However, a comparison of Figure 17 and Figure 18 shows the final product was considerably less sophisticated than originally envisaged. The robot speaks but does not accept voice commands. Instead, HelpMate is dispatched to stations by the push of a button and it has no manipulator. Moreover, the fetch and carry task of Figure 16 had to be simplified so that a human loads and unloads the compartments in the robot. Despite Engelberger's leadership in the robotics field, he found, like the household robot companies of the 1980s and many other robotics practitioners, that it was easier to modify the application than it was to build a sufficiently flexible robot to handle the complete application.

In 1997, there were 109 Helpmates in service in US and Canadian hospitals. Despite building simpler robots than originally planned, the HelpMates are not cheap. Persuading hospitals to purchase or lease the robots has proved to be difficult and the large number in use was achieved by renting rather than selling. This imposed severe strain on the company's working capital especially since the rental fees were below cost and Helpmate Robotics at the time of its 1997 annual report had a negative net worth of US$369,000 (HelpMate Robotics Inc 1998b). Since the 1997 annual report, the share price has continued to decline from around US$1.00 to under US$0.20 in June 1998.

An independent report (Rossetti and Felder 1998) has indicated that the robots can be significantly cheaper than human couriers for both pharmacy and clinical laboratory delivery but the comparison is not simple. For example, a single robot will be unacceptably slower than a courier, so that several robots are needed to replace a single courier. Robot unit costs reduce as more are employed; therefore, the economics become positive for the robots when they are assigned more tasks in a large hospital. In one modelled scenario, six robots were employed to replace three couriers providing both pharmacy and clinical delivery services. The combined delivery had a 75% decrease in cost and a 38% decrease in delivery variability while virtually matching the average delivery time of the courier. They conclude that "through simulation we were clearly able to demonstrate that mobile robots can cost effectively meet the delivery requirements of mid-sized hospitals" (Rossetti and Felder 1998). However most hospitals using HelpMates have only one robot and the largest number, in seven hospitals, is only three (HelpMate Robotics Inc 1998a).

Service robots discussion

The predictions of Engelberger have been compared with developments in robotics in the following ten years. This has contemporary relevance because many of the ideas put forward by Engelberger are mainstream in today's robotics research community and all the areas suggested by Engelberger have attracted the interest of researchers. Robots have been successfully applied in some service applications including space and undersea remotely operated vehicles. As well, robots are commonly used for hip replacement surgery and mines have adopted teleoperated control of underground mining equipment. However, other applications seem no closer now than they did when Robotics in Service was first written, in particular, the household robot.

Many projects have encountered the same problems experienced with the industrial application of robotics. That is, the problem has frequently been more difficult to solve cost effectively than was anticipated. Engelberger's robotic nursing assistant is a typical example where the initial concept had to be simplified to the point that much of the utility of the original idea was lost. Moreover, the high capital cost of a HelpMate nursing assistant meant that to attract customers the machines had to be rented at below cost rather than sold. As a result, the company has a high probability of going the way of the household robot companies of the 1980's. While use of service robots will continue to increase, it seems unlikely they will replace industrial applications as the high growth area for robotics. Although very successful in some areas, robots are still struggling in many industrial applications even though the environment is more structured, the tasks more repetitive and high capital costs less of an issue than for most service applications.

Why have the ambitions not been met?

It would be incorrect, as is sometimes claimed, to suggest that robotics has not been a success. The statistics presented in section 2.3 show it to be an industry that has come into being, grown very quickly, continues to expand at a respectable rate, and is moving steadily into new application areas. However, many researchers see robotics as an unfulfilled promise. In Engelberger's view, "At this writing the industrial robot has become a 'ho hum' product. To some a robot is no more than a mechanical computer peripheral"(Engelberger 1989:13). This sums up the view of many researchers in the robotics community who would like to see a far more visionary approach to robotics and the adoption of technology that they have devoted considerable effort to developing. But perhaps this is not a realistic view.

The larger, more ambitious projects such as the Autonomous Land Vehicle project and the sheep shearing research project have shown just how difficult it is to apply robotic technology to entirely new tasks. In the sheep shearing project the technical problems were solved and hundreds of successful trials and demonstrations took place. This is, however, only a small part of building a commercially viable product. To gain the benefits of the automated shearing process there needs to be automated sheep handling to get the sheep to and from the robot, automatic fleece handling and automatic fleece inspection. The robot must operate quickly to shear large numbers of sheep over a short period. To be economically viable the entire system has to be reliable and maintainable by operators in remote areas with limited training. Because of the high capital cost of the equipment, it must be mobile rather than a permanent fixture at each farm. The solutions require straight engineering but the engineering effort is substantial and estimated to require a far greater investment than the development of the sheep shearing robot itself.

The sheep shearing project was an example of developing the robot but not having the resources to develop the supporting infrastructure. The Autonomous Land Vehicle project, however, is a good example of the other major difficulty facing robotic applications, that being the large gap between almost solving a problem and a working system. The two overland vehicles demonstrated successfully in 1997 were not developed to the point where they could be deployed in a working environment and it is planned that the highway vehicles which were able to traverse 12,500 miles with less than 1% in manual mode in 1997 will not be developed to prototype stage before 2002.

Another problem encountered with the deployment of robotic solutions is the complications of working with existing systems and the length of time it takes to change attitudes. An example of this is automatic fuel refilling for road vehicles discussed in section 2.4.4. The three systems in use by 1996 were only available for specially modified fleets. Schraft et al (1996) then developed a system which can refuel over 80% of vehicles with their filler cap on the right hand side of the vehicle. Having largely solved the problems associated with a large variety of vehicle designs there remains the issue of cost and convincing the public to use the system.

The obstacle most often thwarting the ambitions of roboticists and underlying most of the problems discussed above is complexity. Robots can be constructed to satisfy any given input-output specifications, ie. we could construct robots that would behave in any way we desire under any environmental circumstances. However, the number of combinations of inputs quickly becomes immense. For example in a system with n sensors with each having a binary state running for K time steps, there are 2nK possible states for which an appropriate output is required. To reduce the combinations to a manageable number, a large proportion of these states can be grouped, as they require the same response and many relationships can be developed between others. As with the earliest robots this is still the way that robotic systems are constructed. However in the natural environment, the number of input-output specifications is frequently too large for any engineering solution. The response of engineers has been to control the environment in which the robot operates and limit the tasks it performs in order to limit the number of possible input-output combinations to a number and level of complexity amenable to engineering solutions. They must then provide sensors able to detect the relevant input conditions and finally develop the logic to handle all possible input-output combinations.

One solution proposed, to avoid engineers having to develop control logic, is controlling robots with neural networks. The neural network is trained by defining an appropriate response to each combination of inputs. The neural network will then learn by creating the logic to reproduce the response whenever the input is repeated. The limitation of this technique is that, when time is included as a variable, the combination of circumstances encountered usually becomes too large for the neural network to be adequately trained. A robot controlled this way would be forever encountering new situations negligibly different from those encountered previously but for which it has no appropriate response. Culbertson (1963) recognised neural networks were inadequate for dealing with the real world and subsequently developed a theory of consciousness which he thought could be applied to robotics. Great effort has since been devoted to artificial intelligence research, but in 1998 artificial intelligence still only exists in highly restricted forms unsuitable for the control of robots. What is needed to achieve the ambitions of roboticists are robots that are able to learn, build up an understanding of their environment, self correct and accept high level commands in a similar manner to humans. The technology to do this with machines remains unavailable.

Is Replacing humans the right objective?

In the early literature on robotics, the major objective was to replace human labour with robots. Current robot installations and research pursue other objectives such as precision control, consistent product quality and increased human safety. Nevertheless, frequently the motivation for robotic applications remains the desire to lower costs by replacing human labour with cheaper machines. The motive is based on the premise that human labour is expensive. In fact this is not always true. The cost of human labour varies enormously. The variation is partly due to skill levels, partly to the social standing of particular occupations, and partly to the industry in which a person is employed. Another determinant of labour cost is geographic location, for example, those employed in remote locations are usually compensated for the harsh environment. However, a major determinant of the cost of human labour is political boundaries. People in some countries earn much less for the same work as people in others which can be shown by comparing annual earnings for people engaged in manufacturing in different countries. Table 4 shows the annual earnings in US dollars for workers engaged in manufacturing for a group of countries chosen to show the range. The list is ranked from the country with the lowest incomes to that with the highest incomes. Error is introduced by the requirement to convert the source data to annual earnings and United States dollars to enable a comparison between countries. However, the variation in annual earnings is so substantial that inexactitude does not detract from the evidence that there is a wide variation in annual incomes associated with political boundaries. For example, Table 4 shows the average German manufacturing worker enjoys 50 times the annual earnings of the average Chinese manufacturing worker.

Many of those living in poor areas seek to improve their situation by moving to a location where the circumstances are more favourable, but immigration restrictions frequently limit people's ability to move. Another method by which those in poorer nations seek to take advantage of the discrepancy in labour costs is by selling their labour in the form of manufactured products in locations where labour costs are higher. However, such trade is often limited by natural barriers, for example the distance to markets and artificial barriers, for example tariffs. In addition, manufacturing and trade requires entrepreneurs, capital, political stability and infrastructure in the form of ports, roads, power etc which are often unavailable to those in poorer areas. Apart from manufacturing, there are many other fields of endeavour in wealthier nations including mining agriculture and service industries which people in poorer nations would like to contribute to if they were given the opportunity.

Those in poorer areas or seeking to avoid harsh environments would be better off if they were able to sell their skills in locations where labour costs were higher without having to be physically present. The intelligence required to carry out a task could be separated from the infrastructure required to physically perform a task. That is, people can provide the intelligence from a convenient location for teleoperated reduced instruction set robots, that are located where the work is required. Communicating intelligence requires only minimal local infrastructure if satellite communications are used. This is an alternative to developing machines that can learn, understand, self correct and accept high level commands in the same way as a human.

A way forward for robotics

A fully autonomous robot would need no instruction or control but this is not what most people have in mind when they seek to build autonomous devices. They envisage limited autonomy and want to be able to issue high level commands such as instructing a robot to clean the house. This is perceived as necessary for a robot to be useful. I have suggested that, as long as machines cannot be made to think like humans, the principle limitation of robotic solutions is the development of appropriate responses for all the possible combinations of circumstances which the robot will encounter. This formidable hurdle continues to constrain robotic applications as it has done since robots were first developed.

An alternative is to provide solutions for subsets of situations the robot will encounter and leave the remainder to be solved by a human as they occur. This could be termed, a reduced instruction set approach to robotics. It lessens the subset of circumstances for which an engineer must develop solutions to a manageable number. Solutions need not be comprehensive. The instruction set can even be reduced to the point that nothing useful is achieved by invoking any particular instruction then relying on a human to generate a series of these instructions to complete a task.

To compare a reduced instruction set solution with a traditional solution consider the example of instructing a wheeled robot of the type in Figure 19 with odometry on the driven wheels to proceed to a new location.

An algorithm to implement the instruction is:-

1. Drive the axles simultaneously in opposite directions at the same speed until the odometers have measured sufficient rotation for the robot to be facing the appropriate direction.
2. Drive both axles forward at the same speed until the odometer has measured sufficient rotations for the robot to travel the requested distance.

This is a simple algorithm, easily implemented, and once completed there is an expectation that the robot will be in the desired location. However there will always be some error due to inaccuracies in the odometry and there are circumstances where this algorithm will fail entirely. These include an obstruction, an uneven surface that produces wheel slip or a broken axle. Thus, the algorithm is simple but not comprehensive. However, the algorithm will be comprehensive if the environment is controlled. The axle can be made strong enough not to break, the floor can be made flat, with a surface that will not allow wheel slip, and all obstructions can be removed. Alternatively, sensors can be added to locate the robot by triangulation and to sense obstructions. Then the algorithm can be improved to plan a path around the obstruction. Traditionally an engineer uses a combination of techniques to build a system that will complete the task with a high degree of reliability. A reduced instruction set approach would consist only of the simple algorithm, accept that it will frequently fail and rely on the human giving the instruction to determine the appropriate response.

The concept is also applicable at higher levels. For instance, assume in the case of the above example that an algorithm has been implemented so that the robot will arrive at a specified location with a high degree of reliability. The requirement might now be to fetch an object. The complexity of the problem has increased enormously. The object must be found, recognised, grasped, transported and placed. A reduced instruction set approach will rely on a human to recognise the object, tell the robot where it is, determine how it should be grasped, where it should be transported and how it should be placed. This is accomplished with a string of simpler instructions to the robot with the human deciding what should be next and monitoring the progress from time to time.

A reduced instruction set approach to robotics simplifies the task for the system builders but the robot constructed needs frequent human attention. A human can get a robot to achieve useful functions by providing the missing "intelligence" and issuing new or modified low level instructions to the robot as circumstances require. Teleoperation provides a method for supplying human intelligence from a location which is remote from the robot. The big advantage is that we are able to avoid the problem of building intelligent machines, which I assert is the principle constraint on the use of robotic solutions. Paul et al have also proposed telerobotics as a method of providing the "missing component of intelligence of the robotics pantheon (Paul et al. 1996:13)" and have coined the term "operabotics" to describe the concept. They have applied the concept to two applications, undersea robotics and control of an experimental multiple agent supervisory control system both of which required purpose built operator interfaces. Telerobotics has been used extensively in hazardous environments but providing purpose built operator stations and communications has made it prohibitively expensive for all but cost insensitive applications such as the Sojourner Mars Rover. However, the World Wide Web provides an opportunity to eliminate the expense associated with operator stations. An appropriately configured reduced instruction set robot can be controlled with human intelligence by anyone with access to a web browser. There are millions of operator stations and the opportunity for people to apply their intelligence at any location regardless of their own. It also changes the direction of telerobotics fundamentally, opening whole new application areas. For example, Sheridan points out that: "There has been surprisingly little interest in human supervisory control for industrial robots and the availability of a nearby human operator ... has been taken for granted (Sheridan 1992:107)". But, as discussed in section 2.6, cheap operator workstations offer the opportunity for those in poor countries to sell their labour where wages are higher. There are also advantages in making expensive industrial machinery available to people who can not afford to own it, opportunities to provide entertainment and opportunities for service applications, such as the control of underground mining equipment. Consider for example the HelpMate robots discussed in section 2.4.10 which provide courier services in hospitals. They are instructed to proceed to preset locations by the push of a button, they then they proceed autonomously announcing themselves upon arrival. It is a service likely to be used by many different people all of which have to understand the controls. As well, the HelpMates have a limited number of preset places to go, the item carried might be intended for someone who is not there and the HelpMates might not be in the right place at the right time. Imagine how much the service would be enhanced by providing the HelpMates with human intelligence so that they could accept voice commands as originally intended. The humans in command could then direct the HelpMates to anywhere in their range, search for people or reinterpret commands as required and anticipate demand. As the HelpMates can perform most courier tasks autonomously, modifying the HelpMates for web teleoperation would allow one person from anywhere in the world to supervise several cheaply. This would enable the HelpMates to provide an intelligent courier service.

With web teleoperation there are a host of new questions raised associated with a large number of frequently inexperienced users rather than a small coterie of highly trained operators, for instance do people find controlling a telerobot entertaining? There are new opportunities for analysing the actions of large numbers of users to learn how they interact with the system rather than designing experiments to be conducted on small but well controlled sample groups. There is also an extensive range of peripheral issues such as how to promote a web-based teleoperable service and how those using the service can be persuaded to contribute to its development. The remainder of this thesis deals with control schemes for web teleoperation, building three web telerobots and investigating the new issues that web telerobotics raises.