1. Bibliographic Information

1.1. Title

Exploring Perception-Based Techniques for Redirected Walking in VR: A Comprehensive Survey

1.2. Authors

• Bradley Coles (Affiliation typically associated with the University of Central Florida based on co-authors, marked with circle symbol in text)
• Yahya Hmaiti (Affiliation typically associated with the University of Central Florida based on co-authors, marked with circle symbol in text)
• Joseph J. LaViola Jr. (Professor at the University of Central Florida, marked with number 1 in text)

1.3. Journal/Conference

This paper appears to be a preprint, likely submitted to a venue such as IEEE Transactions on Visualization and Computer Graphics (TVCG), given the formatting (Index Terms) and the citation style, or is currently hosted on arXiv. The source metadata indicates it was published (or uploaded) at UTC 2025-05-21.

1.4. Publication Year

2025

1.5. Abstract

This paper presents a comprehensive survey of perception-based Redirected Walking (RDW) techniques in Virtual Reality (VR). The authors review 232 papers (selecting 165 for final analysis) to construct a novel taxonomy. This taxonomy breaks down RDW algorithms into four key components: Gains, Gain Application, Target Orientation Calculation, and Enhancements. A significant contribution is the explicit categorization of "Target Orientation Calculation," a previously overlooked component. The paper analyzes current research trends using this framework and identifies under-explored areas to guide future work.

1.6. Original Source Link

https://arxiv.org/abs/2505.16011v1 (Preprint/Source)

2. Executive Summary

2.1. Background & Motivation

• The Core Problem: One of the greatest challenges in Virtual Reality (VR) is Locomotion. While Virtual Environments (VEs) can be infinite in size (e.g., a massive open world), the Physical Environments (PEs) where users actually stand are limited (e.g., a living room).
• The Solution - Real Walking: "Real walking" (physically moving one's legs to move in VR) provides the highest sense of presence (the feeling of "being there"). However, physical walls limit this.
• Redirected Walking (RDW): This technique solves the space mismatch by subtly manipulating the user's view. For example, when a user walks straight in the real world, the VR system might rotate the world slightly, causing the user to unconsciously curve their path to stay within the physical room bounds, all while thinking they are walking straight in the virtual world.
• The Gap: While many RDW algorithms exist, prior surveys categorized them broadly (e.g., Reactive vs. Predictive). The authors identify a lack of granularity in understanding how these algorithms determine where to guide the user. Specifically, the component of Target Orientation Calculation—deciding the ideal physical direction to steer the user towards—has not been explicitly categorized in previous taxonomies.

2.2. Main Contributions / Findings

• Novel Taxonomy: The paper proposes a component-wise framework for RDW, introducing "Target Orientation Calculation" as a primary classification axis alongside Gains and Gain Application.
• Comprehensive Survey: An analysis of 165 papers from an initial pool of 232, categorized under this new system.
• Key Data Insights:
  • Dominance of Reactive Steering: 54.5% of works use Steering-based orientation calculation, and 63.2% use Reactive application.
  • Identified Research Gaps: There is a notable absence of "Predictive Alignment" techniques (predicting future movement to align physical and virtual features).
• Strategic Guidance: The framework serves as a guide for developers to mix and match components (e.g., combining a specific heading calculation with a specific gain application) to design custom RDW systems.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

To understand this paper, one must grasp the mechanics of Redirected Walking (RDW). RDW relies on the human perceptual system's dominance of vision over the vestibular system (balance/motion).

• Gains: These are the coefficients used to manipulate the mapping between physical and virtual motion.
  • Translation Gain ($g_T$): Modifies the distance traveled. If $g_T > 1$, a small physical step moves the user a large distance in VR.
  • Rotation Gain ($g_R$): Modifies the angle of rotation. If the user turns physically by $90^\circ$ but the VR camera rotates $100^\circ$, the user perceives a $100^\circ$ turn but physically turned less.
  • Curvature Gain ($g_C$): The most famous RDW technique. As the user walks a straight path in VR, the system slowly rotates the world. To keep walking "straight" visually, the user unconsciously turns their body, walking in a curve or circle physically.
  • Bending Gain: Similar to curvature, but applied to paths that are already curved in the virtual world, increasing or decreasing their curvature radius.
• Detection Thresholds (DTs): The limits of these manipulations. If a gain is too high, the user notices the trick (e.g., "The world is spinning") or gets motion sick. RDW aims to keep gains subtle (below DTs).
• Overt vs. Subtle: Subtle RDW is imperceptible. Overt RDW (like a "Reset") forces the user to stop and turn around when they are about to hit a wall.
• Pose: Defined by position (x, y) and orientation $(\theta)$.

3.2. Previous Works

• Razzaque et al. (2001, 2005): Introduced the seminal Steer-to-Center (S2C) algorithm.
  • Concept: Always gently steer the user toward the center of the physical room. This minimizes the chance of hitting walls.
• Steinicke et al. (2008, 2010): Established the standard Detection Thresholds (DTs) for rotation, translation, and curvature gains, providing the mathematical constraints for imperceptible redirection.
• Nilsson et al. (2018) & Li et al. (2022): Previous surveys.
  • Limitation: These largely categorized algorithms as Reactive (react to current state), Predictive (guess future path), or Scripted (pre-planned). This paper argues this is insufficient because two "Reactive" algorithms might use completely different logics to pick a safe direction (e.g., one steers to the center, one uses repulsive forces from walls).

3.3. Technological Evolution

1. Early Era: Simple Steering (S2C, Steer-to-Orbit) using reactive gains.
2. Middle Era: Introduction of Artificial Potential Fields (APF) which treat walls as repelling magnets to push users away.
3. Modern Era: Alignment techniques (matching physical objects to virtual ones for haptics) and Machine Learning/Reinforcement Learning approaches that learn optimal redirection policies.

4. Methodology

The authors employed a systematic literature review methodology following PRISMA guidelines to construct their corpus, and then derived a new taxonomy to analyze it.

4.1. Survey Methodology

• Sources: ACM Digital Library, IEEE Xplore, Springer.
• Keywords: ("Redirected Walking" OR "RDW") AND ("VR" OR "MR" OR "AR" OR "XR" etc.).
• Filtering:
  • Initial pool: 1,824 papers.
  • Inclusion Criteria: Focus on RDW or Locomotion Gain Detection.
  • Exclusion Criteria: Non-English, posters/short papers, non-human locomotion, or Environment-based RDW (techniques that physically change the virtual world layout, like impossible spaces, were excluded to focus purely on perception manipulation).
  • Final Corpus: 165 papers.

4.2. The Taxonomy Framework

The core contribution is a modular framework for RDW algorithms. The authors argue that any perception-based RDW system is composed of four distinct components.

The following figure (Figure 1 from the original paper) illustrates this taxonomy:

该图像是一个示意图，展示了感知基础的重新行走（RDW）技术的分类框架。图中包括了对增益、目标方向计算、增益应用和一般增强等关键要素的分类，帮助理解RDW算法的设计与应用。

4.2.1. Component 1: Gains

This component defines what kind of manipulation is applied.

• Subtle Gains: Applied continuously below detection thresholds. Includes Rotation, Curvature, Translation, Bending, and newer types like Strafing (sideways) or Jumping gains.
• Overt Gains (Resets): Applied when subtle gains fail and a collision is imminent. The user is explicitly told to stop and turn (e.g., a "2:1 turn" where a $180^\circ$ physical turn creates a $360^\circ$ virtual turn).

4.2.2. Component 2: Target Orientation Calculation (New Contribution)

This is the "brain" of the redirection. Before applying a gain, the system must decide "Which direction in the physical room is safe or optimal?" The authors classify this into three main types:

1. Steering: Directs the user toward a specific target in the Physical Environment (PE).
   • Steer-to-Point: E.g., Center of the room (S2C).
   • Steer-to-Path: E.g., An orbit around the center (Steer-to-Orbit, S2O).
2. Avoidance: Uses repulsive forces to push users away from boundaries, rather than pulling them to a single point.
   • Artificial Potential Fields (APF): Walls and obstacles emit "repulsive forces." The user is steered along the vector sum of these forces (the path of least resistance).
   • Non-APF: Optimization algorithms (e.g., Model Predictive Control) that calculate a path with the lowest "cost" (risk of collision).
3. Alignment: Aims to align the user's state in the PE with a state in the VE to enable Passive Haptics.

   • Concept: If there is a physical chair in the room and a virtual chair in the game, Alignment redirection steers the user so that when they arrive at the virtual chair, they physically arrive at the real chair, allowing them to sit down.

     The following figure (Figure 2 from the original paper) visualizes these concepts: (a) Steering to a target, (b) Steering to an orbit path, (c) Avoiding multiple obstacles (Avoidance), (d) is not explicitly described in the caption but conceptually relates to path planning.

     该图像是示意图，展示了感知基础的重定向行走技术的不同情景，包括目标获取和位置调整的示意。四个子图中，(a) 展示了目标引导，(b) 显示了沿圆周行走的路径，(c) 说明了在众多目标中如何进行导航，而(d) 则展示了路径的选择与障碍物的避免。

The following figure (Figure 3 from the original paper) further illustrates alignment: matching the user's path in the Virtual Environment (left) to a specific optimal path or target in the Physical Environment (middle) to achieve alignment (right).

该图像是示意图，展示了虚拟环境与物理环境之间的关系以及重叠对齐的表现。左侧为虚拟环境，标示着用户的位置和路径；中间为物理环境，表示用户实地走过的区域；右侧展示了两者的对齐情况，便于理解如何在有限空间内实现虚拟行走。

4.2.3. Component 3: Gain Application

This component defines how the system decides the magnitude of the gain to reach the Target Orientation.

1. Reactive: Only looks at the current frame (user's current position and heading). Simple and robust but short-sighted.
2. Predictive: Predicts the user's future path (e.g., "User is likely to walk to the door").
   • Graph Prediction: Simplifies the world into a skeleton graph (hallways/nodes).
   • Heuristic: Uses gaze direction or head motion trends.
   • Reinforcement Learning (RL): Trains a neural network to predict movement.
3. Scripted: Uses pre-defined paths. Requires the user to follow a specific narrative or tour, allowing perfectly pre-calculated redirection (since the future is known).

4.2.4. Component 4: Enhancements

Optional modules to improve performance:

• Multi-User: Preventing collisions between two real users in the same physical space.
• Gain Masking: Using "Distractors" (e.g., a flying drone) to make the user turn their head, allowing for larger, unnoticeable rotation gains.
• Saccadic Redirection: Applying rapid shifts during eye blinks or saccades (fast eye movements) when vision is temporarily suppressed.

5. Experimental Setup

Since this is a survey paper, the "Experiment" is the statistical analysis of the literature corpus.

5.1. Datasets

• Corpus: 165 academic papers on perception-based RDW published up to approximately 2024/2025.
• Data Extraction: Each paper was manually reviewed and tagged according to the four components of the taxonomy (Gains, Target Orientation, Gain Application, Enhancements).

5.2. Evaluation Metrics

The authors evaluate the state of the field using Saturation Analysis:

• Percentage of Works: The proportion of papers falling into each taxonomy category (e.g., $\frac{\text{Count of Steering Papers}}{\text{Total Papers}} \times 100$).
• Cross-Tabulation: Analyzing the intersection of categories (e.g., How many papers are both Predictive and Alignment-based?).

5.3. Key Reference Data (Thresholds)

The paper aggregates Gain Detection Thresholds (DTs) from literature. These are crucial baselines for designing any RDW system.

• Interpretation: A threshold of $r > 16m$ for curvature means if the curve radius is smaller than 16 meters (i.e., the curve is tighter), the user will notice it.

• Rotation Gain: Typically ranges around $0.67 - 1.24$. A gain of 1.24 means the VE can rotate 24% faster than the user without detection.

  The following are the results from Table I of the original paper, summarizing these thresholds:

  Gain Type	Comment	Thresholds	Source
  Curvature		$r > 16m$	Steinicke et al. (2008)
  	2m path normal motion before redirection	$r > 24m$	Steinicke et al. (2008)
  		$r > 22m$	Steinicke et al. (2010)
  	Constant Stimuli	`r > 6.41m`	Grechkin et al. (2016)
  	maximum likelihood	`r > 11.61m`	Grechkin et al (2016)
  	threshold for men	`r > 10.7m`	Nguyen et al. (2018)
  	threshold for women	`r > 8.63m`	Nguyen et al. (2018)
  	tested left and right curves separately	`r_{right} > 10.20m`, `r_{left} > 7.81m`	Li et al. (2021)
  	post-order path	`r_{right} > 12.821m`, `r_{left} > 10.309m`	Li et al. (2021)
  	left side-step	`r_{right} > 6.02m`, `r_{left} > 13.19m`	Cho, Min, Huh, et al. (2021)
  	right side-step	`T_{right} > 9.92m`, `r_{left} > 4.65m`	Cho, Min, Huh, et al. (2021)
  	similar sensory sensitivity to average	$g_C = 0.048$	Matsumoto and Narumi (2022)
  	higher sensory sensitivity than average	$g_C = 0.037$	Matsumoto and Narumi (2022)
  	similar sensory avoidance to average	$g_C = 0.048$	Matsumoto and Narumi (2022)
  	much higher sensory avoidance than average	$g_C = 0.033$	Matsumoto and Narumi (2022)
  	control (no device stimuli)	$g_C = -0.055 \text{ to } 0.054$	Hwang et al. (2023)
  	Noisy Galvanic Vestibular Stimulation	$g_C = -0.055 \text{ to } 0.061$	Hwang et al. (2023)
  	Directional Galvanic Vestibular Stimulation	$g_C = -0.068 \text{ to } 0.7$	Hwang et al. (2023)
  	Bone-conduction Vibration	$g_C = -0.061 \text{ to } 0.07$	Hwang et al. (2023)
  Rotation	Caloric Vestibular Stimulation	$g_C = -0.061 \text{ to } 0.068$	Hwang et al. (2023)
  		$0.59 - 1.10$	Steinicke et al. (2008)
  	Subsequent rotations	$0.76 - 1.19$	Steinicke et al. (2008)
  		$0.64 - 1.24$	Steinicke et al. (2010)
  	360-degree video teleoperation	$0.892 - 1.054$	Zhang et al. (2018)
  	FOV 40 all participants w/out distractor	$0.68 - 1.22$	Williams and Peck (2019)
  	FOV 110 all participants w/out distractor	$0.67 - 1.44$	Williams and Peck (2019)
  	FOV 110 women only w/out distractor	$0.65 - 1.32$	Williams and Peck (2019)
  	FOV 110 men only w/out distractor	$0.70 - 1.56$	Williams and Peck (2019)
  	FOV 110 all participants w/ distractor	$0.57 - 1.57$	Williams and Peck (2019)
  Translation	Jump rotation	$0.50 - 1.44$	Hayashi, Fujita, Takashima, et al. (2019)
  	BiRD - oscillating head rotations	$0.84 - 1.24$	Xu, Chen, Gong, et al. (2024)
  		$0.78 - 1.22$	Steinicke et al. (2008)
  		$0.86 - 1.26$	Steinicke et al. (2010)
  	360-degree video teleoperation	$0.942 - 1.097$	Zhang et al. (2018)
  	Jump distance translation	$0.68 - 1.44$	Hayashi, Fujita, Takashima, et al. (2019)
  	Jump height translation	$0.09 - 2.16$	Hayashi, Fujita, Takashima, et al. (2019)
  	Backward step	$0.84 - 1.33$	Cho, Min, Huh, et al. (2021)
  Bending	radius of real curve: `r_{real} = 1.25m`	`3.25`	Langbehn et al. (2017)
  	radius of real curve: `r_{real} = 2.5m`	`4.35`	Langbehn et al. (2017)
  Slope Curvature	real = 2%	$1.56 - 5.25$	Hu et al. (2019)
  	real slope = 4%	$1.98 - 4.28$	Hu et al. (2019)
  	real slope = -2%	$1.00 - 7.36$	Hu et al. (2019)
  	real slope = -4%	$1.22 - 3.44$	Hu et al. (2019)
  Strafing		$5.57^\circ$ right diagonal, $4.68^\circ$ left diagonal	You, Benda, Rosenberg, et al. (2022)
  Deviation		`1.74-5.60` rad/m	Mayor, Raya, Bayona, et al. (2022)

6. Results & Analysis

6.1. Analysis of Target Orientation Calculation

The paper provides a breakdown of how existing research utilizes different methods to calculate target orientation.

• Dominance of Steering: 54.55% of papers use Steering methods (S2C, S2O). This is the oldest and simplest approach, explaining its prevalence. It performs well in simple, open spaces but struggles with obstacles.

• Rise of Avoidance: 29.87% use Avoidance (e.g., APF). This gained traction around 2019, offering better handling of complex rooms by treating walls as repulsive forces.

• Emergence of Alignment: 15.58% use Alignment. This is a newer trend focusing on integrating physical objects (passive haptics) into VR.

  The following are the results from Table II of the original paper:

  Calculation Type	Sources	Percentage of Work
  <strong>Steering</strong><br> Steer-to-Point(s)<br> Steer-to-Path	[16], [103], [104], [105], [106], [107], [108], [109], [110], [111], [77], [98], [11], [112], [39], [93], [8], [89], [113], [114], [115], [96], [116], [117] [104], [118], [119], [120], [109], [121], [122], [123], [124], [125], [126], [49], [127], [128], [8], [7], [129], [130], [131], [132], [133]	54.55%
  Avoidance APF Non-APF	[9], [107], [108], [134], [135], [136], [96] [16], [137], [138], [139], [140], [141], [95], [142], [143], [144], [102], [83], [145], [146], [12]	29.87%
  Alignment	[138], [147], [148], [10], [149], [150], [151], [152], [153], [154], [155], [24]	15.58%

  6.2. Analysis of Gain Application

  • Reactive Dominance: 63.16% of methods are Reactive. These require no pre-processing or environmental knowledge, making them "plug-and-play" for general developers.

  • Predictive & Scripted Tie: Both hold 18.42%. Predictive methods are powerful but complex (often requiring Machine Learning). Scripted methods are simple but brittle (fail if the user deviates).

    The following are the results from Table III of the original paper:

    Calculation Type	Sources	Percentage of Work
    Reactive	[103], [16], [9], [104], [137], [106], [107], [138], [118], [105], [108], [135], [134], [109], [121], [110], [111], [77], [123], [140], [141], [95], [142], [136], [93], [49], [8], [89], [113], [114], [115], [130], [83], [148], [96], [10], [150], [151], [152], [153], [154], [155], [116], [132], [145], [146], [133], [24]	63.16%
    <strong>Predictive</strong><br> Graph Prediction<br> Heuristic Prediction<br> Reinforcement Learning	[122] [16], [139], [120], [102], [156], [12] [97], [39], [117] [157], [11], [112], [143]	18.42%
    Scripted Pre-Calculated Path Environmental Mapping	[16], [144], [127], [128], [7] [119], [147], [98], [124], [125], [126], [129], [131], [149]	18.42%

    The following figure (Figure 4 from the original paper) summarizes these distributions visually:

    该图像是一个饼图，展示了视知觉重定向行走 (RDW) 技术的不同组合类型及其在目标方向中的分布。图中显示了不同技术在分析中的比例，强调了各类方法的应用情况。

    6.3. Research Gaps & Meta-Analysis

    By cross-referencing these tables (as discussed in the text), the authors identify a critical gap:

    • The "Predictive Alignment" Gap: There are zero techniques classified as using Alignment Target Orientation with Predictive Gain Application. Current alignment methods are almost exclusively reactive.
      • Significance: Passive haptics (touching real objects) requires high precision. A predictive algorithm that knows the user is walking toward a virtual table could start aligning them to the physical table earlier and more smoothly than a reactive one, potentially improving the success rate of haptic feedback.
    • Scripted Steering: Most scripted methods use steering logic ($13.6\%$) because static planning aligns well with predefined paths.

    7. Conclusion & Reflections

    7.1. Conclusion Summary

    This paper successfully establishes a new, rigorous taxonomy for perception-based Redirected Walking. By isolating Target Orientation Calculation as a distinct module, it clarifies the internal logic of RDW algorithms. The survey of 165 papers reveals that while simple Reactive Steering (like S2C) remains the industry workhorse, the field is moving toward Avoidance (APF) and Alignment techniques to handle complex environments and haptics.

    7.2. Limitations & Future Work

    Authors' Identified Limitations:

    • Scope: The survey excluded Environment-based techniques (e.g., impossible spaces/non-Euclidean geometry), which might skew the representation of Scripted methods.
    • Keyword Sensitivity: Different search terms might have revealed missed papers.
    • Gain Classification: They attempted to classify papers by "Gain Type" but found the data too messy/inconsistent to be useful in tabular form.

    Future Directions:

    • Predictive Alignment: Developing algorithms that predict user intent to better align physical and virtual objects.
    • Cognitive Load: Evaluating RDW not just on "did they hit the wall?" but "did it make them mentally tired?"
    • Multi-User Fairness: Ensuring that in a remote multiplayer game, one user isn't disadvantaged by having to reset more often due to a smaller room.

    7.3. Personal Insights & Critique

    • Value of the Taxonomy: The separation of "Where to go" (Target Orientation) from "How to get there" (Gain Application) is a significant conceptual upgrade. In previous conceptualizations, an "APF Algorithm" was often seen as a monolithic block. This framework allows us to imagine an "APF-Predictive" hybrid or an "APF-Scripted" hybrid, fostering innovation by mixing components.
    • Practical Utility: Table I (Thresholds) is an invaluable "cheat sheet" for any VR developer. It consolidates scattered psychophysical data into a single lookup table.
    • Critique: The paper relies heavily on the "number of papers" as a metric for importance. While useful for trends, quantity does not equal quality. A saturation of Reactive Steering might simply mean it's the "Hello World" of RDW research, not necessarily the best solution for modern VR. The lack of standardized open-source benchmarks (mentioned briefly in toolkits) remains a hurdle for the field to objectively compare these 165 methods.