1.2 Literature Overview
Primarily driven by applications in vehicle control and aircraft piloting, the first wave of human factors research on input controllers started in the 1940's and reached its apex in 1950's and early 1960's. Many issues discussed in the previous section, even though in the context of 1 or 2 rather than 6 DOF devices, were studied in this period of time. Orlansky provided one of the earliest comprehensive analyses of the human factors issues to be concerned with for the design of input controllers. His article (Orlansky, 1949) analysed factors such as maximum forces that may be exerted by a human pilot, the gradient of control forces and the manner of human movement. However those analyses were not supplemented with empirical experiments.
Researchers from the Applied Psychology Unit of the Medical Research Council in Cambridge, England, including K.J.W. Craik (Craik, 1943, 1944, posthumously published as Craik and Vince, 1963a, 1963b, after Craik's death in 1945) , C.B. Gibbs (1954; 1962) , E.C. Poulton (1974) and others, took leading roles in the early research on controls. These researchers were concerned with human performance affected by various type of controls. Gibbs, for example, hypothesised that isometric devices (force sticks) provide strong "proprioceptive discharge" in the human limb and therefore help the human operator's performance. Poulton, on the other hand, took a position opposite to that of Gibbs.
Another notable group of researchers, the "Ohio School", including P.M. Fitts (1951) , H.P. Bahrick (Bahrick, Fitts, and Schneider, 1955b) , D. Howland and M.E. Noble (1953) (primarily from the Ohio State University) made the most impressive theoretical contributions to the understanding of controls. Their central thesis was that human proprioception can be modelled by laws of physics. According to their theory, elastic loading on a control device augments the perception of displacement, due to the fact that the resistance force of a spring is proportional to displacement (Hooke's law). When a control device has viscous resistance, the human perception of velocity will be enhanced, due to the fact that viscous resistance is linearly related to velocity. Similarly, as revealed by Newton's second law, inertial resistance is proportional to acceleration, therefore the mass of a control device should augment the human perception of acceleration. This physics based model of proprioception was supported by a series of analyses and experiments (Fitts, 1951; Howland and Noble, 1953; Bahrick, Bennett, and Fitts, 1955a; 1955b; Bahrick, 1957). Notterman and Weitzman (1981) later confirmed this proposition in a more systematic manner.
The early research on controls was often concerned with dynamics. Aircraft, submarines and other vehicles all have complex dynamics. Birmingham and Taylor (1954, cited in Notterman and Page, 1962) hypothesised that human tracking performance would remain unchanged, despite variations in control device properties, if the overall transfer function relating the force applied to a control device to the system output remains unchanged. Notterman and Page (1962) conducted an experiment, however, that rejected Birmingham and Taylor's hypothesis. They studied systems that have the same overall transfer function (second order dynamics) but differ in where the dynamics was located within the control loop. In one system, second order dynamics was embodied in the input device's mechanical properties (elasticity, viscous damping, and inertia). In the other two systems, the input devices had negligible dynamics but the same second order dynamics was simulated in an analogue computer between the input device and the display. Notterman and Page demonstrated that the human operator had better performance with the first system, although mathematically the total system transfer function was comparable with the other two systems. They argued that the "local" (proprioceptive) feedback in the first system helped the subjects, since they could not only see the dynamic response from the visual display but could also "feel" the dynamics from the physical device.
Because of the importance of dynamics in early engineering systems, how humans handled the plant dynamics became more of a central theme in manual control research than the properties of input device themselves. Engineering models (particularly classical and modern control theories) were applied to describe and predict human behaviour in such a context. Sheridan and Ferrell (1974) provided a comprehensive summary of such efforts. In more modern control systems, however, automation of machines has reduced concern for the dynamics aspect of manual control. Much of the low level dynamics can now be handled by automatic controllers and the human's role has been increasingly elevated to supervisory tasks (Sheridan, 1988, 1992b) . Today's design of controls is therefore concerned mostly with facilitating human information input (or spatial instructions) into computer systems.
Research on input control has a strong two-way connection with the study of human motor skills. On the one hand, knowledge from human motor control research can clearly be applied to the design of control interfaces. On the other hand, many researchers have used different input control devices and manual control paradigms as vehicles for studying human motor control behaviour. The above mentioned research by Gibbs and by Fitts and colleagues all aimed at enhancing the understanding of human motor behaviour. Based on tracking research, Krendel's and McRuer's successive organisation of perception (SOP) theory hypothesised the general trend of human skill shift from closed loop to open loop behaviour (Krendel and McRuer, 1960) . Also based on tracking research, Pew (1966) proposed the hierarchical organisation of human motor control.
Interest in research on the properties of controls decreased in the mid-1960's, however. A. A. Burrows (1965) made a plea to continue studies on "control feel" and its related variables. He argued that "one would expect the relationship of the hand to the controlled element, being at the one time both an input and output, to be a fruitful area for research", but the reality is that little was well understood. He pointed out that the reluctance to conduct research in this area is understandable in view of the immensity of the possible interactions among the many dimensions of control feel.
In 1974, E.C. Poulton published his comprehensive review book on human tracking skills and manual control. The book (Poulton, 1974) covers much of the early research on design of controls. It was written in a very empiricist style, placing heavy emphasis on experimental data rather than theoretical issues and models. This was criticised at the time by other researchers (e.g. Pew, 1976) . In retrospect, Poulton's inclination towards empirical results was not necessarily unwise. Models and theories in research often change with the varying cultures in the scientific community but empirical data remain valuable. Taking human motor control as an example, cybernetic models were widely applied in early research, as evident in (Brooks, 1981) which surveys motor control research in the 1960's and 1970's, but decreased dramatically in later journal publications. Instead, artificial neural network models are currently on the rise.
Another important feature of Poulton's book is his critical discussion of "asymmetrical skill transfers" likely caused by within-subjects designs of experiments in the research literature. In within-subjects experiments, the same group of subjects is assigned to all experimental conditions; that is, each and every subject performs all experimental conditions. In between-subjects experiments, on the other hand, the subjects are divided into subgroups. Each subgroup of subjects perform in only one experimental condition. A within-subjects design needs fewer subjects than a between-subjects design and is therefore more commonly used. Apparently, in within-subjects experiments, subjects may carry over some effects, such as skills or fatigue, from earlier conditions to later conditions. In order to overcome this possible transfer effect, the sequencing of the experimental conditions in within-subjects designs is usually ìbalancedî by assigning subjects to the conditions in such a way that all experimental conditions have an equal number of times of being first, second, etc., or last condition. Poulton argues that although such an arrangement may balance the sequence of the conditions, it does not guarantee that the actual skill transfer from one condition to another is ìsymmetricalî. When transfer is asymmetrical, biased results can be produced. Poulton claimed that "once the biased results (due to asymmetrical skill transfer) are discarded, there emerges a clear and sensible description which differs in many respects from current views and practices". Asymmetrical skill transfer could indeed be a problem, but whether its effect is as important as Poulton believed is debatable. His repeated warnings (Poulton, 1966, 1969, 1973, 1989) have not been widely accepted by psychologists and human factors researchers, as within subject designs continue to be used frequently in experimental research.
Since the late 1970's, another wave of studies on input controls have been carried out as part of the research on human computer interaction (HCI). Card, English, and Burr (1978) conducted one of the most well known studies on the performance differences between various computer input devices (mouse, trackball, joystick, stylus, etc.). Card and colleagues also established the Fitts' law paradigm as the de facto standard task for computer input device research, even though Fitts' law is only one of the many theoretical products of decades of human motor control studies.
The often bewildering diversity of devices that can be possibly used for computer input has interested many HCI researchers. Significant effort has gone into building taxonomies for classifying the devices according to user behaviour. Buxton's taxonomy (1983) is among the best known taxonomies of computer input devices. It has been recently expanded by Card, Mackinlay, and Robertson (1990) . Bleser (1991) and Lipscomb and Pique (1993) are two other examples of taxonomy research on input devices.
Viewing human computer interaction as a human dialogue process with computers, Buxton introduced many concepts such as chunking and phrasing into research on computer input (Buxton, 1986, 1990, in press). Recent interest in computer input device research also includes utilising two hands in a co-operative manner, based on models from human motor control theory (Guiard, 1987; Kabbash, Buxton, and Sellen, 1994) .
As computer interfaces went beyond the limitations of flat 2D screens, input devices with multiple DOF caught the interest of HCI researchers. One question that needs to be addressed is what tasks are suitable for multiple DOF device applications. Extending the concepts of perceptual integrality and separability (Garner, 1974) , Jacob and colleagues suggested that the perceptual structure of the task should be considered in determining whether multiple DOF devices should be used (Jacob and Sibert, 1992; Jacob, Sibert, McFarlane, and Mullen, 1994). According to their view, when the multiple dimensions of a task are perceptually integrated, such as translations along each dimension of a 3D space, integrated multiple DOF devices are an advantage. When the multiple dimensions of a task are perceptually separated, on the other hand, such as adjusting the colour as well as the location of an object, it is easier to manipulate the separated variables one at a time with lower DOF devices. Similar thoughts are also reflected in other studies (see Fracker and Wickens, 1989; Wickens, 1992 for reviews). It has been found that when the dynamics of each DOF are alike, integrated control is superior to separated control. In contrast, when the two dimensions have different dynamics, for example when one is zero order and the other is second order, separated control works better than integrated control.
The newly arrived discipline of virtual environments (VE), and
its older sister discipline , teleoperation, are naturally concerned
with the performance of various types of multiple DOF control
devices. A variety of control devices for teleoperation are reviewed
by Brooks and Bejczy (1985) and by Jacobus, Riggs, Jacobus, and
Weinstein (1992) . Similar to the contrast between isomorphism
and tool using discussed earlier, there have been generally two
streams of designs for teleoperation and virtual environments.
One is the master-slave structure, in which the remote slave robot
and the master controller with the human operator are geometrically
isomorphic. The other is the tool-like hand controller approach.
The merit of each of the approaches has been debated in the virtual
environment community (Green, Bryson, Poston, and Wexelblat, 1994) .
Sheridan's essay (1992a) defines the research questions on teleoperation
and VR input control very well:
The research question here is: how do the geometric mappings
of body and environmental objects, both within the perceived (virtual)
environment and the true one, and relative to each other, contribute
to a sense of presence, training, and performance? Control by
the human operator, which requires some such mappings, may be
easy, while others may be difficult. Some own-body to teleoperator/remote-environment
control tasks or own-body to virtual-operator/virtual environment
tasks may demand a high degree of isomorphism. In some cases there
may be a need to deviate significantly from strict geometric isomorphism
because of hardware limits, or constraints of the human body.
At present we do not have design/operating principles for knowing
what mapping or remapping from the lower set of vectors to the
upper, or back again for feedback, is permissible, and which degrades
performance.
In summary, a large body of literature is concerned with human factors in input control design. However, many issues listed in the beginning of this chapter have still not been agreed upon in the literature. There are multiple reasons for this dilemma. First, research is usually based upon available technologies and driven by applications relevant at the time of the research. One such limitation in the literature has been the number of degrees of freedom of the input devices. This thesis deals with 6 DOF manipulation, whereas the existing literature is concerned mostly with 1 or 2 DOF devices. A 6 DOF device can not simply be regarded as a summation of multiple lower DOF devices, due to the co-ordination required between the multiple degrees of freedom. For example, overall operating speed with 6 DOF devices may conceivably be slower than with low DOF devices. Hence, the bandwidth of a 6 DOF device may become less important while properties that facilitate co-ordination, the ability to simultaneously control all degrees of freedom, may become more important.
Secondly, the literature itself is not conclusive. Although a large body of data exists, each subset of these data may depend upon the particular experimental devices and methodology used by the researchers. The generalisability of experimental conclusions is a common open issue in human factors research.
Thirdly, the related literature is very scattered and comprehensive reviews are few and far between. It is therefore difficult even for researchers in the field to grasp a major part (not to mention the entirety) of the literature. For instance, many very important works, such as those of the Ohio school cited earlier, were overlooked in Burrowsí critical paper on the state-of-the-art of research on control feel (Burrows, 1965). The later research on input devices in HCI, and research on controls for telerobotics and virtual environments also have made very little reference to the first wave of research on controls.
Finally, as Burrows (1965) pointed out, the subject of input
control is much more complex than it first appears and firm conclusions
are typically hard to draw. Human performance is a function of
the interactions among many dimensions involved in the making
of a control interface. For example, when taking certain mechanical
properties of the control device as the variable of interest,
subjects' performance may also change as a function of other dimensions,
such as control gain, device bandwidth, learning experience, fatigue
and so on. Furthermore, these variables may interact in complex
ways. For instance, the optimal control gain could change with
learning experience; that is, as subjects acquire more experience,
the optimal control gain may be quite different from levels adopted
during novice phase. Experimental studies addressing all possible
combinations of different amounts of experience, different control
gains, and so on therefore become too complex to conduct. However
daunting this complexity is, researchers should not give up systematic
study and thus fall behind technology.