a more useful application developed and ideas tried out with it. Applications that spring to mind include controlling a forklift for materials handling, operating an engraving machine or operating underground mining equipment but there are many others.

One issue that has not been addressed in this thesis is the time taken to complete a task and there is scope for considerable work in this area. The time required to complete a task has always been a major issue with telerobotics as it generally takes much longer to complete a task with a teleoperator than when the task is carried out locally. The low bandwidth available to web teleoperators makes this more of a problem than it usually is with traditional telerobotics. I believe supervisory control will allow many tasks to be carried out as quickly by remote web teleoperation as by local operators but this is unproved and it is an area that needs considerable work.

The web lends itself to collaboration in developing applications from components provided by different groups in a way that has not been possible previously. Some services provided by others were incorporated into the telerobots built including a counting service by Webside story described in section 5.4 and a computer generated image of the robot in its current pose provided from the University of Braunschweig, Germany and described in Chapter 4. There is considerable potential for greater collaboration in this area. For example, one group could specialise in image analysis for scene interpretation and perhaps another in modelling environments. Someone building an application with a robot can link the services together so that images acquired by the application developer are interpreted by a computer run by the scene interpretation group with the scene data provided to a model uploaded as an applet from the modelling group's server. This level of cooperation requires investigation and development standard methods for interacting between these components. Barney Dalton is continuing work on these ideas with the ABB1400 telerobot and some other groups for example Leu et al (1998) are pursuing this exciting opportunity also.

It has been argued that robotics has made a significant contribution to society. However, the reason that fewer robots are deployed, than initially anticipated is the complexity of developing robotic solutions. It remains to be seen whether the solution proposed, web telerobotics, will have a major impact on the contribution of robotics to society but it has already generated considerable interest.