Tele-presence aided teleoperation of semi- autonomous work ...


Published on

  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

Tele-presence aided teleoperation of semi- autonomous work ...

  1. 1. Helsinki University of Technology Department of Automation and Systems Technology Automation Technology Laboratory Jussi Suomela Tele-presence aided teleoperation of semi- autonomous work vehicles AS-84.147 Kurssimateriaali
  2. 2. CONTENTS CONTENTS _______________________________________________________ 2 1. Introduction ___________________________________________________ 4 1.1 Background_________________________________________________ 4 1.2 Motivation of this work ______________________________________ 7 2. Teleoperation (and supervisory control)_________________________ 9 2.1 History _____________________________________________________ 9 2.2 Applications _______________________________________________ 10 2.2.1 Space __________________________________________________ 10 2.2.2 Underwater ____________________________________________ 11 2.2.3 Military and antiterrorist _________________________________ 12 2.2.4 Medicine _______________________________________________ 13 2.2.5 Heavy Work Vehicles ____________________________________ 14 3. Definitions ___________________________________________________ 16 4. Teleoperation Interfaces ______________________________________ 20 4.1 Direct _____________________________________________________ 20 4.2 Multimodal / multisensor ___________________________________ 20 4.3 Supervisory control ________________________________________ 21 4.4 Novel _____________________________________________________ 21 5.1 Direct teleoperation (short delay) ___________________________ 23 5.2 Move and wait teleoperation (long delay) ____________________ 24 6. Tele-presence ________________________________________________ 26 6.1 Vision _____________________________________________________ 26 6.2 Hearing ___________________________________________________ 27 6.3 Touch _____________________________________________________ 27 6.3.1 Force feedback (kinesthetic information) __________________ 27 6.3.2 Haptic feedback (tactile information) _____________________ 28 6.4 Vestibular sensors__________________________________________ 28 6.4 Virtual and augmented presence ____________________________ 28 6.5 Enhancement Level of Tele (virtual)-presence________________ 30 6.6 Problems of Tele (virtual) presence _________________________ 30 7. Semi-autonomous Work vehicles _______________________________ 32 7.1 Motivation _________________________________________________ 32 2
  3. 3. 7.2 Related Work ______________________________________________ 32 8.Teleoperation experiments ____________________________________ 33 8.1 Introduction _______________________________________________ 33 8.2 Experiments in an unstructured environment (Papers IV and VI) 33 8.2.1 Test Equipment _________________________________________ 33 8.2.2 Experiments _________________________________________ 34 8.2.3 Results_________________________________________________ 36 8.3 Experiments in a structured environment (Papers V and VI) ___ 38 8.3.1 Test vehicle and equipment ______________________________ 38 8.3.2 Tele-existence equipment________________________________ 39 8.3.3 Test persons ____________________________________________ 40 8.3.4 Tele-existence configuration _____________________________ 40 8.3.5 Test runs _______________________________________________ 41 8.3.6 Evaluation methods _____________________________________ 41 8.4 Results ____________________________________________________ 42 8.4.1 Servo camera and monitor _______________________________ 42 8.4.2 HMD with mono vision ___________________________________ 43 8.4.3 HMD with stereo ________________________________________ 43 9. Visual flow in teleoperation ___________________________________ 45 9.1 Test Setup_________________________________________________ 45 9.2 Test drives ________________________________________________ 46 9.3 Preliminary results _________________________________________ 47 10. Conclusions__________________________________________________ 49 11. References __________________________________________________ 50 12. Appendices __________________________________________________ 53 3
  4. 4. 1. Introduction 1.1 Background Traditionally teleoperation has been used in applications where normal on-board manual operation/control cannot be used or where it would be too hazardous or expensive. Typical examples are the handling of nuclear materials (dangerous), control of small models (impossible) and space and underwater exploration (too hazardous and expensive). The history of modern teleoperation began at the end of the 1940’s when the first master – slave manipulator was developed in the Argonne National Laboratory for chemical and nuclear material handling [Vertut and Coiffet, 1985]. After that, the development of teleoperation was fast. Adaptation of video technology and force feedback to teleoperation made the first telepresence systems possible. Computer technology brought the advanced control loops into the remote (teleoperator) end of the system, and finally brought virtual reality into teleoperation. Despite progress in the technology, the traditional idea of teleoperation was based on the idea that the human operator would at all times be available to exercise more or less direct control. Meanwhile, computer technology made it possible to automate complicated factory processes like those performed in chemical plants, by paper machines and in different batch processes. Little by little, automation technology spread almost unnoticed into mines, farms, forestry, and construction sites also. However, unlike the factory tasks, the difficult and dynamic work tasks of work machines operating in changing outdoor environments were not easy to automate, and, even today, the huge majority of work machines are still manually controlled from the cockpit of the vehicle. Automation technology is still used to assist the driver only. However, some of the tasks can be automated. Typical examples can be found in mining, where automation is perhaps at the highest level among heavy work vehicles. The easiest tasks to be automated in mining are the hauling and dumping tasks that are performed during the work cycle of an LHD machine. The only work for which a driver is needed is the loading task. Another example is a modern drilling machine. The driver only brings the machine to the drilling place; the actual drilling is done automatically. The driver mainly monitors the task and intervenes in the case of malfunction. These machines are called semi-autonomous. In work-vehicle automation – like in all automation technology – the target is to reduce production costs and improve quality as much as possible. Normally – in the long run – this leads to a fully automatic system where the human only supervises the process. This has already happened in the process industry. In Figure 1, the history and a future vision of the automation level in forestry vehicles is illustrated. This vision can be adapted to any task where heavy mobile machines are used. Teleoperation is, more or less, only an intermediate phase in the development of fully automated systems. However, the need for teleoperation will last at least for the few decades before fully autonomous work vehicles emerge, and even then human supervision will always be needed. 4
  5. 5. aut onom ou sit y / int ellig enc e aut onom ou s t hinning and br ushing r obot soc iet ies au t onom ou s f or w ar d er sem i-aut onom ou s r em ot e oper at ed t eleop er at ed m u lt i-m ac hine w heel/t r a c k f or est har vest ing m an or leg based w or k sit e & har vest er m an wal k ing & har vest er w heel/t r ac k based m an har vest er & MAN t r ak t or & HORS E t im e 1950 1970 2000 2010 Figure 1: A vision for future development in forestry. Adapted from [Halme and Vainio, 1998] The reason for the lack of teleoperation in industrial work vehicles is clear. In industrial tasks there are practically no advantages in utilizing teleoperation. There is no financial connection between cases where the driver is driving in the cockpit of the vehicle and teleoperating from a control room, if one machine takes his full attention. The elimination of the cockpit would – of course – bring savings, but these would be compensated by the cost of teleoperation equipment and the high bandwidth data transmission infrastructure. There are, however, some applications where direct full time teleoperation is used. For example, in some mines there are areas that are productive but – for some reason – not totally safe. These areas, where human operators are not allowed to go, can be mined by teleoperated machines. Semi-autonomous work vehicles raised a commercial need for teleoperation. When the simple tasks of a work vehicle are automated, the human operator is enabled to focus his efforts on the more demanding tasks. This means that, for all the time the machine is running autonomously, the operator is free to do something else, for example to control an other machine. In traditional operation, this doesn’t work, because a change over from one vehicle to another is impossible because of time and – in most cases – safety. If these tasks that need manual control are teleoperated, the driver can easily, speedily and safely operate two or even several machines. This so called part time teleoperation combines the advantages of the traditional direct teleoperation with those of the advanced automation to decrease the need for manpower in work-vehicle operation. Part time teleoperation, as a time and money saving method, is a relatively old invention. Especially in unmanned electric distribution stations, simple tasks like 5
  6. 6. reconnecting switches, etc. are usually teleoperated in order to speed up the process and to save the time of service people. Workmachines High Bandwidth Bus Control Station Figure 2: Principle of semi-autonomous work vehicles Partially teleoperated semi-autonomous working machines are without doubt the next step on the way to developing automation in worksites like mines or construction sites. Simple tasks like drilling or hauling ore in mines can be automated, whereas more sophisticated tasks like selecting the start point of the drill or loading the ore need the hand of an experienced operator. When most of the time in the work cycle of working machines is automated, one operator can easily handle from two to five machines, assuming that all the machines can be teleoperated from one place. In Figure 2, a typical layout of a semi-autonomous work machine system is shown. In the future, the development of robotics will bring service robots really among humans, not only in industry, but also in homes and other places where service work is needed. Typical examples of these new service robots are Honda’s humanoid Asimo [] and HUT’s centauroid WorkPartner [Halme et al., 2000], [Halme et al., 2001]. These robots set totally new demands for teleoperation. Again, it will take years or decades before service robots will be able to perform autonomously even the simplest tasks. Meanwhile, there have to be teleoperation methods to teach new tasks to the robot, or to help it in faulty or exceptional situations. In the case of commercial services, where there are several robots working under the same “employer”, the situation can be more or less similar to part time teleoperation. One human can control a group of robots from a control room utilizing all the practical teleoperation methods like tele-presence and high bandwidth closed loop control. 6
  7. 7. Figure 3: Asimo and Workpartner robots [Asimo adopted from:] What to do when there is only one service robot? This is a typical situation in homes and in most cases where operator and robot are working together. Now the operator is usually relatively near to the robot and can carry only very light and small equipment for robot control, or none at all. Novel control methods like speech and gestures [Paper III] or even brainwave [Amai et al., 2001] control provide humanlike interaction with the robot, but the autonomy of the robot has to be improved. 1.2 Motivation of this work The evolution of teleoperation has generated sophisticated tele-presence systems where the operator can really feel that he is present in the teleoperation site. When looking at the related research in the area of teleoperation, it may be noted that most of the research has been done in order to provide better and more effective teleoperation methods for difficult work and manipulation tasks where stereo vision and anthropomorphic manipulators with force feedback are needed in order to perform the task. This development is, of course, very natural for any technical area. The (wretched) motivation in the industry is to earn more money. This usually means that not the best system in terms of technology is chosen, but rather the most suitable system, or the most cost efficient. In heavy work vehicles the subsystems are usually not the most state-of-the-art, but simple, well-tested and reliable systems that will work for thousands of hours without failure. Teleoperation systems are still so rare among work vehicles that there has not been much research at the required level of teleoperation or telepresence. The obvious thing is that the required level of telepresence depends on the task. The operator needs a different amount of information in the case of operating a road roller from the amount needed in the case of operating a harvester in unknown forest. The aim of this study is to test and compare different levels of telepresence equipment in the operation of common work-vehicle tasks. When an adequate and/or the best composition is found, it is analyzed into the properties of the task that define the level of presence needed. Also, the effect of learning and the different performance of the drivers are studied. The study is based on tests with both an 7
  8. 8. experimental and a real-work vehicle. The test results are evaluated both objectively and subjectively. 8
  9. 9. 2. Teleoperation (and supervisory control) 2.1 History The poking of fire (Figure 4) might have been one of the first general teleoperation tasks in the history of mankind. To be exact, the poking of fire is tele or remote manipulation, which was the earliest type of teleoperation. This task is also a good example with which to demonstrate the difference between teleoperation and tool utilization. A human hand is a perfect tool for setting the firewood better and, in fact, usually the unfired wood is set by hand in the fireplace. After the fire has been set, the environment is so hostile that a more bulky tool must be used in order to protect the hand. Tools make it possible to perform a task like cutting (a knife), or to improve work like digging (a spade). Figure 4: Primitive teleoperation task The first modern master - slave teleoperators were mechanical pantographs. The group working under R. Goertz developed these manipulators in the late 1940s at the Argonne National Laboratory, where Enrico Fermi developed the first nuclear reactor [Vertut and Coiffet, 1985]. The need was obvious. The radioactive nuclear material has to be manipulated safely. The nuclear material was placed in a “hot cell” where the operator could manipulate it outside the cell by remote handling (Figure 5). The visual contact with the target was through a protective window and/or a mirror. 9
  10. 10. Figure 5: R. Goertz and the first mechanical master - slave manipulator. Chemicals are manipulated remotely behind a protective glass. [Adopted from: Vertut and Coiffet, 1985] The mechanical manipulators were soon replaced by electro mechanical servos. In 1954, Goertz’s team developed the first electro mechanical manipulator with feedback servo control. After this, the teleoperation of manipulators and vehicles extended rapidly to new branches where the advantages of teleoperation techniques could be utilized. One of the first areas where teleoperation techniques were utilized was deep-sea exploration. The deep oceans are even today regarded as so hostile that most of the deep-sea operations are made with teleoperated submarines. These submarines are called – even today - Remote Operated Vehicles (ROV), though the term could be equally well understood as referring to ground, or even flying, vehicles. Often ROVs are equipped with telemanipulators in order to perform underwater work tasks. Teleoperation is also typically used for space and military applications. In both cases, the environment is hostile for humans, while in space applications there is the additional point that the extra equipment needed for the human pilot is more expensive than a sophisticated teleoperation system. 2.2 Applications 2.2.1 Space Space applications provide several good reasons for teleoperation. Safety – in all space operations there are big risks, which can lead to the loss of astronauts’ lives. There have also been scenarios in which teleoperated mining sites have been built in space. Costs – in space operations the equipment needed for human passengers is much more expensive and weighs more than teleoperation technology. 10
  11. 11. Time – long space missions can take several years, which is not possible for manned flights. The first successfully teleoperated vehicle on the moon was Russian Lunakhod 1 (Figure 6) at the start of the 1970’s. Lunakhod ran 11 days over 10 kilometers. The Lunakhod mission also faced the problem of teleoperation over long time delay, which is typical for space missions. Already the delay of several seconds to and back from the moon made the fast closed loop control impractical because of the resulting instability. Instead of closed loop control, a “move and wait method” is used. The vehicle is operated by open loop commands without immediate feedback. After one or more commands have been executed, the operator waits for the confirmation and feedback. A much longer delay was faced at the end of 1990 when NASA’s Sojourner landed on Mars. Despite the 10 - 20 min. control delay, Sojourner was successfully operated over the planned 7-sol (Martian day) period. Figure 6: Teleoperated space rovers: Lunakhod 1 and Sojourner In Paper 1, a teleoperated rover for planetary soil sampling missions is presented. This experimental RoSA2 is designed to be controlled with a “move and wait” strategy improved with a 3D-model of the environment and a virtual model of the vehicle, which can be closed loop controlled. 2.2.2 Underwater As mentioned before, underwater operations were one the first mobile applications where teleoperation techniques were adopted. Today these ROVs probably represent the largest commercial market for mobile vehicle teleoperation. ROVs are used in surveying, inspections, oceanography and different simple manipulation and work tasks, which were traditionally performed by divers. ROVs are generally tethered to a surface ship and controlled using video monitors and joysticks. The most recent system can also perform some autonomous tasks such as station keeping or track following. The French research submersible “Victor” (Figure 7) can dive down to 6000m. 11
  12. 12. Figure 7: Submersible Victor [adopted from:] 2.2.3 Military and antiterrorist The military field provides endless possibilities for teleoperated systems. It was not a surprise that one of the first automotive teleoperators was developed for military applications. Mobile military teleoperators/-robots cover the whole scale from ROVs to the Unmanned Air Vehicles (UAV). In between, there is a wide range of teleoperated ground vehicles. Modern UAVs like US Air Force Predator (Figure 8) are remotely piloted by radio or satellite links. They can also have the capability to fly autonomously with the help of GPS and inertial navigation. Their typical tasks are reconnaissance and target identification. Unmanned Ground Vehicles have a wide application field in military operations. Typical tasks are reconnaissance, surveillance, target acquisition, route clearing, ordnance disposal and landmine detection. The first UGVs were fully teleoperated with closed loop control. The newest models like SARGE (Figure 9) are equipped with vehicle localization (GPS, Inertial navigation) and supervisory control to improve the performance. Military UGVs are often supplied with the state of the art teleoperation equipment like stereovision telepresence etc. to provide best possible feedback in fast and dangerous operations. UGV development has been led by the US military. Figure 8: US Air Force Predator [adopted from:] 12
  13. 13. Figure 9: The Surveillance And Reconnaissance Ground Equipment (SARGE) by Sandia National Laboratories [adopted from:] Increasing criminality and terrorism have created a new variety of military type teleoperators [Davies, 2001], [Hewish, 2001]. These so called terrobots are used for bomb disposal, surveillance in police operations and even assault against dangerous targets. These vehicles are teleoperated with closed loop control over radio or cable connection. Typical of vehicle equipment are color camera(s), infrared cameras, manipulators, hydraulic guns, shotguns and non-lethal guns. Figure 10: Terrobots [adopted from: Davies, 2001] 2.2.4 Medicine In medicine, teleoperation usually takes the form of micromanipulation. In endoscopic operations, the cutting equipment and endoscope are taken to the target through a small hole, while the operator cuts the target causing only minimal damage 13
  14. 14. to the surrounding tissue. The risks are smaller and the recovery time remarkably shorter than in the traditional open wound surgery. Figure 11: Endoscopic nose operation [Adopted from:] Micromanipulation is also a common tool in biochemistry, especially in genetic manipulation. A typical example is cloning when the genotype located in the nucleus of a cell is replaced. 2.2.5 Heavy Work Vehicles In mining, teleoperation has already been in use for two decades in cases where the mining area was not totally safe. Drill vehicles and loaders are driven manually in the safe parts of the mine, but teleoperated in areas where safety can’t be guaranteed [Shyu, 1997]. Teleoperated rescue vehicles have also been developed for mines. Ralston and Hainsworth [Ralston and Hainsworth, 1998] present a mine emergency response robot called Numbat. It operates in conjunction with rescue teams as they enter a mine after an emergency. It moves ahead of the teams under remote control through the mine and transmits to the surface video or infrared images of the hazardous areas, as well as data on the atmospheric conditions. 14
  15. 15. Figure 12: Mine emergency response robot Numbat [adopted from] 15
  16. 16. 3. Definitions Definitions are to help the reader to understand what the writer has meant with the word in this document. Most of these definitions are generally approved by the robotics research community but, as is the case with all definitions, there are usually different opinions. Robot 1: A robot is a re-programmable, multi-functional manipulator designed to move material, parts, tools, or specialized devices through variable programmed motions for the performance of a variety of tasks. [Robot Institute of America] Robot 2: Any automatically operated machine that replaces human effort, though it may not resemble human beings in appearance or perform functions in a humanlike manner. The term is derived from the Czech word robota, meaning “forced labor.” [Encyclopedia Britannica] Autonomous robot is something, which is not available now, and will be extremely difficult to build in the future. Animals and humans are autonomous. To be autonomous a robot has to have consciousness, which cannot be created by existing computer and software technology. (This is the writer’s own, rather pessimistic, view.) Usually the term autonomous robot is used for a robot that can execute its task(s) autonomously without an operator's help. The degree of difficulty of the task does not affect this. Operator = A human operator is the person who monitors the operated machine and takes the control actions needed. Teleoperator is the teleoperated machine. A sophisticated teleoperator can also be called a telerobot. Teleoperation means simply to operate a vehicle or a system over a distance [Fong and Thorpe, 2001]. However, more exact definitions are needed to separate the poking of fire from high-level supervisory control. The first teleoperation tasks like poking fire or manipulating nuclear material can be classified as remote operation or remote manipulation. The word "remote" emphasizes that the controlled vehicle or system is, for most of the time, in the view area of the operator. Today, in “normal teleoperation” there is no visual contact with the controlled machine. The visual feedback is made (usually) by a camera – monitor combination. Control commands are sent electrically by wire or radio. Where the connection between the manipulator and operator is mechanical, the term "remote manipulation" means mechanical manipulation. In tele-manipulation, this connection is electrical. Here the word "teleoperation" has been used both in its widest sense covering all meanings, and in the sense meaning just the “normal teleoperation” defined above. Between the simple mechanical manipulation and high-level supervisory control, there are several systems of different technical levels included under the term teleoperation. 16
  17. 17. Mechanical manipulation: The control commands are transmitted mechanically or hydraulically to the teleoperator. Visual feedback can be direct or via a monitor. Remote operation/control: The operator has direct visual contact most of the time with the controlled target. Control commands are sent electrically by wire or radio (Figure 13). Figure 13: Remote control of a drilling machine To clarify the wide concept of “normal teleoperation”, this can be divided easily into three different sublevels: Closed loop control (Direct teleoperation): The operator controls the actuators of the teleoperator by direct (analog) signals and gets real-time feedback. This is possible only when the delays in the control loop are minimal. A typical example of this is a radio controlled toy car (Figure 14). Figure 14: RCtoy car (Tamiya) and controller (Futaba) example of closed loop teleoperation [adopted from: and] 17
  18. 18. Coordinated teleoperation: The operator again controls the actuators, but now there is some internal control - remote loop (the blue line in Figure 15) - included. However, there is no autonomy included in the remote end. The remote loops are used only to close those control loops that the operator is unable to control because of the delay. A typical example of this is a teleoperator for whom the speed control has a remote loop and, instead of controlling the throttle position, the operator gives a speed set point. Digital closed loop control systems almost always fall into this category. In supervisory control [Sheridan, 1992], the remarkable part of the control is to be found in the teleoperator end (compare coordinated teleoperation). The teleoperator can now perform part of the tasks more or less autonomously, while the operator mainly monitors and gives high-level commands. The term task based teleoperation is sometimes used here, but it is more limited than "supervisory control". In Figure 15 the red loop - operator loop - demonstrates feedback from the HMI computer. This can be a virtual model, estimated parameters, etc. OPERATOR OPERATOR OPERATOR display controls display controls display controls HMI computer HMI computer HMI computer transmission transmission transmission teleoperator’s teleoperator’s teleoperator’s computer computer computer sensors actuators sensors actuators sensors actuators TASK TASK TASK Figure 15: The first figure demonstrates closed loop control; the second and third, supervisory control Telepresence (tele-existence): When a sufficient amount of sensor information (vision, sound, force) is brought from the teleoperator site to the operator, then he or she feels physically present in the site. 18
  19. 19. Virtual presence (or virtual reality) is similar to telepresence, except the environment where the operator feels to be present (and the sensor information) is artificially generated by a computer (Red loop in Figure 15). Augmented presence (or augmented reality) is a combination of real world and virtual reality. A typical example of this is a real camera image with additional computer generated virtual information. 19
  20. 20. 4. Teleoperation Interfaces Interfaces partly overlap the definitions presented in the previous chapter. However, interfaces are an essential part of teleoperation and a more profound definition is justified. The presented classification of vehicle teleoperation interfaces, direct, multimodal/multisensor, supervisory control and novel, is adopted from [Fong and Thorpe, 2001]. Often, the applied system can be clearly classified according to these classes, but frequently there are features included from two or more classes. In part time teleoperation, typically at least direct and supervisory controls are included. Despite these problems, this classification clarifies the concepts in the field of teleoperation interfaces. 4.1 Direct The traditional and most common method of vehicle teleoperation is direct control. The operator controls the vehicle via hand controllers (joysticks, or steering wheel) and watches the feedback video from vehicle-mounted cameras. The operator feels as if he is inside the teleoperator, looking out. Direct control is appropriate when the real-time decision making of a human operator is needed continuously. The restrictive feature is the requirement of high-bandwidth and low-delay communications. Even though communication techniques have developed a lot in recent years, delay still exists, especially in digital signal transmissions. The presence of delay is tedious and fatiguing for the operator. Also, the mismatch between different senses during control can create simulator sickness, especially when the operator’s feeling of presence is improved by using telepresence. 4.2 Multimodal / multisensor When a complex robot moves into a dynamic situation, the operator can have difficulties in perceiving the environment and robot’s state, or in performing control actions. Multimodal interface provides the operator with a variety of control modes. Typical examples are separate control of individual actuators with graphical feedback and coordinated motion. Feedback displays also contain multimodal information in graphics and text. A multimodal interface of a legged robot is illustrated in Figure 16. Multisensor interfaces collect information from several sensors and combine it into one integrated display. In vehicle teleoperation, these displays are often used to improve the operator’s depth-judgment or attitude feeling. 20
  21. 21. Figure 16: Multimodal control interface of WorkPartner robot 4.3 Supervisory control Supervisory control has already been defined in Chapter 3. Nowadays, when, in practice, all teleoperation is based on computers and digital data transfer, almost all teleoperation cases can be regarded as cases of supervisory control. 4.4 Novel The term “novel” is somehow misleading because it is relative. Most of the teleoperation interfaces, as well as every technical invention, has once been novel. Thus, it is most likely that the “novel interfaces” presented here will not be called novel in the future. Nevertheless, at the moment they can be classified as such. Some interfaces are novel because the input method is unconventional. [Amai et al., 2001] presents a vehicle-driving controller based on brainwave and muscle movement monitoring. Paper III describes a “cognitive teleoperation interface” where control commands are given by speech or by gestures that are recognized either by image processing or hand-tracker. Direct control of both vehicle and two- hand manipulator is performed with the hand-tracker interface (Figure 17). [Fong et al. 2000] describe the Haptic driver, which enables the drive-by-feel control, and the Gesture driver, which is based on gestures (mapped with the robot camera). [Heinzmann and Zelinsky, 2001] also use gestures, but, instead of hands, they use face gestures and gaze control. 21
  22. 22. The web-based teleoperation interfaces (Figure 18) can also be classified as novel, although they can be classified under the multimodal interfaces or supervisory control as well. The Web is very interesting because it provides global low-cost data transfer for long distance teleoperation [Schilling et al., 1997]. There are also problems, like unpredictable, varying bandwidth and delay, which are characteristic of the Internet. Figure 17: Hand-tracker interface for mobile machine control in outdoor conditions Figure 18: Teleoperation interface in Web [adopted from: http://redrover.ars.fh-] 22
  23. 23. 5. Problems in Teleoperation The main problems in teleoperation are related to delays and the human - machine interface (HMI). In semiautonomous machines, the HMI is even more problematic because, in addition to the direct teleoperation state, there is also a supervision state, and – most difficult – the transition between these two states. The delay problem can be divided into two different cases: direct teleoperation (short delay) and move-and- wait control (long delay). 5.1 Direct teleoperation (short delay) There are always delays in a teleoperation loop, which is like any control loop with controller, process and feedback measurement(s). According to the law of Shannon, the process can be measured (and controlled) only when the measurement frequency is at least 2 times higher than the nominal frequency of the measured process. In teleoperation this means that the delay in the control loop – from control action to the feedback of the action effect – should be at least two times the nominal frequency of the controlled process, otherwise the process frequency has to be decreased. The delay in teleoperation equipment (the human delay is not included) consists of several parts (Figure 19). Nowadays, the digital signal processing causes the major part of the delay. However, the advantages of digital processing are so great that there is no use for analog technology. feedback delay transmission teleoperator delay delay Control delay Figure 19: Delays in teleoperation At the operator end, the control delay is computational and consists of digitizing the values of the control equipment, i.e. of the steering wheel, joystick and pedals. This delay - a maximum of tens of milliseconds - is usually not significant. Even the electric signal has about the speed of light, so there is always delay in the transmission. Digital transmission contains extra delays compared to analog transmission. Typical delays in the digital transmission are between 10 and 100ms. In addition to the control information, the feedback also has to be transmitted back to the operator. Usually this information contains both image and data. Image information is the most critical. Even today, in most cases, analog video links are used and they work with practically no delays. The image compression and decompression delays are included into the feedback delay to separate them from the transmission delay. They can be significant (>100ms) in the digital video transmission. The biggest delay – with the transmission – occurs at the teleoperator end. The control of robot actuators (remote loop), as with the steering and throttle, includes both the control delay and the delays characteristic of the controlled process like mechanical time constants. In most cases, the remote loop (closed control loop in the teleoperator) helps the teleoperation, but one must remember that in all cases where the remote loop is included into the direct teleoperation loop it increases the delays. 23
  24. 24. A very good example of the remote loop delay can be found in Paper V. The loader used has articulated steering with a hydraulic actuator. The manual control of the loader is performed with a joystick, which directly controls the servo valve of the steering actuators, i.e. the position of the joystick is relative to the angular velocity of the steering, not to the angle. In the experiments, the loader has three different (teleoperation) steering configurations: joystick, orbitroll and “servo”. Joystick control was identical with the manual control. In orbitroll control, the angular speed of the steering wheel (frequency of encoder pulses) was transformed to the control current of the valve, i.e. the angular speed of the steering wheel is relative to the angular speed of the steering. Servo steering imitated normal car steering where the steering wheel position is relative to the steering angle. To perform this, a remote controller was needed in the loader. A PID controller was added to the system to control the steering angle according to the set point from the steering wheel. At the start of the experiments, the drivers were offered the possibility of testing all configurations, and they chose the one for the rest of the tests. None of the drivers chose the servo steering. The reason was mainly “too sensitive steering”, but also, all drivers noticed the small but noticeable delay compared to the other two methods. As mentioned before, the delay has to be proportional to the frequency of the controlled process. In vehicle control, the process frequency is relative to the vehicle kinematics and the driving speed. In most work vehicles, the driving speed is low (<40km/h), and the dynamic driving is utilized only in winter conditions. In the experiments made by the author, the vehicle control is trivial when the total delay is less than 100ms and relatively easy when it is below 0,5s. The human part – cognitive and decision making ability – is the most important in the control loop. In case of delay, human learning ability plays the main role. It seems that the control delay is something that humans are used to. In all experiments, the control delay was noticeable (0,3 – 1s), and test drivers learn to compensate for it in only a few minutes. However, it has to be stressed that compensation is possible only when the process is controllable with the existing delay. Virtual models of a telerobot and its environment can also be used for short delay compensation. These so called predictor displays [Sheridan, 1995] present the estimated movements of the robot resulting from a control action in real-time. In this case, the operator can see both the estimated immediate response and the real delayed response (augmented reality). This model-based compensation is more typical in cases of long delay (see next chapter). 5.2 Move and wait teleoperation (long delay) When the transmission delay increases enough, there is no possibility of direct teleoperation. This is a typical situation in space applications where distances are so long that the speed of light is the limiting factor in the delay. The only possibility in these conditions is to increase the autonomy of the robot and use task-based move- and-wait methods. From the operator's point of view, it would be easiest to make the teleoperator highly autonomous and give it only long and demanding tasks to avoid unnecessary operator control. However, in space conditions, the vehicle should be as 24
  25. 25. simple as possible, and only 100% sure tasks can be commanded because errors cannot be allowed. The positive thing in space operations is that time is usually not a limiting factor. In Figure 20, (and Paper I), a configuration of a teleoperated Mars robot is shown. The robot is connected by a tether to the lander. All communication between the operator and the robot goes through the lander. The feedback of the robot movements and the environment comes from the cameras, which are located in the lander. The control sequence is as follows: 1. A stereo image from the rover and the environment is transferred to the operator. 2. A 3-dimensional environment model is created from the image data. This model can be created either in the lander or on the ground. 3. The operator looks at the image and the model and plans a short trajectory for the robot. This can be, for example, “drive 10 cm forward” or “turn 30deg”. In the trajectory planning, the most important thing is to be sure that the task can be executed without problems. 4. The robot executes the task and a new image is taken and transferred to the operator. 5. The operator gets visual feedback from the image, and plans a new task. One must now remember that in the case of Mars, the transmission delay between the robot and operator is at least 15mins., and that the low speed of the transmission (about 1200bit/s) increases the delay, especially in the case of visual feedback. Figure 20: Operational scenario of a Mars robot The models of the telerobot and its environment can be used in task planning like that presented in the previous chapter. 25
  26. 26. 6. Tele-presence Tele-presence simply means that the operator feels that he is present at the teleoperator site. Already the simple camera monitor combination creates some level of presence, but usually a more sophisticated system is called for in order to call it telepresence. The most typical ways to create telepresence are cameras that follow the operator’s head movements, stereovision, sound feedback, force feedback and tactile sensing. To provide a perfect telepresence, all human senses should be transmitted from the teleoperator site to the operator site. A good example of multi- sense telepresence is presented by Caldwell [Caldwell, 1996]. His system provides both hearing and vision in stereo mode, head tracking, tactile, force, temperature and even pain feedback. The vision, hearing and sense are relatively easy to transmit, but smell and taste are more complicated. Fortunately, these two senses are rarely important in machine teleoperation. 6.1 Vision Humans get more than 90% of their perception information via vision. The human vision sensors – eyes – are very complex opto-mechanical systems. They allow stereovision, focusing, fast pointing, and a very wide field of view. The human field of view is 180 deg horizontally and 120 deg vertically. The focused area is only a few degrees, but movements and other interesting targets can be noticed from the whole field. It is extremely difficult to manufacture a teleoperation system that can imitate human vision and provide the operator with the same amount of information as he could get in the teleoperator place. In most visual feedback configurations, from simple monitor to complex telepresence systems, the field of view is reduced because of the camera and monitor technology used. In all monovision systems, the perception of distances is limited because of the lack of depth view. In most cases, there is no need to build up any complex telepresence systems. The simple vision feedback with static camera and monitor is enough for most cases. As in the delay compensation, the learning will also compensate for the limitations in visual feedback. In some cases, an advanced telepresence is needed. To create the presence, a human operator has to be cheated into feeling that he is present in the teleoperator place. To “cheat” a person is primarily to cheat his or hers vision – to see is to believe. It was Goertz who first showed that when the monitor is fixed relative to the operator’s head, and the pan and tilt movements of the head drive the pan and tilt of the camera, the operator feels as if he were present at the location of the camera. Already, the head mounted display with tracked pan and tilt provides clear telepresence for the operator. If the roll, and even the eye movements [Sharkey, Murray, 1997], is tracked, the feeling is even more real. [Tachi et al., 1989] were amongst the first who developed a high performance hardware system for telepresence (Tachi called it tele-existence) experiments. Tachi’s system had a very fast 3 DoF head tracking which – together with a high class HMD - provided a very good feeling of presence []. 26
  27. 27. 6.2 Hearing The total range of human hearing is between 16 – 20000 Hz. The smallest audible intensity depends on the frequency; the minimum is between 1000 and 6000Hz and increases for lower and higher frequencies. In the control of a heavy work vehicle, the noise of the machine is usually so high that the driver uses hearing protectors, and can in general observe only the sounds of his vehicle. Despite damping, these sounds are extremely valuable for the driver. In Paper II, a teleoperation experiment with a mine drill machine is described. It was amazing how the operator could operate the drill by hearing almost only the sounds of the machine. In the experiments of Papers III – VI, it was also noticed that sound was even more important to teleoperation than to manual driving because the drivers got no touch response from the vehicle. The loading of the engine during a loading task, for example, could be felt during manual loading, but only heard while teleoperating. In teleoperation, the electrical transmission of sounds also makes it possible to tune the intensity, and filter the non-informative noise away. It is difficult to create a tele- presence without sounds. 6.3 Touch From a very fundamental point of view, the touch or feel is the most important human sense. Without vision or hearing, the human being can survive amazingly well, but without the sense of touch, he would die relatively soon. Human touch sensors – mechanoreceptors – are activated by touch, i.e. by pressure on the tissues. These sensors are located throughout the human body. They sense the positions and movements of joints, tension in muscles and touch on the skin. These tactile sensors can be divided into two basic classes [Durlach, Mavor, 1995]: 1. tactile information, referring to the sense of contact with the object, mediated by the responses of low-threshold mechanoreceptors innervating the skin (say, the finger pad) within and around the contact region and 2. kinesthetic information, referring to the sense of position and motion of limbs along with the associated forces conveyed by the sensory receptors in the skin around the joints, joint capsules, tendons, and muscles, together with neural signals derived from motor commands. Touch is needed in all kinds of work where the human being is mechanically interfacing with tools and environment – practically in every work except thinking. Even the computer work is difficult without the sense of touch in the fingertips. However, in case of tools or machines, like heavy work vehicles, touch is not focused on the actual task but the control equipment of the machine. In some teleoperation tasks like manipulation, the feedback of touch can help the operator to a remarkable degree. Touch feedback can be divided in two types: force feedback and haptic feedback. 6.3.1 Force feedback (kinesthetic information) Force feedback means that the force generated by the teleoperator, usually a manipulator, is fed back to the operator in order to generate a real response in gripping and manipulation tasks. Among mechanical manipulators, this feature was inbuilt because it was the force of the operator that was using the manipulator. When hydraulic and electrical servos replaced the straight mechanical contact, force feedback was no longer used. Now the feedback was generated artificially by 27
  28. 28. measuring the force from the actuator of the robot and generating it with an additional actuator to the control equipment. In the manipulation tasks, force feedback is essential for a good telepresence. Force feedback can also be used in virtual environments to generate the feeling of presence. 6.3.2 Haptic feedback (tactile information) In the wide sense, both the force and tactile feedback come under the term "haptic feedback". In teleoperation, the main difference between a haptic interface and a force feedback interface is the touch point. In force feedback, the muscular (kinesthetic) sensors give the response to the operator. In the haptic feedback, the tactile skin sensors have the main role. Usually in haptic interfaces, the tactile sensing of the robot manipulator is fed back to the fingers of the operator. But it can also be the vibration of the vehicle or the intensity of the camera that is fed back to the human skin. 6.4 Vestibular sensors Vestibular sensors are located inside the inner ear, and they are sensitive either to angular acceleration and thus rotation, or to linear acceleration in the horizontal and vertical plane, i.e. to gravity. This allows the position and movements of the head to be detected. Vestibular sensing is important in all dynamic work tasks. The driver of a heavy work vehicle gets a lot of information via his vestibular sensors, but also a lot of annoying movements like vibrations. In teleoperation, the vestibular feedback is not used because the control can be made without feedback in almost all situations, and the vestibular feedback needs expensive mechanical structures. The lack of vestibular feedback in the case where the operator is using a head-mounted display generates a conflict of senses, which can generate simulator sickness (see Ch. 6.7). Vestibular feedback is usually used in simulators to provide very natural feeling of presence (see Ch. 6.5). 6.4 Virtual and augmented presence In virtual presence, the operator feels he is present in an environment that has been artificially generated by a computer. Pure virtual environments are usually used only in simulators and games. The most typical examples are flight simulators, which are used both for entertainment and the real training of pilots. Flight simulators were also the first systems where virtual reality was utilized. In flight simulators, the accelerations of the “plane” are also simulated by moving the simulator (and operator) with hydraulic actuators (Figure 21). 28
  29. 29. Figure 21: Finnair’s DC-10 flight simulator In teleoperation, usually the virtual reality is used to augment the telepresence. This is called "augmented presence" (or "augmented reality"). Augmented reality can be used for prediction and planning, for example, in cases where the long time delay disturbs the teleoperation. In prediction, the existing environment is modeled, and, when the operator is operating, the estimated actions of the teleoperator are shown virtually. The real actions are shown in the same display after the delay. This way the operator can do direct teleoperation tasks despite the time delay. However, the estimated actions must be corrected with the real feedback information every now and then. In Paper 2 and [Halme et al., 1997], an augmented reality system is presented where the operator can create and correct the virtual model in telepresence. The telepresence system consists of HMD, head-tracker and 3DoF servo head with stereo cameras and laser pointer. The virtual model is created in a PC with the WTK program. A virtual (3D) model is fixed within the real world by updating it according to the head-tracker information. The real image and virtual model are overlaid in a video mixer, which provides the possibility of overlaying the two sources steplessy. Additionally with this model, the virtual image also contains a graphical user interface (GUI), which can be used with a mouse. When a new, unmodeled object is noticed, the operator divides it into basic pieces, which are modeled one by one. The basic pieces are: box, sphere, cylinder and cone. Modeling is performed by naming the object and measuring a group of points from the surface of the object with the laser pointer. The modeling software calculates and draws a model of the object and places it in the overlaid image. If the size or the place of the object is not matching totally the user can move, rotate and scale the object by using a mouse and the user interface included in the model. 29
  30. 30. Figure 22: a) Real image, measured points and GUI. b) Real image and the model based on the measured points 6.5 Enhancement Level of Tele (virtual)-presence In this work, the manner in which the level of the telepresence affects the performance of the operator in different work-vehicle tasks is studied. The word “level” here refers to the level to which the telepresence system is advanced. In the experiments, the telepresence was concentrated on the vision. The level was changed from a stable camera - monitor combination to servo cameras and HMD via three sublevels. Does the increase in its level also mean an increase in its quality? How can the quality be measured objectively? Schloerb [Schloerb, 1995] presents an evaluation method for telepresence systems where the objective evaluation is to indicate how well the defined task is performed. This leads inevitably to a situation in which the defined task also affects the evaluation result. In some tasks, it might be possible that the lack of presence is actually improving the performance. Schloerb’s subjective evaluation is based on the feeling of how good the presence is from the operator’s point of view. It seems that if there is clear mismatch between the objective and subjective evaluation, the evaluated task makes the difference. Despite its slight limitations, Schloerb’s evaluation uses the best criteria for measuring the quality of the telepresence. Drascic [Drascic, 1991] compares the performances of mono- and stereovision in a simple robot teleoperation task. In the experiments, the camera is stable but stereovision clearly provides a higher enhancement level (feeling of presence). 6.6 Problems of Tele (virtual) presence If technical problems like delay, lack of bandwidth, etc. are not considered the biggest problem in telepresence based and virtual aided teleoperation, then simulator sickness (SS) is. Simulator sickness is very similar to motion sickness and the symptoms also resemble those of motion sickness like: apathy, general discomfort, headache, stomach awareness, nausea, etc. The difference is that SS can occur without any actual motion of the operator. SS problems are encountered especially when HMD type displays are used. 30
  31. 31. Individual Simulator Task age binocular viewing altitude above terrain concentration level calibration degree of control ethnicity color duration experience with real- contrast global visual flow world task experience with field of view head movements simulator (adaptation) flicker fusion flicker Luminance level frequency threshold gender inter-pupillary unusual maneuvers distance illness and personal motion platform method of movement characteristics mental rotation ability phosphor lag rate of linear or rotational acceleration perceptual style position-tracking self-movement speed error postural stability refresh rate sitting vs. standing scene content vection time lag (transport type of application delay) update rate (frame rate) viewing region Table 1: Potential Factors Associated with Simulator Sickness in Virtual Environments [adapted from Kolasinski, et al., 1995] Kolasinski [Kolasinski, et al., 1995] has researched the SS especially in simulator (virtual) environments but the results can be transferred to telepresence environments also. The most typical reason of SS is the cue conflict. In cue conflict different nerves get different information from the environment. Typical case, which will occur in teleoperation also, is the conflict between visual and vestibular inputs. Other possible reasons can be the quality of displays especially when HMD is used and the time lags in vision and control. Potential factors associated with SS are shown in Table 1. 31
  32. 32. 7. Semi-autonomous Work vehicles The automation level of work vehicles is still far removed from the level of factory automation. The machines are still designed around the driver, and any automation is only to improve performance, while the idea of replacing the driver has not been taken seriously by the industry. 7.1 Motivation As described before, the main motivation in semiautonomous work vehicles is money. The full autonomy is still more or less a pipe dream in the case of most machines. Increasing the automation level is sensible as long as it boosts the performance – either time or quality – of the work task. When the part of the task that the driver can perform as well as the computer is automated, then there will be spare time for driver. In the factory, this spare time could be used for extra work, but in the case of a work vehicle, the spare time is more or less useless because drivers do not usually have any other tasks to do. This problem can be solved by using teleoperation, which allows the driver to use his spare time to control another machine, or to do something else. If a work vehicle can work autonomously more than 50% of its work cycle, one operator can control two or more machines and save the man power costs. From the point of view of economics, the productivity of the automation investment can be simplified to the calculation of costs (automation and teleoperation investments) and savings (decreased man power costs and savings in the driver infrastructure). The fastest development in work-vehicle automation has been in the field of mining. Also the first industrial experiments and products in the semiautonomous machines have been done in the mining industry. 7.2 Related Work In the LKAB’s mine Kirunavaara, Sweden [Erikson and Kitok, 1991] made the first real environment experiments with a semiautonomous LHD machine. The system approximated to the one presented in Paper V. The teleoperation was based on radio control and video feedback from static cameras to monitor, while the automatic driving was based on an underground signal cable and inductive sensor coils in the vehicle. The loading was performed by teleoperation from a control room and the hauling, dumping and driving back to the loading place were driven autonomously by the vehicle. The reaction of the test operators, who were all experienced LHD drivers, was mainly positive due to the improvements in the working environment. The loading and driving of an LHD machine by teleoperation didn’t cause any great problems for the inexperienced operators. On the other hand, they noticed that it takes a considerable amount of time for an automatic LHD system to manage to reach a satisfactory functionality level. The production average speed was 80-90% of the manually driven LHD in the same mine. There was no mention of the operation of several machines by one operator. 32
  33. 33. 8.Teleoperation experiments 8.1 Introduction A full telepresence system with all trimmings is technically demanding. The system easily becomes complex and expensive. In the case of heavy working vehicles, the price is important, and so is the robustness of the system. Depending on the work and the environment in which it is done, a simplified system is often sufficient, and in some cases is even better than a more complex one. It is not very clear, however, what the main factors affecting this are. In what follows, we try to illustrate some related problems through a number of field tests. The first part of the tests is done with a teleoperated test vehicle by simulating different possible tasks in an unstructured environment. In the second part, the same experiments are conducted in a structured environment with real vehicles doing real tasks. 8.2 Experiments in an unstructured environment (Papers IV and VI) 8.2.1 Test Equipment The study was started with general teleoperation experiments involving different tasks where heavy working vehicles might be used. These tasks can be found from forestry, earth-moving, construction etc. As the use of real machines was not possible, the tasks were performed using a Honda all-terrain vehicle(ATV) called Arska [Koskinen et al., 1993]. The teleoperation of the ATV, shown in Figure 23, was implemented with a steering wheel and pedal combination that provided the feeling of driving a normal car. The operator station is shown in Figure 3. Communication between the control station and ATV was made with one pair of half duplex radio modems. The telepresence equipment included a stereo-HMD, a head- tracker, two monitors, two cameras, a laser pointer, a 2 DOF servo head (Figure 4), two pairs of short range video links, a half duplex radio, and a pair of radio modems for transmitting and receiving the head tracking data from tracker to servo head. Figure 23: The test vehicle “Arska” 33
  34. 34. To study the effect of equipment level, the telepresence hardware was configured to five different systems respectively representing different enhancement levels in vision and camera control: 1. Full telepresence (SYSTEM A) stereovision, sound, 2DOF head tracking 2. Monovision telepresence (SYSTEM B) monovision, sound, 2DOF head tracking 3. Monitor based telepresence (SYSTEM C) image on screen, sound, 2DOF head tracking 4. Manual telepresence (SYSTEM D) image on screen, sound, manual 2DOF camera control 5. Standard teleoperation (SYSTEM E) image on screen, sound, fixed camera Sound was left in each alternative because leaving it out was observed to deteriorate the system radically. 8.2.2 Experiments Driving experiments were made in the university test field that consists of an uneven ground surface mostly covered by hard sand, stones and vegetation. The experiments included tasks that simulated real world driving tasks involving material handling and transportation. The test tasks were the following: 1. “Corridor” driving Driving on a defined route similar to a road, like ore transport in mine tunnels, etc. The test was conducted by driving along a winding path that included narrow gates. Overriding the path border and collisions with obstacles were counted as errors (Figure 24). Figure 24: “Corridor” driving 34
  35. 35. 2. Unknown terrain driving Driving over an unknown area where there are a lot of obstacles and no specific route, typical of forestry machines. The test was carried out by driving to an unknown forest area were the operator had to follow a natural path after perceiving it. 3. Loading tasks The vehicle must take a load into its manipulator and move during the loading, typical of different kinds of loaders. The test was carried out by pushing boxes from one line to another by the aid of the beak assembled on the front of the vehicle. 4. Maneuvering tasks Maneuvering in close places, typical of forestry and loading machines. The test was done on a “slalom” track, where the driver must dodge piles, stop on a line and park in a given slot. 5. Fast driving Driving with velocities of more than 5 m/s, typical of transporting tasks. The tests were carried out by driving fast and stopping the vehicle in a given position on an open field. 6. Off-road driving Driving in areas where both obstacles and surface conditions can stop the vehicle. The test was carried out by driving over uneven ground and crossing obstacles, like ditches and stones. Five different people aged 26 - 40, all men, were used as test operators. Two of them were classified as experienced and the rest just amateurs. The following properties were evaluated from each system: • Ease of driving with continuous motion • Ease of driving accurately • Ease of navigation • Perception of obstacles and unexpected objects in the environment • Possible ergonomic drawbacks The results of each test were evaluated by measuring overall execution time and the number of errors during the test. Only the results obtained by the experienced operators were counted. Verbal assessment concerning the properties of the system was obtained from all operators. When repeating a particular test several times the effect of learning can be clearly seen. This affect is independent of the system used. In order to eliminate the effect, only the best results obtained after a training period were taken into account. Another point was the quality of vision in different system configurations. Due to some problems of cross hearing the video channels, the quality of stereovision was not so high as it could have been. This effect influenced both execution times and verbal assessments. 35
  36. 36. 8.2.3 Results As pointed out in Chapter 5.3, the effect of learning is very noticeable in tasks that are repeatable in nature. This is illustrated in Table 2, where the first three execution times of Test 1 (Corridor driving) by one of the experienced operators are given. The system configuration is C, which was used when driving the test the very first times. Effect of learning 0:03:36 0:02:53 0:02:10 0:01:26 0:00:43 0:00:00 1 2 3 Table 2: Execution times (min.) in Test 1 when an operator drives the test the first three times. As to the different systems, the operators reported soon after starting the experiments that System D with manual camera control was much more difficult to use than the others and ergonomically bad, so it was eliminated from further evaluation. A comparison of the rest of the systems is given in the Table 3. Here the fastest execution times are presented for Tasks 1, 3 and 4. It can be seen that the times are somewhat shorter when using the helmet, but also the simplest system - system E - has good values. The different systems can be also evaluated according to error sensitivity. Table 4 illustrates error sensitivity, which has been given for each system as the relative amount of errors calculated from the total amount of errors registered during the whole test period. From this data, it can be clearly seen that systems A, B and C with head tracked camera are less sensitive to errors than system E with fixed camera. The conclusion that the stereovision helps somewhat for improving the error sensitivity may be also drawn. Operators could accomplish Test 2 (unknown terrain driving) only when using systems A, B and C. The biggest problem in this test was loosing the path so badly that it was not possible to find it again. 36
  37. 37. 03:36 02:53 A B 02:10 C D 01:26 E 00:43 REF 00:00 Corridor Maneuver Loading Table 3: The shortest execution times of Tests 1, 3 and 4 with different system configurations. REF is the reference execution, which was driven manually. (Note: system D was used only in corridor driving) Error expectancy 3.00 2.50 2.00 A NoE 1.50 B 1.00 C 0.50 E 0.00 Corridor Maneuver Loading Table 4: Error sensitivity of different systems. Note: relative number of errors in three test cases. The turnable camera was necessary to overcome this problem; system E was not suitable for this purpose. It has to be stressed that the number of drives was too limited for proper statistical analysis. More information was derived from the subjective evaluation, which was made by interviewing the drivers. The following points were made during the evaluation: 1. It was difficult to say that the stereovision clearly helped in tasks (3,4), which required accurate driving, but indications in this direction exist. The situation could be operator dependent. It should also be noticed that the HMD was relatively old and that the video links were “different pairs”, i.e. the quality of stereo pictures was not equal to mono and monitor pictures. 2. Mono-vision was considered better than stereovision because of the better quality of picture. In cases of the same quality of pictures, previous tests [Drascic, 1991] showed that stereo was faster when performing difficult tasks 37
  38. 38. for the first time. After training, the difference between mono and stereo decreased. Both mono- and stereovision caused simulator sickness in some drivers. 3. The use of monitor + head tracking was considered clumsy, especially if the whole workspace of the camera had to be used. However, the image was better than when using the HMD, and simulator sickness was not observed at all. 4. The manual use of the camera while driving was not easy, and ergonomically it was not feasible. 5. The fixed camera + monitor combination was judged as the best alternative when the task and environment was familiar. To summarize from the trials in an unstructured environment: 1. When operating in an unknown environment for the first time, the use of a head-tracker and possibly servo cameras is justifiable. 2. HMD-displays can cause simulator sickness. 3. In most cases, after learning, the fixed camera + monitor is enough. 8.3 Experiments in a structured environment (Papers V and VI) Previous tests produced generic information for teleoperation of basic work tasks, but also left many questions that can only be answered in real tests with a real machine. Tests with the same type of equipment were carried out with a full scale LHD-machine. The difference was that this time the environment was an underground construction site that formed clearly structured surroundings. 8.3.1 Test vehicle and equipment The loader used weighed 40 tons and had a loading capacity of about 5 m3. The vehicle was diesel powered with hydrostatic power transmission. The steering was an articulated type, i.e. with a frame divided into two parts of approximately the same size, which were connected by the steering joint. The bucket had two degrees of freedom: lift and tilt. The loader was equipped for full teleoperation. All the required driving actions could be performed remotely from a remote control station. The actual user interface in the remote control station was the control chair with a steering wheel and pedals (Figure 25). The steering could be controlled either with the steering wheel or with the joystick on the left handle of the chair. The throttle and brake were respectively controlled with two pedals. Engine start, gears, camera, etc. were controlled with the buttons, switches and the other joystick on the right chair handle. The feedback data was shown on the monitor of the control PC. The video image from the vehicle was shown on a separate monitor. Also, sound from the vehicle was available. The control PC read the data from control devices (pedals, joysticks, buttons, etc.) and sent it to the vehicle. 38
  39. 39. Figure 25: Control chair, steering wheel and pedals The control data was transmitted with a pair of radio modems between the vehicle and the control station. A leaky feeder cable system was used in order to cover the whole test route of the construction site. Two video channels and one sound channel were required for the tests. The leaky feeder system used did not support video frequencies so two analog video links with 2,4GHz frequency were used. To ensure the connection throughout the whole test route, three pairs of video receivers were located around the route. The right receiver was chosen manually during a run. 8.3.2 Tele-existence equipment The telepresence equipment was basically the same as that used in previous tests [Halme et al., 97]. Head Mounted display (HMD) The old HMD used in previous tests was replaced with a new one. It had two 1,35” active-matrix TFT-LCDs with VGA (640 x 480) resolution. The quality of the image was evaluated as good, especially in terms of resolution, colors and sharpness. The only negative evaluation related to a clear distortion of the image, which was probably caused by the HMD optics. Head-tracker The position of the operator’s head was tracked with the same 6 DOF mechanical head-tracker, which provided fast and accurate position data, but the mechanical connection to the operator limited his mobility and disturbed his concentration. Cameras The cameras were compact integrated lens video cameras with auto focus and 12 x optical zoom. The field of view was 47º horizontal (wide) and 4º horizontal (tele). In the tests, the wide mode was used all the time. Cameras were evaluated as very good. Servo head The servo head was also a new one. It was designed in the automation laboratory especially for teleoperation applications. The small sized robust head had a size of 300 x 300 x 300 mm. It provided turning angles of 180º in the vertical direction, and a full 360º in the horizontal. The positions of the servo head could be sent through an RS-232 or a CAN-bus. The performance of the head was equal to normal head movements. 39
  40. 40. 8.3.3 Test persons A (age 40) was a professional loader driver. He had 20 years' experience operating loaders and was also well experienced in loader teleoperation. He had driven several thousands of buckets with a system that was similar to the test system with the fixed camera configuration. B (age 43) was also a professional loader driver with several years of experience. He also was well experienced with other types of work vehicles. He did not have any teleoperation experience before these tests. C (age 35) was a professional loader driver. He had no teleoperation experience. D (age 30) was a research engineer in the automation laboratory (HUT). He had no experience of loaders. He had teleoperated a laboratory test bed 'Arska' with all the same telepresence configurations that were used in these tests. He also had a large experience of video and computer games. He was an amateur pilot. 8.3.4 Tele-existence configuration The telepresence system was configured to four different enhancement levels. These levels were basically the same as in the previous experiments, except that the manual control was omitted: 1. Fixed cameras and monitor The image came from the fixed cameras pointing straightforward and backward. Cameras were located on the centerline of the vehicle. The front camera was fixed to the front part and the back camera to the back part of the vehicle. The image was shown on a monitor ahead of the driver. The image was automatically switched between the front and the back camera depending on the gear position. There was also a possibility of taking a look in the opposite direction by pressing a button on the handle of the operating chair. 2. Servo camera and monitor The image was coming from the right camera of the servo head, which was located on the top of the cabin about 1m from the vehicle centerline. The operator controlled the camera movements with the ADL head-tracker. The image was shown in a monitor like in the previous configuration. The main pointing direction of the servo head could be rotated 180° by a button. 3. Servo camera and mono HMD Like the previous configuration but with the monitor replaced by the HMD. The image from the right camera is divided between the two displays in the HMD. 4. Servo camera and stereo HMD Like the previous configuration, but both cameras were used. The images from the cameras were transferred to the two displays of the HMD to create a stereo image. The sound feedback was considered to be so important that it was included in all the configurations. 40
  41. 41. 8.3.5 Test runs The test runs were divided into two parts. In the first part, drivers drove three runs with all three steering configurations by using the first telepresence configuration (fixed camera + monitor). The steering tests are not reported here. After that, they chose the steering configuration they liked most. This was criticized because all the drivers then learned the driving with the same telepresence configuration. However, there was no time for driving with all the configurations, and test runs with this defect were approved. In the second part, they drove the remaining test runs with three telepresence configurations while using the chosen steering configuration. The problems in the leaky feeder communication forced the first two drivers A and B to drive the shorter route. Test drivers C and D drove the full route. In both cases, the communication was not working properly all the time. The vehicle control system generates an emergency stop if two consecutive data packets are missed. This caused several stops during test runs. The runs were continued after the stops but this took a lot of time because the engine of the vehicle had to be started again after an e-stop. 8.3.6 Evaluation methods Schloerb [Schloerb, 1995] presents an evaluation method for telepresence systems where the objective evaluation is to conclude how well the defined task is performed, while the subjective evaluation is based on the feeling of how good the presence is from the operator’s point of view. In our case, the evaluation was also divided to two parts: objective and subjective evaluation. The objective evaluation was quite near to Schloelrb’s definition; the subjective evaluation was based on subjective comments of the drivers. These comments concerned both the “goodness of the telepresence” and the performance in the tasks. In previous experiments, it was noticed that with the resources available it was difficult to drive enough to get a sufficient quantity of data for proper statistical analysis. In the objective evaluation, the performance was evaluated on the basis of performance data. In the subjective evaluation, the comments of the drivers during driving were recorded. Each driver was interviewed after each run, and when all the runs had been completed. In the objective evaluation, the following data was collected in order to evaluate the performance of the driver: 1.Time of the run: The time of each run was measured. The e-stop interrupts were subtracted. The problem here was that the number of the runs did not allow a proper statistical analysis. 2. Errors: All the errors, which were mostly hits to the walls or instances of emergency braking by the safetyman, were calculated 3. Logging the operator driving data: The steering and the throttle movements made by the operator were logged with 100Hz frequency. The data that is shown for the time phase, and, especially, for the frequency phase, shows very clearly how nervous the driving was. 4. Logging the vehicle positional data: The test vehicle also had navigation equipment. The 2D position of the vehicle was calculated 41
  42. 42. from optical gyro and velocity, speed, motor speed, angle of the middle joint, time and angular speed were measured. 8.4 Results After the first (steering) tests, the fixed cameras were changed to the servo camera(s). All the drivers tested all the three configurations with the chosen steering system. The result of those tests was a surprise. Unlike in the previous tests, the use of the servo cameras didn’t help the driver at all. In fact, even with the stereovision, the results were clearly worse than with the fixed cameras. The number of collisions with the walls and the driving times increased. In [Halme et al., 97] and [Drascic, 91] it was shown that after the task was learned the performance difference between the different enhancement degrees disappeared. Now, despite learning the performance, in all cases involving the servo camera(s), performance was clearly worse. In the evaluation, the logged operator data was noticed to be very informative. Steering and throttle data showed clearly how fast and smoothly the driver could drive. Figure 26 shows the operator data from the different drivers. It can be clearly seen how drivers A and C drove faster with a smaller number of steering movements (less nervous) than drivers B and D. 3 3 2.5 2.5 2 2 1.5 1.5 1 1 0.5 0.5 0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160 5 5 4 4 3 3 2 2 1 1 0 0 0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160 Throttle (up) and steering data of drivers A and B 3 3 2.5 2.5 2 2 1.5 1.5 1 1 0.5 0.5 0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160 5 5 4 4 3 3 2 2 1 1 0 0 0 20 40 60 80 100 120 140 160 0 20 40 60 80 100 120 140 160 Throttle (up) and steering data of drivers C and D Figure 26: Raw throttle and steering data for all drivers (sticks, fixed cameras) 8.4.1 Servo camera and monitor It was found unergonomic to turn the cameras by means of the head and still look at the fixed monitor. The driving times increased, and then started to approach the times of driving with fixed cameras. The times decreased when drivers start to keep their heads (and cameras) fixed. 42
  43. 43. Figure 27: Tele-existence aided driving. Driver (front) is using a servo camera and normal monitor. Virtual passenger (back) is looking at the stereo image from servo cameras with HMD. The greatest change compared to the fixed camera configuration was the different camera place. The fixed front camera was on the middle line on the front part of the vehicle. The fixed back camera was on the back part. The servo cameras were located in the left edge of the cabin (back part). This helped the loading, because the bucket could be seen a little better when it was in the lowest position. When cameras were located to the side of the centerline, it was difficult to center the vehicle in the tunnel. In the servo camera operation also, “information increased the pain”. When the driver had the chance to see how near the corner was in a turning, he started to correct the situation, even though the vehicle would have managed without his doing so. 8.4.2 HMD with mono vision Again errors and times increased from the fixed camera runs. In Figures 28 and 29, the data from driver C with all four telepresence configurations is shown. It can be seen how especially the throttle is used much more careful in HMD + mono configuration. During driving, the drivers started again to keep the cameras as fixed as possible. The times were still worse than with the fixed cameras. Driver C got nausea and the others complained that they could not drive for long periods of time with the HMD. It seems that an HMD - even a good quality one – is not suitable for use over a long period of time. 8.4.3 HMD with stereo Again errors and times increased from those of the fixed camera runs. Figures 28 and 29 show that the data for the runs with the HMD + stereo is slightly better than that for the runs with the HMD + mono, but clearly worse than for those with the fixed cameras. However, at this point, the drivers also made positive comments about the servo vision, even though it strained their eyes. It created a real telepresence and helped drivers in their distance estimation. This especially helped in the loading. 43
  44. 44. 3 2 1 0 100 200 300 400 500 600 700 3 2 1 0 100 200 300 400 500 600 700 3 2 1 0 100 200 300 400 500 600 700 3 2 1 0 100 200 300 400 500 600 700 Figure 28: Throttle data; diver C by using sticks. Fixed camera (top), servo camera + monitor, HMD + mono, HMD + 5 stereo. 0 0 100 200 300 400 500 600 700 5 0 0 100 200 300 400 500 600 700 5 0 0 100 200 300 400 500 600 700 5 0 0 100 200 300 400 500 600 700 Figure 29: Steering data; driver C, steering by using sticks. Fixed camera (top), servo camera + monitor, HMD + mono, HMD + stereo. An HMD always provides a feeling of existence. With the real stereo, the feeling is very strong. In a monitor the movements of the vehicle are movements of the image. With an HMD, the operator feels himself moving while he is sitting in a stationary chair. This conflict of senses can cause simulator sickness (see Ch.5.4). The stereovision also causes a problem. When the mutual angle of cameras is fixed, the operator has to correct the distortion with his eyes. This gives rise to eyestrain after a while. 44
  45. 45. 9. Visual flow in teleoperation The “easy” driving in a tunnel with fixed camera and normal monitor raised further questions as to how the walls do actually affect teleoperation. In [Halme et al., 1997], it was noticed that on a normal route, where route markings are on the ground, driving is easy, until tight maneuvers are needed. In the bends of the route, the route disappears from the view of the camera if there is no possibility of turning the camera (this depends on the camera’s field of view (FoV)). Especially on an unfamiliar route, this renders driving remarkably difficult. In a tunnel, this situation doesn’t arise because the walls are always visible in the bends in spite of the camera FoV. But are there any other effects in tunnel teleoperation? Srnivasan [Srnivasan et al., 2000] demonstrates how honeybees flying in a tunnel center themselves by sensing the frequency differences between the textures of the walls (optical flow). Although there is a great difference between the visual perception of bees and humans, it was thought interesting to see whether the same phenomena have a role in human driving. This interesting test was imitated by replacing the honeybees with a RC-car teleoperated by a human driver. As mentioned above, the eye of a human is totally different from that of a honeybee. Most probably the cognition of the two is also different. Nevertheless, teleoperation in a tunnel remains comparable to the experimental tasks of Srnivasan. 9.1 Test Setup The first idea was to use conveyor belt type walls where a “driver” could adjust the camera between the walls. However, this was considered too artificial, and a full teleoperation environment was created on a RC-car (Figure 30). The “normal” RC-car controller with two sticks was considered too complicated for people who are used to driving cars only; therefore the sticks were replaced with a car type (and size) steering wheel and gas pedal. The test tunnel was made from corrugated cardboard coated with white paper with black stripes (Figure 8). The frequency of the stripes was different on each side of the tunnel. The length of the tunnel was 20 m. The test drives were conducted in tunnels 50 and 70 cm wide respectively. (The width of the car was 20cm). 45
  46. 46. Figure 30: Small scale RC-car equipped with camera, videolink and headlight 9.2 Test drives The test drives were carried out by 7 persons, each driving the car forth and back in 8 tunnel configurations. The parameters of test drives are set out in Table 5. Tests were started in normal “daylight” conditions (Figure 31). However, it seemed that drivers could see a lot of possible landmarks other than stripes. To remedy this, the “stripe effect” was strengthened by adding night (or real tunnel) driving to the test program. In the vehicle’s headlight, only striped walls could be seen. 46