FingerMapper: Mapping Finger Motions onto Virtual Arms to Enable Safe Virtual Reality Interaction in Confined Spaces

TL;DR

FingerMapper enables safe and precise virtual arm movements in confined VR spaces by mapping small finger motions, reducing fatigue and collisions while maintaining presence and enjoyment.

Contribution

This paper introduces FingerMapper, a novel method that maps finger motions onto virtual arms, allowing safe and comfortable VR interactions in confined spaces.

Findings

Reduces physical motion and fatigue compared to hand tracking.

Increases perceived safety and reduces collisions in confined spaces.

Maintains similar levels of presence and enjoyment as hand tracking.

Abstract

Whole-body movements enhance the presence and enjoyment of Virtual Reality (VR) experiences. However, using large gestures is often uncomfortable and impossible in confined spaces (e.g., public transport). We introduce FingerMapper, mapping small-scale finger motions onto virtual arms and hands to enable whole-body virtual movements in VR. In a first target selection study (n=13) comparing FingerMapper to hand tracking and ray-casting, we found that FingerMapper can significantly reduce physical motions and fatigue while having a similar degree of precision. In a consecutive study (n=13), we compared FingerMapper to hand tracking inside a confined space (the front passenger seat of a car). The results showed participants had significantly higher perceived safety and fewer collisions with FingerMapper while preserving a similar degree of presence and enjoyment as hand tracking. Finally, we present three example applications demonstrating how FingerMapper could be applied for locomotion and interaction for VR in confined spaces.

Equations

$$P_w=P_s+r\cdot l_{arm}\cdot {\hat{\mathbf{v}}}_{cs},whereris0.15<r<1and{\mathbf{v}}_{cs}=P_c-P_s$$

$$P_w=P_s+r\cdot (l_{arm}/2)\cdot ({\hat{\mathbf{v}}}_1+{\hat{\mathbf{v}}}_2),whereris0.15<r<1$$

$$R_{new}=\left\{\begin{array}[]{ll}R&{if\ R\ <\ D}\\ R+k(R-D)^{2}&{otherwise},\ k\ =\ 0.6\end{array}\right.$$

Full text

FingerMapper; Confined Spaces; Body Re-Association in VR

††journalyear: 2023††copyright: acmlicensed††conference: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems; April 23–28, 2023; Hamburg, Germany††booktitle: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI ’23), April 23–28, 2023, Hamburg, Germany††price: 15.00††doi: 10.1145/3544548.3580736††isbn: 978-1-4503-9421-5/23/04††ccs: Human-centered computing Virtual reality

1. Introduction

Current Virtual Reality (VR) experiences often require whole-body movement to increase presence (Slater et al., 1998; Usoh et al., 1999). Whole-body movement is a significant factor in engagement and sensory immersion (Isbister and Mueller, 2015), suggested by research on physicality in games. Although having whole-body movement in VR enables a higher presence and enjoyment, these techniques also require a larger physical space (Marwecki et al., 2018).

As VR Head-Mounted Displays (HMDs) are getting more mobile (e.g., Oculus Quest), users can immerse themselves wherever they wish. One prominent example would be using VR in transportation (e.g., planes (Williamson et al., 2019), cars (Hock et al., 2017; McGill et al., 2017), and others (McGill and Brewster, 2019; McGill et al., 2020)). However, these scenarios often happen inside confined spaces (Gugenheimer et al., 2019; Li et al., 2021), where whole-body movement is usually undesirable since VR users do not have enough space for interaction and could even accidentally collide with bystanders or surroundings.

In this paper, we present FingerMapper—a concept that maps small-scale finger motions onto large-scale virtual arm motions (Figure 1ab). FingerMapper enables VR experiences in confined spaces (e.g., the passenger seat in a bus, Figure 1c) with fewer physical motions and collisions while preserving presence and enjoyment. Although confined spaces are unavailable for whole-body movements, we can benefit from the beyond-real interaction in VR (Abtahi et al., 2022) by creating mapping metaphors between the VR user’s finger and the virtual arm. Additionally, users can learn different relationships between tracked and rendered motion in VR to control an avatar (Won et al., 2015). We implement two design variations (Attach and Direct) and present three applications (Figure 1d-f) that support VR interaction, including large swinging motions (FingerSaber), object interaction in space (FingerGrabber), and locomotion (FingerClimber).

To understand the usability of FingerMapper, we first conducted a within-subject user study (n=13) including both design variations (Attach and Direct) with two baselines (hand tracking and ray-casting) in a target selection task. While results show that FingerMapper had a higher task completion time and lower Virtual Body Ownership (VBO) compared to hand tracking, both mappings reduce the physical path length of participant’s hands and fatigue. Next, we conducted a user experience study in a confined space (n=13) where participants played FingerGrabber with FingerMapper (Attach) and hand tracking inside the front passenger seat of a car. Results show that FingerMapper had fewer collisions and a similar degree of presence and enjoyment to hand tracking. Overall, participants reported a significantly higher perceived safety using FingerMapper in a confined space compared to hand tracking. Most participants (10 out of 13 in Study 2) reported they prefer using FingerMapper in confined spaces to avoid colliding and keep feeling safe during the interaction. We see FingerMapper as an alternative for scenarios where the user wants whole-body motion without having the necessary physical space for interaction. Finally, we discuss the potential scenarios of using FingerMapper for future VR experiences (e.g., long-duration usage, context adaptive input).

Our work has three contributions. 1) The concept and implementation of FingerMapper leveraging small finger motions onto virtual arms and hands to enable VR interaction in confined spaces. 2) Insights from two studies show that FingerMapper has less physical movement, less fatigue, and fewer collisions compared to hand tracking. Participants also reported a similar degree of presence and enjoyment, a higher perceived safety, and preferred to use FingerMapper in confined spaces (10 out of 13 in study 2). 3) The implementation of three applications (FingerGrabber, FingerSaber, and FingerClimber) that demonstrate how to integrate FingerMapper for different interaction scenarios.

2. Related Work

Our work is related to confined spaces in VR and improving ergonomics for VR interaction. For the second subsection, we discuss prior research based on the framework of Abtahi et al. (Abtahi et al., 2022).

2.1. Confined Spaces as a New VR Context

The prevalence of VR products (e.g., HTC Vive, Oculus Rift) pushes the VR experience from a well-controlled environment (e.g., laboratory, demo booth) to the user’s private space. Plenty of new contexts emerge while people use VR technologies in their homes. One prominent context is social interaction, for example, including bystanders in the VR experience (Gugenheimer et al., 2017, 2018) or managing spaces between VR and non-VR users (Yang et al., 2018). Recent HMDs such as the Oculus Quest are mobile and standalone systems. They are easier to be carried around to new environments. Prior research has explored using VR in different forms of transportation (e.g., planes (Williamson et al., 2019), cars (Hock et al., 2017; McGill et al., 2017), and others (McGill and Brewster, 2019; McGill et al., 2020)).

In this work, we focus on the confined space, where a user cannot fully extend one’s limbs. This includes cars, planes, and other forms of public transportation. An example would be maintaining presence and productivity in VR while the user is confined in the rear seat of a car (Li et al., 2021). In these cases, users are usually restricted to a small physical space (e.g., car and shuttle) or surrounded by passengers (e.g., train), where whole-body motion is almost impossible. We explore how to keep presence and enjoyment in VR experiences within the context of the confined space by using beyond-real interaction—mapping finger motions onto virtual arms.

2.2. Improving Ergonomics for VR interaction with Hands and Arms

Reality-Based and Illusory Interaction

Directly using reality-based interaction (e.g., manipulation using hand tracking) benefits from our real-life experiences. Compared to ray-casting with controllers (or other metaphors), manipulating with hands is more natural and intuitive. Hand tracking benefits from the Degrees of Freedom (DoF) in manipulating virtual objects. Therefore, it becomes an ideal input for VR. On the contrary, ray-casting needs additional extensions to allow for rotation and retraction (e.g., controlling the depth of the ray (LaViola et al., 2017)) because the manipulation of ray-casting is not hand-centered. Additionally, simulating whole-body movements we have in the real-world experience also enhances presence (Slater et al., 1998; Usoh et al., 1999) and engagement (Isbister and Mueller, 2015) in VR.

Although reality-based interaction is natural, they sometimes cause fatigue and require physical space for whole-body motions. Prior works leverage visual dominance over proprioceptive cues (Burns et al., 2005) to create interaction techniques that reduce fatigue without breaking the Sense of Embodiment (SoE) and the reality-based metaphor (i.e., illusory interaction in (Abtahi et al., 2022)). SoE has three working constructs: self-location, sense of agency (being able to act and control), and body ownership (Kilteni et al., 2012). Recent research investigates how to reduce the physical path length and fatigue of limbs within the maximum arm’s reach while preserving VBO. Examples include redirecting visual targets (Montano Murillo et al., 2017), gradually directing the user’s hand to a comfortable posture (Feuchtner and Müller, 2018), or creating an offset between the physical and virtual hand (Li et al., 2018; Wentzel et al., 2020). However, these techniques still require sufficient physical space for whole-body motions and may be unavailable in the confined space.

Using a self-avatar with an input metaphor similar to reality enhances the sense of presence and embodiment in VR. High-fidelity avatars have better distance estimation in the near field (Ebrahimi et al., 2018), and it takes time for VR users to adapt to the dimensions and capabilities of self-avatars in VR (Day et al., 2019). Although using a self-avatar has been state-of-the-art, recent research explores avatars beyond reality. Won et al. (Won et al., 2015) studied how participants control an extended third arm controlled by feet in VR and showed they could learn and achieve higher task performance. In a recent study, Kato et al. (Kondo et al., 2020) showed using visual-motor synchrony can enable small VBO when mapping the right-hand thumb motion to the left virtual arm in VR. FingerMapper is inspired by mapping smaller finger motions to the virtual arm. Our approach can be an alternative when interacting in a confined space but the user still wants to preserve SoE.

Beyond-Real Interaction as a Solution for the Confined Space

VR interactions can go further beyond replication of reality (Abtahi et al., 2022). Previous research explores the beyond-real concept to enable distant interaction (Poupyrev et al., 1996; Feuchtner and Müller, 2017) and object manipulation (Frees et al., 2007). More recent examples aim at reducing physical movements also fatigue for tapping (McIntosh et al., 2020) and target selection tasks (Schjerlund et al., 2021). FingerMapper reduces physical path length and fatigue within the maximum arm’s reach of the VR user. Inspired by the beyond-real concept, we focus on mapping small-scale finger motions onto the virtual arms for an energy-efficient interaction, reducing physical movements to avoid fatigue for VR and collisions inside confined spaces. In addition, we are inspired by the small-scale mid-air gestures like Gunslinger (Liu et al., 2015), which redirects a ray-casting interaction from an energy-efficient position at the user’s hip towards a pointing gesture. Dukes et al. (Dukes et al., 2013) design a novel 3D interaction that scales the stroke survivor’s impaired arm movements into full movements in VR to provide expert action observation for post-stroke neurorehabilitation. Although these approaches are not resolving physical constraints, FingerMapper shares the concept of mapping smaller motions to more physically demanding interactions. When the physical space is limited, FingerMapper maps smaller finger motions to the virtual arm to enable a whole-body experience in VR. By reducing physical movements, we expect FingerMapper to have fewer collisions and enhance the VR user’s perceived safety.

3. Interaction Design

Our design is to enable virtual hands and arms for VR interaction in confined spaces. Hand tracking and ray-casting are unavailable because the former may obstruct the surroundings, and the latter only provides a pointing metaphor. FingerMapper aims to have the same functionality as hand tracking—having arm motion for object interaction—but less physical motion and fatigue.

3.1. Mapping Functions

Two mapping functions (Attach and Direct) for controlling the virtual arm using finger motions were implemented. The maximum arm reach was represented by a sphere having an origin at the shoulder ($P_{s}$) and a radius of the user’s arm length ($l_{arm}$, Figure 2a). We set $l_{arm}$ (from shoulder to wrist) as 60 cm according to the average arm length (Pheasant and Haslegrave, 2005). The goal was to use hand tracking data (e.g., position and rotation of finger joints and wrist) to calculate the virtual wrist position ($P_{w}$) and create virtual arm movements that can cover the VR user’s maximum arm reach.

Attach: Casting the virtual wrist onto the surface of the maximum arm reach.

The Attach mapping design was based on homuncular flexibility in that humans are flexible at learning a different relationship between tracked and rendered motion (Won et al., 2015). First, we cast a ray along the direction from the metacarpophalangeal (MCP) joint to the proximal interphalangeal (PIP) joint. This ray reaches the surface of the maximum arm reach sphere (Figure 2a). We chose the proximal phalanx since it is less intervened when bending the finger inwards and called this intersection point the casting position ($P_{c}$). The wrist of a hand model locates at $P_{c}$. Here, the user’s index finger is in the pointing gesture, so the virtual arm is also fully stretched. They can move their proximal phalanx of the index finger or their wrist to cast $P_{c}$ on the surface of the maximum arm reach, changing the virtual arm’s direction and position.

Next, to reach positions inside the maximum arm reach sphere, we retracted the virtual wrist along the direction between the casting point ($P_{c}$) and the shoulder ($P_{s}$). A retraction fraction $r$ was the distance between the fingertip and MCP joint ($d$) divided by the length of the index finger. As the user bends their index finger, $r$ decreases and the virtual wrist moves closer to the shoulder (Figure 2bc). Because the virtual wrist should not pass through the VR user’s shoulder, the value of $r$ ranges from 0.15 to 1. Therefore, the virtual wrist position ($P_{w}$) was the retraction amount multiplied by the arm length and the unit vector from the shoulder position to the casting point (Equation 1). Note that $P_{c}$ and $P_{w}$ are identical when $r$ equals one. While the user retracts their index finger, $P_{w}$ moves toward the user, along the direction from $P_{c}$ to $P_{s}$.

[TABLE]

Finally, because Attach only calculates $P_{w}$, we compute the direction of the virtual upper and lower arm by solving inverse kinematics, in which $P_{w}$ is the target, and $P_{s}$ is the origin. We constrained the virtual upper and lower arm movements according to human arm anatomy. The idea is to have more natural virtual arm movements. We applied 1€ Filter (Casiez et al., 2012) to smoothen virtual arm movement.

Direct: Re-associating finger segments to virtual arms.

The Direct mapping function is inspired by body re-association and the notion that providing visuomotor synchrony could induce a portion of VBO (Kondo et al., 2020). By using the similarity of index finger joints and virtual arm joints, users control the virtual arm movements by bending their index fingers. The idea is to use this similarity of posture between finger and arm to give the user a better understanding of agency and VBO.

Two vectors, MCP to PIP joint (${\mathbf{v}_{1}}$) and PIP joint to fingertip (${\mathbf{v}_{2}}$), are mapped onto the upper and lower virtual arms (Figure 3a). When the user’s index finger is straight, the virtual arm also remains in the same direction and reaches the maximum arm reach. While the user bends their index finger (Figure 3b), the virtual arm follows the direction of two vectors on the index finger and bends inwards simultaneously. Equation 2 shows the virtual wrist position ($P_{w}$) is the sum of ${\mathbf{v}_{1}}$ and ${\mathbf{v}_{2}}$ from the shoulder position ($P_{s}$). We scaled up by multiplying half of the arm length ($l_{arm}$).

Figure 3c shows the arm length decreases by a fraction $r$ as the retraction amount of virtual arm length while bending the finger. Here, $r$ equals the distance between the fingertip and MCP joint divided by the length of the index finger, ranging from 0.15 to 1. Because the virtual upper and lower arm lengths are equal, the new length equals $r$ multiplied by half of the arm length ($l_{arm}/2$). The virtual wrist position ($P_{w}$) becomes the sum of two unit vectors of the index finger and multiplies the retracted arm length from $P_{s}$ (equation 2). The user can control the virtual arm direction by moving their fingers, and the whole virtual arm rotates when the user rotates their wrist along the direction of the index finger. The difference between Attach is that the arm direction follows the index finger’s angle. Our implementation applied 1€ Filter (Casiez et al., 2012) to hand tracking data for smoothening virtual arm motions.

[TABLE]

3.2. Spatial Extension

While Attach and Direct enable virtual arm movements in three dimensions within the maximum arm’s reach, we found that the sense of agency over the virtual arm breaks when moving the wrist without the corresponding virtual wrist movement. We observed this effect during the early prototyping process of FingerMapper. The sense of agency is a construct of being able to control and manipulate, and it is sensitive to any temporal discrepancies between the execution of a self-generated movement and visual feedback (Kilteni et al., 2012). Therefore, we developed a spatial extension function that changes the virtual arm length to create virtual wrist movements, maintaining the sense of agency and overall SoE while using FingerMapper. We applied this function to both Attach and Direct.

The spatial extension was implemented with a similar approach to the Go-Go technique (Poupyrev et al., 1996). The Go-Go technique is designed for distant interaction in the virtual environment (from mid-range to far-range) and requires enough space for whole-body movement (need to extend arm). On the contrary, FingerMapper enables virtual arm interaction in the “close and mid-range” space (within arm span) using mainly finger mapping. The two techniques have distinct design objectives. In FingerMapper, users can move only their fingers to reach a position in space and their wrists to support their motion. The observed behavior often consists of both. Depending on the context of usage, one of them would become more dominant but never exclude the other. FingerMapper only applies spatial extension to preserve the sense of agency while interacting within close and mid-range. The part of Go-Go is only triggered during larger motions, and the close-range interaction is supported only by FingerMapper. The Go-Go technique in a confined space would become hand tracking in our scenario.

Figure 4a shows the default posture that the user takes a sitting position and fixes their elbows to both lateral sides of the waist. The default posture simulates using VR in a confined and restricted interaction space. The sitting shoulder height is 20 cm below the eye, and the sitting elbow height is 33.5 cm below the shoulder, approximated using the average body dimensions (Pheasant and Haslegrave, 2005). The chest position is 37 cm below the HMD, and the chest height is in the middle between the sitting shoulder and elbow height. We calculate the projected distance from the physical wrist to the chest ($R$), as shown in Figure 4b. This distance is projected onto the plane that always has its normal vector pointing upwards from the user’s head since we need the forward and backward spatial offset. The projected distance ($D$) of default posture is the criterion of spatial extension. We set the value of $D$ as 18 cm through our first experimentation. As shown in Figure 4c, when the user’s hand is within the range of $D$, The spatial offset is $D-R$. When the user extends their hand out of the range of $D$ (Figure 4d), we compute $R_{new}$ using the Go-Go function shown in equation 3. The spatial offset then becomes $R_{new}-D$. These design decisions of the spatial extension were made during the implementation process within the group of authors. The offset is added to the radius of the shoulder sphere of Attach. Therefore, the user can perceive the arm extending or shortening when they move their hand forward and backward. For Direct, we increase or decrease the arm length by this offset.

[TABLE]

3.3. Finger Motion

In an early prototype, we experimented with a rigid hand model as the virtual hand that either points or grabs. However, this static representation felt more awkward since the virtual hand model did not resemble the finger motions we have in reality. To increase the sense of agency and VBO, we map the tracking information of physical fingers to virtual fingers. Although index fingers are occupied for positioning the virtual arm in space, the current implementation still has more precise finger movement at this new location and performs simple manipulations.

3.4. Selection

Selection is essential in VR interaction. Despite having the actual finger tracking data attached to the virtual hand, the current implementation partially impedes the ability to select objects with a pinch since bending the index finger also moves the virtual arm. To overcome this limitation, we added functionality based on two hand gestures (ThumbButton and GrabSelect) to enable selection with FingerMapper.

In ThumbButton, the user moves their thumb to touch the knuckle of the middle finger, similar to pressing a button (Figure 5a). We track the distance from the thumb fingertip to the middle finger knuckle (PIP joint). The user triggers selection once detecting the thumb fingertip contacts the middle finger knuckle. Figure 5b shows GrabSelect, where the user holds their hand with the middle, ring, and pinky fingers at the same time. We compute an average distance from the middle, ring, and pinky fingertips to the wrist, leveraging the metaphor of grabbing an object to trigger selection.

We tested both trigger mechanisms within the group of authors during the implementation process of FingerMapper to make design decisions. While performing the motion in GrabSelect feels more natural and closer to grabbing an object, tucking in the middle, ring, and pinky finger results in a higher distortion of the index finger position. These joint movements with the index finger are because the fingers are controlled by the same muscle group (Schwarz and Taylor, 1955). This distortion causes the Heisenberg effect of spatial interaction (Wolf et al., 2020)—a user points or selects virtual content using spatially tracked input devices, and a discrete input such as a button press disturbs the tracking position, resulting in a different selection point. Here, GrabSelect disturbed the index finger position, so this effect exists. Therefore, We selected ThumbButton since the muscle group involved in the thumb motion is independent of the index finger, resulting in less disturbance of the selection.

3.5. Hardware Requirement

Our current setup requires 6-DoF tracking on the wrist, joints, and fingertips. We use Oculus Quest hand tracking to acquire all the necessary position and rotation data. The concept is also compatible with other VR systems providing similar hand-tracking data.

4. Applications

We highlight three usage scenarios of FingerMapper with different interactions (e.g., object interaction, arm swinging motion, and locomotion). The goal is to show that when the physical environment is unavailable for whole-body movement, the user can switch to FingerMapper as an input metaphor for confined space. Three applications were originally designed for whole-body interaction. We only replace the regular hand tracking (or controllers) with FingerMapper to show VR users can interact in a confined space using FingerMapper in different scenarios.

4.1. FingerGrabber

FingerGrabber is an application inspired by VR sandbox games like Job Simulator. The interaction in this type of game is often selecting, grabbing, and manipulating any object in the virtual environment. This interaction style can represent any generic selection and manipulation task a user has to conduct in VR. The virtual objects in these VR applications are usually scattered within the arm span of the user. Here, we argue that FingerMapper could be a good solution by reducing physical movement to enable the gaming experience in confined spaces.

Since such games often strive for a well-written narrative, FingerGrabber added game mechanics to have a more engaging experience. The user is in front of a pit that is successively filled with different types of balls (e.g., basketball, volleyball, and football shown in Figure 6a). Figure 6b shows that the user grabs each ball and place it in one of the three corresponding baskets. We implement the game using Attach mapping and ThumbButton, allowing the player to conduct the full range of interactions necessary for this type of game, i.e., select, grab, and manipulate.

4.2. FingerSaber

The VR rhythm game, BeatSaber, inspired us to use fingers to perform the arm swing motion (Figure 7a). In BeatSaber, different colored cubes move toward the user and follow the beats of the music. The user’s goal is to slice every cube in the direction indicated on them. The original concept of BeatSaber revolves around having a whole-body experience with large arm movements and sometimes even moving their whole body. To enable FingerSaber, we chose the Attach mapping and positioned two virtual lightsabers in the user’s virtual hand. Since no selection is needed, we kept the lightsabers constantly attached to the user’s hand without using a grabbing motion. Our technique enables the BeatSaber experience to work using small finger motions instead of large arm motions without modifying the game mechanics.

4.3. FingerClimber

FingerClimber is inspired by current VR games like The Climb, where the user goes to the top of a mountain using different climbing grips attached to the cliff. In this game, the user usually stretches their whole arm and sometimes the whole body. We used Direct and hoped this mapping would help the user to control the specific shape of the virtual arm. The goal is to avoid users climbing with constantly stretched-out arms. As the selection mechanism, we used GrabSelect since it imitates the natural motion of holding onto a climbing grip (Figure 8a). The user grabs a climbing grip with the virtual arm and pulls themselves up using a downward motion with the finger (Figure 8b).

5. Study 1: Target Selection

We conducted a target selection study to compare Attach and Direct to common inputs (e.g., hand tracking and ray-casting). The goal was not to outperform the baselines since none of them can achieve our motivation—arm and hand interactions for VR in a confined space. The comparison in this study was to position our prototype techniques between full embodiment (hand tracking) and full efficiency (ray-casting). We observed how good users could control FingerMapper and how they impacted metrics (e.g., presence, VBO).

5.1. Apparatus and Setup

The study was conducted with Oculus Quest 1 either in our lab or remotely at the participants’ homes. Participants were seated on a chair. Five remote participants used their devices at home to avoid unnecessary physical contact and crowds. We instructed remote participants to clear a space that did not cause obstructions while using hand tracking. The study application was given to participants as a .apk file. The other participants used the lab setup on the same headset. To treat both setups as similar as possible, remote and lab participants were given the instructions via a website and could ask questions in person or remotely. We disinfected the HMD cushion for lab participants for hygiene concerns.

5.2. Study Design

The study was a within-subject design with one independent variable (technique), including Attach, Direct, hand tracking, and ray-casting (Figure 9). We were interested in what degree we could reduce physical movements and fatigue, so we compared with hand tracking. Additionally, ray-casting was included as a second baseline for understanding efficiency. Here, we used an implementation similar to (Bowman and Hodges, 1997), where a ray is casting from the fingertip of the index finger (Figure 9d). A red pointer was located on the fingertip to visualize the selection point. In ray-casting, the pointer was on the tip of the ray, indicating the contact point when casting on a target.

We used a world-fixed layout (Figure 10b) in the target selection task of the previous work (Montano Murillo et al., 2017; Wentzel et al., 2020). Here, interactable objects were spatially distributed around the user. We argue that this arrangement represents the application scenario with arm and hand interactions in a confined space. The layout dimension was determined by the participant’s arm span (A). The target layout had 18 spheres on two layers located at the depth of the maximum arm reach (0.44A) and a half of this range (0.22A). The closer layer had 3 $\times$ 4 targets evenly distributed between 0.76A $\times$ 0.4A, and the other layer had 3 $\times$ 2 targets that were placed evenly between 0.25A $\times$ 0.4A from the participant. Targets ranged evenly in the height from 0.4A to 0.8A to simulate a seated interaction space. We calculated the distance between every two targets in the layout and separated them into short-, medium-, and long-range by using the 25th and 50th quantile of distances. Each type of range had an equal amount of tasks in each condition, and they were randomized within 30 tasks of each condition. Each participant had to complete 120 tasks (4 techniques $\times$ 30 tasks). The order of technique was counterbalanced by the 4 $\times$ 4 Latin square.

For dependent variables, we measured task time, the physical path length of the used wrist, and the virtual path length of the pointer in VR for each task. The path length data was divided by the target distance of each task as a ratio for comparison. The interaction volume was calculated by recorded physical fingertip trajectory. In the post-test questionnaire, we used NASA-TLX (Hart and Staveland, 1988) for the subjective fatigue, the avatar embodiment questionnaire (Gonzalez-Franco and Peck, 2018) with three subscales (ownership, agency, and location) for VBO, and Igroup Presence Questionnaire (IPQ) (Schubert et al., 2001; Regenbrecht and Schubert, 2002) for presence. Participants responded with a 7-point Likert scale inside VR. Finally, the tracking quality of the Oculus Quest was recorded through every frame during each task for filtering out the data of tracking loss.

5.3. Procedure and Task

Participants were introduced to the purpose of our study and signed the consent form. We measured the arm length111We define arm length as the distance from the shoulder to the wrist. Please see the figure in the supplementary material., index finger length, and arm span for calibration. After participants put on the HMD and entered the application, they had to re-center themselves in VR and calibrate their body dimensions.

Participants had to maintain the default posture (Figure 4a), placing their elbows on both sides of their waist to simulate a confined space. They touched a white cube in front of them to begin (Figure 10a), and two randomly selected spheres became green as the targets (Figure 10b). They were free to choose their right or left hand for a task, but they had to start and finish the task with the same hand. If they used the right hand to select the first target, the second one had to be selected with the right. Using the wrong hand for the second sphere was considered an error. This approach is common in these types of studies (Wentzel et al., 2020) and allows users to keep the natural behavior. The task started as the participant touched one of the green targets using the red pointer. This target turned blue, and the participant had to move the pointer to the second target fast and steadily, and this task was marked as finished. A successful task came with a sound effect, and participants returned to the default posture, waiting for a two-second countdown, and continued to the next task by touching the white cube.

Participants had to practice and succeed in at least ten tasks before starting each condition. They were asked to practice until they felt confident and comfortable performing the task with each technique. This setup should prevent a learning phase at the beginning of the study exposure for each condition but result in different training duration for each participant. The average task count of practicing was 11.1 times ($SD=1.6$), showing that participants had a similar training time. The overall average success rate of practicing was 98% (Attach: $M=97.3\%,SD=6.5$, Direct: $M=100\%,SD=0$, hand tracking: $M=97.3\%,SD=5.1$, and ray-casting: $M=97.4\%$, $SD=5.1$). These results indicate that participants’ mastery of each technique was similar. After each condition, they filled out a post-test questionnaire in VR and took a 30-second rest. After finishing all the conditions, we asked for their subjective feedback and preferred interaction if they used VR in a confined space. Each condition took about ten minutes, and the whole study finished in one hour.

5.4. Participants

Volunteer participants (n=13) took part in the study via convenience sampling (five females and eight males, age: $M=26.3,SD=3.4$). Six participants had experience developing VR or research, and the rest used VR occasionally.

5.5. Results

The following results were based on 1560 selection tasks. The hand tracking condition had overall more tracking loss than other techniques, probably because FingerMapper and ray-casting had more static and close-up postures, so they were easier to track. In total, 197 (12.6%) tasks were removed. The accuracy of each condition was Attach 98.4%, Direct 99.4%, hand tracking 99.0%, and ray-casting 98.1% (overall 98.7%). Each technique had a high success rate and showed participants could control the techniques to complete the task. We only considered successful tasks and excluded the task time and path lengths that exceeded three standard deviations from the average in each condition. There were 30 tasks (2.2%) removed among 1363 well-tracked tasks.

For task time and path length ratios, one-way within-subject ANOVA with Greenhouse-Geisser correction was performed and followed by the pairwise comparison with Bonferroni correction for the post-hoc analysis. We used the Friedman and Wilcoxon test for the post-hoc comparison of the Likert-scale data.

Task Time. There was a significant difference in technique on the task time ($F_{3,36}=90.03,p<.01,\eta^{2}=0.88$), as shown in Figure 11a. The post-hoc comparison revealed that the mean task time of Attach ($M=1.71s,SD=0.24$) and Direct ($M=2.05s,SD=0.44$) were significantly higher than hand tracking ($M=0.84s,SD=0.25$) and ray-casting ($M=0.90s,SD=0.19$) ($p<.01$). There was no significant difference between Attach and Direct ($p=.09$), neither hand tracking and ray-casting in the post-hoc analysis ($p=.68$).

Path Length Data. All the path length data was divided by the target distance for each task. These metrics represented the relationship between physical movement and virtual distance. Figure 11b shows the path length ratio of the physical wrist. We found a significant difference in the motion of the physical wrist ($F_{3,36}=51.47,p<.01$, $\eta^{2}=0.81$) in technique. In the post-hoc comparison, ray-casting ($M=0.42,SD=0.10$) had a significantly lower path length of the physical wrist compared to the other three conditions ($p<.01$). The results indicated no significant mean difference in the path length of the physical wrist between Attach ($M=0.69,SD=0.14$) and Direct ($M=0.66,SD=0.20$) ($p=.94$). Hand tracking ($M=0.98,SD=0.09$) had a significantly higher path length ratio compared to Attach and Direct ($p<.01$).

There was a significant difference in the path length of the virtual pointer ($F_{3,36}=11.53,p<.01,\eta^{2}=0.49$). The virtual path length ratio was calculated as the virtual path traveled divided by the target distance (Figure 11c). For the virtual pointer of ray-casting, we recorded the path length when the virtual pointer was not intersecting with targets. Because the virtual pointer (cursor) was at the end of the ray, there were additional movements on this virtual pointer, resulting in the average ratio was 2.09 for ray-casting (Figure 11c). These metrics were an indicator of how much unnecessary movement was done. Hand tracking ($M=1.28,SD=0.09$) had a significantly lower path length ratio of a virtual pointer than the other three techniques ($p<.01$). Other comparisons revealed no significant mean difference in the virtual path length (Direct-Attach: $p=.054$, Ray-Attach: $p=.51$, Ray-Direct: $p=.65$). Attach ($M=1.83,SD=0.40$), Direct ($M=2.30,SD=0.89$), and ray-casting ($M=2.09,SD=0.43$) all resulted in significantly more unnecessary motion, indicating overshooting or correction behavior while selecting. Overall, both finger mappings reduced the physical motion but increased the virtual motion compared to hand tracking.

Figure 12 shows the physical position of all participant’s index fingertips from two perspectives (HMD movement included). We observed the pattern of reaching toward the target layout. Participants’ movements were closer to the center in FingerMapper functions and ray-casting than hand tracking. Black points represent the HMD position. Since we conducted Study 1 in an open space without any constraints, participants still had some body movements. We calculated these maximum bounding cubes to indicate interaction volume, showing how much space the participants used with each technique. These volumes allowed us to quantify how much we could reduce this volume. The hand tracking had the largest interaction volume ($0.888\ m^{3}$) followed by Direct ($0.782\ m^{3}$), Attach ($0.602\ m^{3}$) and ray-casting ($0.506\ m^{3}$). Attach could reduce the interaction volume by 32.2% to be comparable to ray-casting. Also, we analyzed the interaction volume data by participants and conditions. By performing a logarithmic transformation, we found a statistically significant difference between technique ($F_{3,36}=32.8,p<.01,\eta^{2}=0.73$). In a post-hoc analysis, hand tracking ($M=0.44,SD=0.14$) had a statistically significant higher volume than Attach ($M=0.21,SD=0.1$), Direct ($M=0.22$, $SD=0.12$), and ray-casting ($M=0.11,SD=0.05$) ($p<.01$ for three comparisons). There was no difference between Attach and Direct, but both had a statistically significant higher volume compared to ray-casting (Ray-Attach: $p=.02$, Ray-Direct $p=.04$).

Fatigue and NASA-TLX. Figure 13a shows the results of NASA-TLX. We were particularly interested in the mental and physical demand, and our analysis revealed significant differences (Mental: $\chi^{2}_{3}=24.64,p<.01,\eta^{2}=0.63$, Physical: $\chi^{2}_{3}=19.73,p<.01,\eta^{2}=0.51$). The mental demand of hand tracking ($M=1.62,SD=1.19$) and ray-casting ($M=2.31,SD=1.80$) were significantly lower than Direct ($M=4.62,SD=1.94$) and Attach ($M=4.54,SD=2.15$) ($p<.05$). For the physical demand, hand tracking ($M=5.85,SD=0.90$) had a higher score compared to Attach ($M=3.85,SD=1.14$), Direct ($M=4.05,SD=1.75$), and ray-casting ($M=2.54,SD=1.66$) because participants had to reach out their arm in the hand tracking condition. Here, comparisons between Attach-Hand and Ray-Hand found significant differences in the post-hoc analysis ($p<.05$). For the other four subscales in NASA-TLX, there were significant differences in temporal demand, effort, and frustration (Temporal: $\chi^{2}_{3}=13.27,p<.01,\eta^{2}=0.34$, Effort: $\chi^{2}_{3}=12.20,p<.01,\eta^{2}=0.31$, Frustration: $\chi^{2}_{3}=8.17,p=0.043,\eta^{2}=0.21$), and there was no significant effect on the subscale of performance ($\chi^{2}_{3}=6.36,p>.05,\eta^{2}=0.16$). One significant mean difference in temporal demand was found in the post-hoc comparison where Attach ($M=2.92,SD=1.44$) has a lower score than hand tracking ($M=4.38,SD=1.39$) ($p<.05$).

VBO and Presence. Figure 13b shows the score of three subscales for VBO. The values were averages of individual questions and ranged between -3 and 3. A score higher than zero represents the feeling of VBO, and a score lower shows the lack thereof. There were significant differences in techniques on three subscales we measured (Ownership: $\chi^{2}_{3}=28.97,p<.01,\eta^{2}=0.74$, Agency: $\chi^{2}_{3}=18.22,p<.01,\eta^{2}=0.47$, Location: $\chi^{2}_{3}=14.56,p<.01,\eta^{2}=0.37$). For the post-hoc comparison in the ownership subscale, we only found hand tracking ($M=2.49,SD=0.46$) was significantly higher than Attach ($M=0.51,SD=1.14$) and Direct ($M=-0.39,SD=1.59$) ($p<.01$). For the agency subscale, hand tracking ($M=1.65,SD=0.58$) was significantly higher than Direct ($M=0.48,SD=1.17$) and Attach ($M=0.83,SD=0.66$) ($p<.05$). No significant mean difference was found in the post-hoc comparisons of the location subscale.

Finally, Figure 13c shows the General Presence in IPQ. Please see the supplementary material for all subscales (Spatial Presence, Involvement, and Experienced Realism). There were significant differences in General Presence ($\chi^{2}_{3}=14.51,p<.01,\eta^{2}=0.37$) and Experienced Realism ($\chi^{2}_{3}=11.5,p<.01,\eta^{2}=0.30$). No mean differences were found in the post-hoc comparison of General Presence and Experienced Realism. No significant difference was revealed in Spatial Presence ($\chi^{2}_{3}=4.42,p=.22,\eta^{2}=0.11$) and Involvement ($\chi^{2}_{3}=3.92,p=.27,\eta^{2}=0.10$).

5.6. Discussion

In the following, we position both FingerMapper mapping functions between two extremes: hand tracking and ray-casting.

Performance. We found using FingerMapper leads to a significant increase in the task time. Parts of this can be explained by the learning curve that we observed. Despite the initial training phase, participants still showed improvements during the study. Additionally, Attach and Direct showed a higher mental demand, effort, frustration compared to hand tracking and ray-casting. Participants also tended to a lower performance222The scale of performance is reversed in NASA-TLX. A low-performance score means the user felt perfect considering how successful (or satisfied) while performing a task, and a high-performance score means the user felt failure while doing a task. using FingerMapper. This is probably due to the novelty of the technique and the initial phase, where users needed to get used to the different mappings. Overall, we argue the task time is still within an acceptable range since eight participants reported in the questionnaire that they would use one of the finger mappings to interact in confined spaces (two Direct, six Attach, and five ray-casting over 13 participants).

Motion and Fatigue. Our goal was to reduce fatigue, physical motion, and interaction volume for virtual arm interaction in a confined space. Results indicated that FingerMapper had significantly lower scores on the physical demand than hand tracking. We also found that ray-casting had an even lower score, positioning Attach and Direct in between both of them.

Attach reduced the necessary physical path length down to approximately 67.8%, positioning it again between ray-casting (57.0%) and hand tracking (considered here as 100%). Additionally, we found that Attach reduced the overall interaction volume compared to hand tracking by 32.2%. For the average of the individual interaction volume, Attach reduced this metric down to 46.7% of hand tracking. We argue Attach (0.602 $m^{3}$) and ray-casting (0.506 $m^{3}$) have similar maximum interaction volumes while allowing the case to perform large virtual arm motions in confined spaces. Also, ray-casting has substantially less DoF, requiring additional extensions to allow for rotation and retraction (e.g., controlling the depth of the ray (LaViola et al., 2017)). Without these extensions, ray-casting cannot enable manipulation as FingerMapper. While hand tracking had little additional virtual motions, both FingerMapper and ray-casting showed a similar (Figure 11c) level of additional virtual distance traveled, indicating participants had a quick burst in the direction of the target with overshooting and had an additional correction phase afterward.

VBO and Presence. Here we mainly consider Attach, Direct, and hand tracking because the concept of “body ownership” is, in general, difficult to apply to an abstract pointing device such as a ray (Lin and Jörg, 2016). However, to control the visual of virtual arms in each condition, we used a tracked visualization of the user’s hand. This probably led some participants to rate VBO toward the hand and not the actual ray as we intended. We still report the VBO score of ray-casting for completeness, but we do not provide further interpretation of this score.

Initially, we expected Attach and Direct can achieve a VBO score similar to hand tracking. However, Direct could not preserve any degree of VBO compared to Attach. Although Direct still preserved some VBO in the agency subscale, it was dominated by the visual of virtual arms. Therefore, Direct had neutral or even negative scores in the ownership and location subscales. It was less preferred when asked whether participants would use it in a confined space. Nevertheless, Attach could preserve a general sense of ownership, agency, and even location scores.

In General Presence, all conditions received an average score above four on the 7-point scale. We believe all techniques could keep a rather high level of presence and did not break the user experience because General Presence scores were all above four. Nevertheless, hand tracking had the highest score followed by ray-casting, Direct and Attach. In Spatial Presence and Involvement, four techniques did not show statistically significant differences in their score. The reason might be all participants were in the same virtual environment and interacted with the same task. Experienced Realism ranged between three to four, probably caused by the virtual environment being almost clear for study control.

Observations and Qualitative Feedback. We observed participants with little to no VR experiences tended to express frustration when they could not fully control FingerMapper during practicing. They also preferred ray-casting for using VR in a confined space. Since the task in the study can be achieved in a simple way physically (i.e., hand tracking), spending extra effort to complete this easy task with FingerMapper indeed troubled some participants.

The participants with more VR experiences could learn and control the mapping faster than others. One participant who frequently uses VR for gaming mentioned (P13), “At the beginning, I struggled a little with controlling, but once I understood the mapping, I think I can use this mapped arm to do more VR interaction.” FingerMapper, despite a higher task time, has lower fatigue and physical motion. Results show that Attach can preserve the user’s presence and partially some degree of VBO.

6. Study 2: User Experiences in the Confined Space

Study 1 was to understand the exact interaction time and experience metrics in a controlled environment. In Study 2, we wanted participants to experience using FingerMapper inside a confined space, compared to hand tracking, the current state-of-the-art for VR interaction. Our focus was on experience metrics (presence, enjoyment) and safety metrics (perceived safety, amount of collisions).

6.1. Apparatus and Setup

While the setup in study 1 was in an open space, study 2 was in the front passenger seat of a standing car (Figure 14, Dacia Logan MCV) parked in our institution. The main goal was to test our technique inside a realistic representation of a confined space. We used FingerGrabber deployed on an Oculus Quest 2 for the scenario because it involves object interaction. For the techniques, we used the same version of hand tracking and Attach in Study 1. We chose Attach because it outperformed Direct in VBO and performance. Participants were seated on the front passenger seat, using the stationary boundary of the Oculus Quest. The distances from the chest to the surroundings were roof: 50 cm, windshield: 85 cm, right car door: 40 cm, and left car door: 110 cm.

6.2. Study Design

The study was a within-subject design with one independent variable techniques having two levels (Attach and hand tracking). Each participant had to experience FingerGrabber (subsection 4.1) with Attach and hand tracking. FingerGrabber aims to simulate VR experiences similar to Job Simulator, where the virtual objects would scatter within the arm space of the user for selection and manipulation tasks. These VR applications may be unavailable in a confined space because the virtual objects may be out of reach. Here, we argue that Attach could be a good solution, reducing physical movement to keep the user safe during the interaction and preserving presence and enjoyment.

We did not include pointing techniques, such as ray-casting, since they require adapting the game or application. We envision FingerMapper as an alternative to generate input stimuli for virtual arm interactions that can be used across all current applications relying on full-scale hand tracking without modifying applications.

The order of the two conditions was counterbalanced. To measure the user experience, the participant had to fill out a questionnaire after each condition, including SUS presence questionnaire (Usoh et al., 2000), avatar embodiment (Gonzalez-Franco and Peck, 2018) for VBO, and Intrinsic Motivation Inventory (IMI) (McAuley et al., 1989) for enjoyment. We added two 7-point Likert scale questions for perceived safety in the confined space: “I feel safe when I was in the game,” and “I feel worried about touching the car interior or other passenger when playing.” The study was video recorded. Therefore, we observed and counted collisions in each video footage to see if Attach could reduce physical space for VR interaction. A collision was defined as the participant physically contacting the interior parts of the car. At the end of the study, participants had to rate (7-point Likert scale) which interaction techniques they prefer for using VR confined spaces.

6.3. Procedure

Participants were introduced to the study and signed the consent form. All participants were new to the system and did not attend Study 1. To reduce the “wow” effect, each participant practiced with the system in a training session prior to the data collection. They were free to practice Attach or hand tracking in an open space until they felt comfortable with the interaction (on average five minutes) until they could successfully select and put the target at the right place three times in a row. This step ensured each participant had a similar level of understanding and skill despite their prior VR experience. When they sat in the car, we gave the following instruction, “imagine you are sitting inside a car of your friend (or a family member), please adjust the seat into a comfortable position as you usually do.” After adjusting the seat, the participant put on the Oculus Quest 2. The game lasted two minutes. Here, the goal was not to run performance-oriented evaluation (done in Study 1) but to observe users interacting in a realistic environment. The impact of the confined space (collisions) and perceived safety were independent of the duration of exposure. Collisions will happen right away, affecting the perceived safety. After each condition, they filled out the prior mentioned questionnaires in VR. When the participants completed both techniques, the experimenter interviewed them on their overall experience in VR and their preference of techniques. The whole study took an hour to complete.

6.4. Participants

Participants (n=13, six females and seven males) ages ranged from 18 to 33 ($M=26.0,SD=4.0$). All participants were recruited from our institute using convenience sampling. On a 5-point Likert scale, participants rated their expertise in VR between below average and average ($M=2.2,SD=0.8$). The frequency of using VR was rated between almost never and sometimes ($M=1.8,SD=0.8$). Overall, the sampled participants were mostly novice VR users. Participants reported normal or corrected vision and no motor impairments. The study was carried out following local health and safety protocols. Each participant was compensated with 10€ for their participation.

6.5. Results

Figure 15a shows Attach received a lower score in VBO compared to hand tracking. Dependent t-test found hand tracking had a higher score in ownership ($Mean_{Diff}=-1.15$, $95\%\ CI[-1.92,-0.38]$), agency ($Mean_{Diff}=-0.98$, $95\%\ CI[-1.55,-0.41]$, and location ($Mean_{Diff}=-1.03$, $95\%\ CI[-1.53,-0.52]$) of the avatar embodiment questionnaire. In each subscale, we revealed a significant difference (Ownership: $t_{12}=-3.27,p=0.007,Cohen^{\prime}s\ d=0.90$; Agency: $t_{12}=-3.73,p=0.003,Cohen^{\prime}s\ d=1.03$; Location: $t_{12}=-4.44,p=0.0008,Cohen^{\prime}s\ d=1.23$). However, the VBO scores of Attach were still above neutral, which is in line with the results in Study 1. Keep in mind that the scale of the avatar embodiment questionnaire ranges from -3 to 3. Our results would be equivalent to a rating of above 4 (above average) on a 7-point Likert scale.

For presence (Figure 15b), a paired t-test found no statistically significant difference between techniques ($Mean_{Diff}=-0.3$, $95\%\ CI\ [-0.9,0.3]$, $t_{12}=-1.21,p=0.25,Cohen^{\prime}s\ d=0.34$). Although there was also no statistical significance in enjoyment ($Mean_{Diff}=-0.1$, $95\%\ CI\ [-0.7,0.5]$, $t_{12}=-0.51,p=0.62$, $Cohen^{\prime}s\ d=0.14$), both Attach ($M=5.0$) and hand tracking ($M=5.1$) received positive scores, showing participants enjoyed using FingerGrabber.

Our results show that Attach benefited from having less physical movement that resulted in fewer collisions with the environment compared to hand tracking. Re-watching all the video recordings of the participants, one author coded each instance (location, time) of a collision with any part of the car. We found that Attach had fewer collisions ($M=0.3,SD=0.5$) compared to hand tracking ($M=4.9,SD=3.5$) (Figure 15c). A dependent t-test found a significant difference between the two techniques ($Mean_{Diff}=-4.5$, $95\%\ CI\ [-6.5,-2.5]$, $t_{12}=-5.02,p=.0003$, $Cohen^{\prime}s\ d=1.39$). Most of the collisions happened on the right car door (84%). Additionally, we found a difference in the participant’s perceived safety scores between the two techniques. Figure 15b shows the response to “I feel safe when I was in the game,” and “I feel worried about touching the car interior or other passenger when playing.” For “feel safe”, participants had a higher rating in Attach (M=5.7) than Hand (M=4.5). For “feel worried”, Attach (M=2.8) received a lower score than Hand (M=5.8). There were significant differences of the feeling of safety ($Mean_{Diff}=1.2$, $95\%\ CI\ [0.3,2.0].t_{12}=2.96,p=0.01,Cohen^{\prime}s\ d=0.82$) and feeling of worry ($Mean_{Diff}=-3.1$, $95\%\ CI\ [-4.0,-2.2].t_{12}=-7.41$, $p<.001,Cohen^{\prime}s\ d=2.05$). These results show that participants perceived a higher safety while using Attach. Finally, 10 out of 13 preferred to use Attach (FingerMapper) over hand tracking when asked which technique they would like to use for VR applications in confined spaces.

6.6. Discussion

In the confined space, our results show that FingerMapper (Attach) benefited from its fewer physical movements and kept presence and enjoyment comparable to hand tracking. Most participants (10 out of 13) mentioned that reducing collisions was their main concern, and they preferred using FingerMapper because it requires less physical space, so they feel safer. FingerMapper was able to almost completely avoid collisions. In our study, only four participants collided one time with the right door at the beginning of the condition. P5 mentioned, “FingerMapper respects the physical environment around you. So you make small movements and do not have collisions with the physical space. This way, we can be really into the VR experience but [avoid] the experience of colliding with something that’s bringing you back to physical reality.” The subjective rating on perceived safety and preferences also reflected that participants preferred FingerMapper as the interaction for confined spaces.

Participants were aware of crossing over the space of the driver’s seat. They felt that hand tracking is almost impossible to be applied in a confined space. P9 said, “if I touched the door or window, I know it’s not that dangerous because if I touch them, there won’t be any accidents. However, if I know there’s a driver on the left-hand side, Then I will avoid [going] too far with my arms.” P10 also mentioned, “If a ball refreshed at the same position where I know, I probably will touch the interior of the car to grab the ball. If some people are sitting there and I touched those people, I probably will just avoid grabbing the ball in that direction and try to do the other direction.” Participants mentioned that during the interaction they had a mental model of the physical environment they were in, and rather avoided an interaction than collide with the environment.

FingerMapper can overcome the confined space and presents a safe (perceived safety, amount of collisions), ergonomic (less physical motion in Study 1), and enjoyable interaction (enjoyment) while preserving presence and VBO. Overall, participants’ subjective rating showed a strong preference towards using FingerMapper in confined spaces for VR interaction ($M=5.3,SD=0.9$) compared to hand tracking ($M=3.2,SD=1.5$).

7. Lessons Learned

As VR technologies become more mobile, users are able to engage in immersive virtual environments wherever they want. The confined space (e.g., commuting in a passenger seat) would be one of the new contexts where it does not allow for whole-body motion. We can observe smartphone behavior to understand what these confined spaces could look like. People use a smartphone and smartphone games during commutes or in small spaces (e.g., a toilette booth). If VR becomes further distributed, users would also want to engage with immersive applications in such spaces.

Despite having a higher task time and smaller VBO, Study 1 shows that FingerMapper could reduce physical movements and fatigue compared to hand tracking. In the confined space of Study 2, FingerMapper had fewer collisions than hand tracking and could enable a similar level of presence and enjoyment inside a VR experience. Additionally, participants had a significantly higher perceived safety using FingerMapper. Although it took a short training phase to master, participants preferred using FingerMapper for VR interaction in confined spaces. In the following, we discuss insights into FingerMapper gained from the design process and two studies.

Reducing Fatigue in Long Duration VR Experiences with FingerMapper

FingerMapper does not necessarily have to be used inside a confined space. Current VR interactions start to optimize for ergonomics to enable a longer and more comfortable experience (Zielasko and Riecke, 2020; Montano Murillo et al., 2017; Wentzel et al., 2020; Evangelista Belo et al., 2021). Our approach can be an option when the user is tired and wants a lighter experience in terms of physical fatigue. This approach challenges the dominant notion that VR user needs to perform the same actions as they do in real life (e.g., walking long distances). FingerMapper provides users the feeling as if they would perform these interactions without actually performing them.

Context Adaptive Interaction: Confined Space and New Metrics

Through the exploration of FingerMapper, we found that participants started to value perceived safety and avoid collisions while using VR in a confined space. Current VR interaction techniques are frequently optimized around metrics such as presence or performance. FingerMapper demonstrated that participants preferred the interaction technique with lower performance metrics but higher safety. We argue that future research metrics of VR interaction techniques will start to include constructs such as safety, social acceptance, and public perception.

8. Limitations and Future Works

Although FingerMapper can enable virtual arm interaction for VR in a confined space while having less physical movement and fatigue, fewer collisions, and higher perceived safety, there are still some limitations to be improved. First, the displacement between the fingers and virtual arms resulted in lower VBO. One potential solution could be adding haptic feedback at the fingertip (e.g., tactile feedback) when the virtual arm intersects with a target to enhance the overall embodiment. Second, we introduce Attach and Direct, but the mapping function can be designed in many different forms, giving rise to future work exploring different mappings.

The current studies focused on the controlled study and did not evaluate FingerMapper inside a public confined space (e.g., a bus). The reason is that we were initially interested in comparing the individual techniques in a controlled environment to have a fair comparison to regular hand tracking. Future work could conduct further studies in public spaces to understand the social implications of finger mappings. Our studies explored the effect of FingerMapper by observing path length with a small sample size, which is one limitation of the current work. Future research could aim for a large sample size experimental setup to verify the effect of re-associating body parts in both position and rotational movement.

To enable VR interaction in confined spaces, one can either adapt the application to require less physical movement (im-henri, 2022; Melnick, 2020) or the input. FingerMapper focused on adapting the input since it allows for having one implementation (might even be on the system level) that creates a mapping from a confined space interaction toward a whole-body interaction in VR. This allows the designers of VR applications to abstract from the context in which the application would be used and always design for a whole-body experience. However, the challenge of adapting the input is to find an appropriate mapping that has the same DoF while using less physical space and still being usable. This inevitably leads to trade-off decisions and creates a set of limitations for the proposed input adaptation.

While FingerMapper provides a mapping for the basic input mechanisms of whole-body hand-tracking interaction (object selection and manipulation), our trade-off is losing efficiency (higher selection time, less precision). Additionally, the kinematic structure of fingers is different to the structure of arms (e.g., a throwing motion with arms is possible because of the DoF of the connecting joints, while the connecting joints of the finger have a different DoF). This results in some actions (e.g., throwing) being more difficult to perform using FingerMapper. However, we argue that FingerMapper focuses on enabling the basic functions of interacting in VR (object selection and manipulation) and presenting a mapping that can be generally applied to most applications and allow for a safe interaction in confined spaces. Future work should explore how complex actions can be mapped to small-space interaction techniques.

9. Conclusion

FingerMapper enables whole-body VR interaction in VR, allowing users to control arm and hand interactions in confined spaces through small finger motions. Three applications highlight the different functionalities and scenarios, including large slicing motions, object interaction, and locomotion. Two studies show that FingerMapper can reduce physical movements, fatigue, and collisions. Osur techniques reach a similar level of presence and enjoyment in the confined space with higher perceived safety and user preferences. We see FingerMapper as an alternative to hand tracking—if the user interacts in a confined space, and ray-casting—if the user needs more functionality and embodiment of the avatar. Overall, we argue input techniques in VR should be similarly easy to adapt to the context as the tracking options are (e.g., room-scale and stationary tracking). The concept of FingerMapper would work with any application when the environment is not allowed for whole-body motion while preserving the VR user’s experience and safety.

Acknowledgements.

We thank our anonymous reviewers for their suggestions on improving this paper and all participants. This work was partially conducted within the Investments for the Future Program (PIA), CONTINUUM (ANR-21-ESRE-0030), funded by French National Research Agency (ANR) and the Living Lab 5G project financed by the Banque Publique d’Investissement (BPI).