7.1 Introduction and goals for the studies
This chapter discusses some experiments which analyze manipulation of physical objects with both hands. This includes an analysis of real-world tasks, such as handwriting and sketching, as well as a formal experimental task which involves using a tool in one hand to point at a target object held in the other hand. An underlying goal of all of these studies is to gather both formal and informal evidence that Guiard's theoretical framework applies to a wide range of tasks of interest for human-computer interaction, as well as the specific tasks supported by the props-based interface.7.2 Experiment 1: Qualitative analysis of handwriting and sketching
My handwriting and sketching analyses were suggested by Guiard's result that exclusion of the nonpreferred hand from a handwriting task reduces the writing speed of adults by some 20% [67]. I believed that a similar result could be demonstrated for drawing or sketching tasks, which of course are important for the many artists, architects, or engineers who sketch out design concepts using digital media. The general implication is that using two hands is relevant to far more than 3D manipulation. Designers of hand-held computers, sketching software, gesture or pen-based interfaces, and drafting-board style displays such as the Immersadesk [49] or the Immersive Workbench [52] should be designing with a careful consideration for the issues raised by two-handed interaction.
The handwriting and sketching studies emphasize qualitative results. Through subject observation and videotape analysis, my pilot studies quickly led to qualitative insights with immediate application to interface design, but I found it was difficult to formalize these results. There are many interacting within and between-subjects factors that must be carefully controlled to produce a clean quantitative result, but many of these factors (which are outlined in section 7.2.1) have little or no relevance to interface design. As a result I decided that discovering, understanding, and measuring or controlling for these factors in formal experimental designs was beyond the intended scope of this thesis.
The handwriting and sketching studies individually consisted of 6 rounds of experimental work with small numbers of subjects. In total, 29 right-handed subjects (9 friends and colleagues and 20 subjects drawn from the psychology department's subject pool) participated. Each round of experimental work tested one or more variants of the experimental tasks of (1) filling a page with handwriting, (2) filling out a form, (3) sketching an object in the environment, or (4) drawing a circle. Rather than focusing on the specific details of each round of tests, this discussion pools my observations across all subjects, and gives the specifics of the experimental tasks where appropriate.
7.2.1 Handwriting
Subjects were asked to make a copy of a passage from a printed page onto a separate blank sheet of paper. I found that subjects almost always place the page for copying to the left of the page for writing, and use the left hand to keep track of their current position in the text. The actual handwriting itself is done unimanually. Clearly, maintaining split attention between the pages is a more difficult task than the handwriting itself, and because of this subjects prefer to recruit the left hand for that task.


Figure 7.2 Writing posture of left-handed inverters and non-inverters [6].

Figure 7.3 Examples of circles drawn versus individual pen strokes3.
As expected, in the BF condition most subjects did naturally use their left hand to rotate the page and thereby extend the working range of the right hand. This strategy allowed subjects to repeatedly draw an almost identical stroke until the circle was complete. Subjects would perform most of the work on the left hand side of the page. To illustrate this tendency, I placed transfer paper (carbon paper) on the back of a specially prepared desk blotter; subjects sketched a circle (fig. 7.3, top row) on a page on top of this blotter. The transfer paper captured the impression of the subject's pen strokes relative to the blotter (fig. 7.3, bottom row). As the examples shown in figure 7.3 demonstrate, most of the arcs are in the same direction and take up a small area relative to the size of the circle itself. Note that figure 7.3 shows examples from two separate subjects performing slightly different tasks (subject A had 4 registration marks to guide the drawing of each circle, while subject B had 8 such registration marks).
In the BR and UR conditions where rotating the page was not possible, subjects often had difficulty in the upper right-hand quadrant of the circle. This is exactly the point where the inflection of the circle goes against the natural biomechanical curves generated by the wrist and fingers of the right hand. To compensate for this biomechanical awkwardness, subjects would try to move their entire torso relative to the page to produce a better angle for the action of the right hand. Having the left hand in the lap, as required by the UR condition, made it even more difficult for subjects to counter-rotate their body to adjust for the fixed page. As one subject commented, "I usually put my hand on the table for support." Thus, the UR condition reveals that the left hand normally plays a postural role as well.

The same issue arises in interface design when using both hands for direct input on a large display surface such as the Active Desk (fig. 7.4). In unpublished work, Buxton tried to reproduce the results of Kabbash [95] using direct input on the Active Desk display. Kabbash (as originally discussed in section 2.6.2.2 on page 36) found that users could connect and color a series of randomly appearing dots much faster with the two-handed ToolGlass technique [15] than with one-handed techniques; Kabbash produced these results using indirect input devices in conjunction with a standard monitor. But on the Active Desk, Buxton found that the targets would sometimes randomly appear in parts of the display occluded by the user's arms, causing subjects to become extremely confused. Thus, interface designs which use two hands for direct input need to be aware of the problems introduced by occlusion of the display surface and potential problems caused by the resulting loss of context4.
As a final example, during one experimental session, I asked two subjects with minors in art to sketch an object which I placed at the rear of the desk surface. Although these subjects were allowed to use both hands, neither subject rotated the page while they were capturing the general outline of the object. However, once they started to shade it and "overstrike" some of their initial strokes, they began to frequently rotate the page with the left hand. Thus, when drawing an object from the environment, rotating the page is a detriment because it would require a mental rotation to align the page and the environment; but when attention is focused primarily on the page itself, these subjects found rotating the page to be a natural and efficient task strategy.
Furthermore, as illustrated by the impressions of individual pen strokes relative to the desk (fig. 7.3), when the left hand helps to manipulate the page, the patterns of hand use are quite consistent and predictable across subjects. By dynamically positioning and orienting the page, the nonpreferred hand extends the working range of the "sweet spot" for the preferred hand.
There are problems which these studies do not address. Many uninteresting factors obscure the time, precision, or postural advantages (if any) of using two hands for these tasks, and a formal experiment would need to control for these factors. The tasks which I tested were also somewhat open-ended, and did not adequately control for time/accuracy trade-offs. This, in addition to the large between-subject variability in terms of handwriting speed, artistic ability, and so forth, means that these studies were not well suited to produce quantitative results.
Furthermore, when physical constraints guide the tool placement, this fundamentally changes the type of motor control required. The task is tremendously simplified for both hands, and reversing roles of the hands is no longer an important factor. Thus, specialization of the roles of the hands is significant only for skilled manipulation.

Figure 7.5 A subject performing the experimental task.
The present experiment was motivated by my experiences with the props-based interface. Informally, I observed that the operation of the interface was greatly simplified when both hands were involved in the task. But the early design stages had to consider many possible ways that the two hands might cooperate. An early prototype allowed users to use both hands, but was still difficult to use. The nonpreferred hand oriented the doll's head, and the preferred hand oriented the cross-sectioning plane, yet the software did not pay any attention to the relative placement between the left and the right hands. Users felt like they were trying to perform two separate tasks which were not necessarily related.
For truly bimanual movement, most psychology and motor behavior experiments have studied tasks which require concurrent but relatively independent movement of the hands. Example tasks include bimanual tapping of rhythms [44][134][185] and bimanual pointing to separate targets [87][117][184]. Since the hands are not necessarily working together to achieve a common goal, it is uncertain if these experiments apply to cooperative bimanual action.5
There are a few notable exceptions, however. Buxton and Myers [27] demonstrated that computer users naturally use two hands to perform compound tasks (positioning and scaling, navigation and selection) and that task performance is best when both hands are used. Buxton [29] has also prepared a summary of issues in two-handed input.
Kabbash [95] studied a compound drawing and selection task, and concluded that two-handed input techniques, such as ToolGlass [15], which mimic everyday "asymmetric dependent" tasks yield superior overall performance. In an asymmetric dependent task, the action of the right hand depends on that of the left hand [95][67]. This experiment did not, however, include any conditions where the action of the left hand depended on the right hand.
Guiard performed tapping experiments with a bimanually held rod [69]. Subjects performed the tapping task using two grips: a preferred grip (with one hand held at the end of the rod and the other hand near the middle) and a reversed grip (with the hands swapping positions). The preferred grip yielded better overall accuracy, but had reliably faster movement times only for the tapping condition with the largest amplitude. Guiard also observed a distinct partition of labor between the hands, with the right hand controlling the push-pull of the rod, and the left hand controlling the axis of rotation.
A number of user interfaces have provided compelling demonstrations of two handed input, but most have not attempted formal experiments. Three-dimensional virtual manipulation is a particularly promising application area. Examples previously described in chapter 2 include the Virtual Workbench [138], 3Draw [141], Worlds-in-Miniature [164], PolyShop [1], and work by Shaw [149] and Multigen, Inc. [121]. There is also some interest for teleoperation applications [163]. In two dimensions, examples include Toolglass and Magic Lenses [15], Fitzmaurice's Graspable User Interface [59] (shown in figure 7.4 on page 139), and Leganchuk's bimanual area sweeping technique [108]. Bolt [17] and Weimer [180] have investigated uses of two hands plus voice input. Hauptmann [76] showed that people naturally use speech and two-handed gestures to express spatial manipulations.
7.5 The Experiment
7.5.1 Task
The subject manipulates a tool (either a plate or stylus) in one hand and a target object (either a puck, a triangle, or a cube) with the other hand (fig. 7.6). Each target object has a rectangular slot cut into in it, at the bottom of which is a small gold-colored target area. There are two versions of the task, a Hard task and an Easy task.
7.6 Experimental hypotheses
The experimental hypotheses were suggested by my experiences with the props-based interface and formalized with the help of Guiard's Kinematic Chain (KC) model. The high-level working hypothesis for this experiment is that the KC model can be used to reason about two-handed 2D or 3D tasks and interface design. 7.6.1 Subjects
Sixteen unpaid subjects (8 males, 8 females) from the Psychology Department subject pool participated in the experiment. Subjects ranged from 18 to 21 (mean 19.1) years of age. All subjects were strongly right-handed based on the Edinburgh Handedness Inventory [127].
7.6.2 Experimental procedure and design
Figure 7.6 shows the overall experimental set-up. The experiment was conducted using instrumented physical objects, rather than virtual objects. Since the purpose of the experiment is to look at some basic aspects of bimanual motor control, using physical objects helps to ensure that the experiment is measuring the human, and not artifacts caused by the particular depth cues employed, the display frame rate, device latency, or other possible confounds associated with virtual manipulation. The physical objects also provided the haptic feedback needed to test hypothesis H4.
The experiment began with a brief demonstration of the neurosurgical props interface to engage subjects in the experiment. I suggested to each subject that he or she should "imagine yourself in the place of the surgeon" and stressed that, as in brain surgery, accurate and precise placement was more important that speed. This made the experiment more fun for the subjects, who would sometimes joke that they had "killed the patient" when they made an error.


For the experimental trials, a within-subjects latin square design was used to control for order of presentation effects. For each of the four experimental conditions, subjects performed 24 placement tasks, divided into two sets of 12 trials each. Each set included two instances of all six possible tool and target combinations, presented in random order. There was a short break between conditions.
7.6.3 Details of the experimental task and configuration
For each trial, the computer display (at the right of the working area) simultaneously revealed a pair of images on the screen, with the objects for the left and right hands always displayed on the left and right sides of the screen (fig. 7.9).
When using the plate, subjects were instructed to use the entire 0.5" wide tip of the plate to touch the target. For the stylus, the subject was told to touch the rounded part of the target area (the stylus was thicker than the other parts of the target, as shown in figure 7.7, and thus only the central rounded part of the target area could be touched without triggering an error).
7.6.4 Limitations of the experiment
There are a couple of factors which limit the sensitivity of this experiment. First, ideally the experiment would present a range of controlled difficulties analogous to the Index of Difficulty (ID) for Fitts' Law [114]. Fitts' law relates the movement time for one hand to two quantities, the amplitude A of the movement and the width W of the intended target. Together, A and W can be used to compute ID, the index of difficulty for the movement. But Fitts' Law applies to movement of one hand, and I am not aware of any adaptations which could handle movement of both hands together. Instead, the experiment uses an easy versus hard difficulty distinction.
Second, the accuracy measurements yield a dichotomous pass / fail outcome. Thus, the apparatus captures no quantitative information about the magnitude of the errors made when the subjects missed the target in the Hard conditions. Even given these limitations, the experimental results are quite decisive. Therefore, I decided to leave resolution of these issues to future work, and to demonstrate some effects with the simplest possible experimental design and apparatus.
7.7 Results
For each condition, only the second set of 12 trials was used in the data analysis, to minimize any confounds caused by initial learning or transfer effects across conditions.
Condition |
Mean |
Std. dev. |
Error rate |
|---|---|---|---|
|
Preferred Easy (PE)
|
0.76
|
0.15
|
--
|
|
Reversed Easy (RE)
|
0.83
|
0.19
|
--
|
|
Preferred Hard (PH)
|
2.33
|
0.77
|
43.9%
|
|
Reversed Hard (RH)
|
3.09
|
1.10
|
61.1%
|
Figure 7.10 Summary of mean completion times and error rates.
7.7.1 Qualitative analysis
Before proceeding with a full statistical analysis, it seems appropriate to first discuss some of the qualitative aspects of the experiment. Some of the subjects were videotaped; the qualitative observations presented here were based on these tapes as well as handwritten notes.
At first glance, it would seem that the primary difference between the RH and the PH conditions was the left hand's unsteadiness when handling the tools. For at least some of the subjects, however, it also seemed that the right hand had difficulty setting the proper orientation for the action of the left hand. So the right hand was best at fine manipulation, whereas the left hand was best at orientating the target object for the action of the other hand.
For the Easy tasks, subjects performed the task quickly and without much concentration, since they could rely on the physical constraints of the slot to guide the tool. Subjects were divided about whether or not the RE task was unnatural. Some thought it was "definitely awkward," others thought it was "fine." At least one subject preferred the Reversed grip; this preference was confirmed by a small Reversed grip advantage in the quantitative data.
Finally, when switching to the Hard task after performing a block of the Easy task, subjects often took several trials to adjust to the new task requirements. Once subjects became used to relying on physical constraints, it required a conscious effort to go back. To assist this transition, I instructed subjects to "again emphasize accuracy" and to "focus initially on slowing down."
Figure 7.11 Significance levels for Main effects and Interaction effects.


Contrast |
F statistic |
Significance |
|---|---|---|
|
PE vs. RE
|
F(1,15) = 3.94
|
p < 0.10, n.s.
|
|
PH vs. RH
|
F(1,15) = 33.56
|
p < 0.0001
|
Figure 7.14 Significance levels for comparisons of experimental conditions.

7.7.3 Possibility of Order or Sex biases
I repeated the ANOVA with between-subject factors of Sex and Order of presentation to ensure that the above experimental results were not biased by these factors. The Order of the experimental conditions did not approach statistical significance, nor did the Order 5 Condition interaction, indicating that the results are not biased by transfer or asymmetrical transfer effects.
Factor |
F statistic |
Significance |
|---|---|---|
|
Sex
|
F(1,14) = 5.55
|
p < .05
|
|
Tool 5 Sex
|
F(1,14) = 12.80
|
p < .005
|
|
Task 5 Sex
|
F(1,14) = 5.23
|
p < .05
|
|
Tool 5 Task 5 Sex
|
F(1,14) = 20.90
|
p < .0005
|
Figure 7.16 Overall sex difference effects.
Figure 7.17 Results of separate analyses for males and females.
7.8 Discussion
On the whole, the experimental results strongly supported the experimental hypotheses as well as the high-level hypothesis that Guiard's Kinematic Chain model can be used to reason about bimanual performance for skilled 3D manipulative tasks. Reviewing this evidence:
H2: For the Easy task reversing roles of the hands will not have any reliable effect. The Grip effect was much smaller for the Easy task, but was significant at the p < 0.10 level, so one cannot confidently conclude there was no Grip effect. Nonetheless, for practical purposes, lateral asymmetry effects are much less important here.
H3: The importance of specialization of the roles of the hands increases as the task becomes more difficult. The predicted Grip 5 Task interaction was highly significant, offering strong evidence in favor of H3.
H4: Haptics fundamentally change the type of motor control required. Taken together, the experimental evidence for H1-H3 further suggests that the motor control required for the Easy conditions, where there was plentiful haptic feedback in the form of physical constraints, fundamentally differed from the Hard conditions.
The evidence in support of this final hypothesis underscores the performance advantages that are possible when there is haptic feedback to guide the task. Subjects devoted little cognitive effort to perform the Easy task, whereas the Hard task required concentration and vigilance.
This suggests that passive haptic feedback from supporting surfaces or active haptic feedback from devices such as the Phantom [147] can have a crucial impact for some tasks. This also underscores the difficulty of using a glove to grasp a virtual tool: when there is no physical contact, the task becomes a hand-eye coordination challenge, requiring full visual attention. With haptic feedback, it can be an automatic, subconscious manipulation, meaning that full visual attention can be devoted to a high-level task (such as monitoring an animation) instead of to the "tool acquisition" sub-task. For example, the PolyShop VR application [1] (fig. 2.5 on page 18) uses a physical drafting table as a support surface for 2D interactions, allowing the user to just touch the table to make selections.
These issues underscore the design tension between physical and virtual manipulation. The design challenge is find ways that real and virtual objects can be mixed to produce something better than either can achieve alone.
2 I would like to acknowledge private communication in which Yves Guiard discussed these issues.
3 The circles shown in this figure were 7.5 inches in diameter (outer circle) and 7.0 inches (inner circle) as originally drawn by the subject.
4 A simple reformulation of Buxton's experiment to produce targets at predictable locations did manage to reproduce the results from Kabbash. I would like to acknowledge personal communication during which Bill Buxton discussed this experiment on the Active Desk.
5 For bimanual rhythm tapping, conceptually the two hands are working together to produce a single combined rhythm. This task, however, does not address the hypothesis of right-to-left reference in bimanual manipulation.
6 This also doubled as a lateral preferences assessment, to ensure that each subject actually did prefer the "Preferred" grip to the "Reversed" grip.