🤖 machine learning

Vibrotactile Preference Learning: Uncertainty-Aware Preference Learning for Personalized Vibration Feedback

This paper introduces Vibrotactile Preference Learning (VPL), an uncertainty-aware system that efficiently personalizes haptic feedback for individual users by leveraging Gaussian processes and an information-gain-based query strategy to learn preference spaces through a low-workload pairwise comparison study.

Original authors: Rongtao Zhang, Xin Zhu, Masoume Pourebadi Khotbehsara, Warren Dao, Erdem Bıyık, Heather Culbertson

Published 2026-04-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rongtao Zhang, Xin Zhu, Masoume Pourebadi Khotbehsara, Warren Dao, Erdem Bıyık, Heather Culbertson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a new video game controller, but the vibration settings feel "off." Maybe the rumble is too weak to feel, or maybe it's so intense it hurts your hand. You know what you like, but you don't know the technical terms (like "frequency" or "amplitude") to tell the computer how to fix it.

This paper introduces a smart system called VPL (Vibrotactile Preference Learning) that acts like a personal sommelier for vibrations. Instead of asking you to fill out a complex engineering form, it asks you simple questions to figure out exactly what your hand likes.

Here is how it works, broken down into simple concepts:

1. The Problem: "I know what I like, but I can't explain it."

Usually, if you want to customize your phone's vibration, you have to slide a bar up and down to change the "intensity" or "speed." This is hard because:

It's exhausting: Trying to rate 50 different vibrations on a scale of 1 to 10 gets tiring, and your rating might change halfway through (you might think a vibration is a "5" at the start, but a "3" after you've felt 20 others).
It's confusing: You don't know what "40Hz rhythm" actually feels like until you try it.

2. The Solution: The "Blind Taste Test" Approach

Instead of asking you to rate vibrations, the VPL system plays a game of "This or That."

The Setup: The system plays two different vibrations (Signal A and Signal B) on your controller.
The Question: "Which one feels better?"
The Secret Sauce: After you pick one, it also asks, "How sure are you?" (On a scale of 1 to 5).
- If you say "I'm totally sure," the system trusts your answer 100%.
- If you say "I'm not sure, they felt similar," the system knows, "Okay, this is a tricky spot. I need to be more careful with my next guess."

3. The Magic Brain: The "Uncertainty Map"

The computer uses a smart algorithm (called a Gaussian Process) that acts like a cartographer drawing a map of your preferences.

The Map: Imagine a landscape where the "mountains" are vibrations you love, and the "valleys" are vibrations you hate.
The Fog: At the beginning, the map is covered in thick fog. The computer doesn't know where your favorite spot is.
The Strategy: The system uses a strategy called "Information Gain." It doesn't just ask random questions. It asks the most useful questions to clear the fog the fastest.
- If the map is very foggy in one area, it sends two very different vibrations to that area to see where the "mountain" is.
- If it thinks it found your favorite spot, it asks very similar vibrations to confirm the peak.

4. The Result: A Perfect Fit in 40 Questions

In the study, the researchers asked 13 people to play this game for 40 rounds (about 12 minutes).

The Outcome: By the end, the system had drawn a very accurate map of what each person liked.
The Proof: When the system recommended a vibration, it matched what the users would have picked if they had spent hours manually tweaking sliders.
The Feeling: Users reported that the process was easy, not tiring, and they were very happy with the final result.

Why This Matters (The Big Picture)

Think of this like tailoring a suit.

Old Way: You try on 50 suits and rate them. It takes all day, and you get tired.
VPL Way: The tailor asks, "Do you prefer the blue or the red?" "Are you sure?" "Do you prefer the slim fit or the loose fit?" After 40 quick questions, the tailor has a perfect pattern for your body.

This technology means that in the future:

Video Games can automatically tune the controller rumble to your specific sensitivity.
Smartwatches can find the perfect vibration pattern for notifications that you can feel but that won't annoy you.
Therapy devices can customize haptic feedback for pain management or rehabilitation without needing a doctor to guess the settings.

In short: VPL is a smart, polite assistant that learns exactly how your skin likes to be touched, using a simple "A or B" game and paying attention to how confident you feel about your choices.

1. Problem Statement

Haptic feedback is crucial for immersion in virtual and remote environments, but its effectiveness is hindered by high individual variability in tactile perception. Factors such as mechanoreceptor density, skin thickness, and psychological preferences mean that a "gentle" signal for one user may be imperceptible or uncomfortable for another.

Current personalization methods suffer from significant limitations:

Manual Tuning (Sliders): Requires users to understand complex parameters (frequency, amplitude) and navigate high-dimensional spaces, imposing high cognitive load.
Absolute Ratings (Likert Scales): Prone to score drift (lack of stable internal reference) and user fatigue during prolonged sessions, leading to noisy data.
Binary Comparisons (A/B Testing): While more robust than ratings, standard pairwise comparisons generate sparse data. Without an intelligent selection strategy, they require too many interactions to converge on a user's true preference.

The core challenge is to efficiently learn an individual's latent vibrotactile preference structure within a limited interaction budget (minimizing user effort) while accounting for the uncertainty inherent in human sensory judgments.

2. Methodology: Vibrotactile Preference Learning (VPL)

The authors propose VPL, an interactive system that combines Gaussian Process (GP) preference learning with uncertainty-aware query selection.

A. Signal Parameterization

The system uses a Microsoft Xbox controller with two asymmetric motors (low-frequency left, high-frequency right). Vibrotactile signals are defined by a 4-dimensional parameter space $\Phi$ :

Intensity ( $I$ ): Overall magnitude.
Motor Balance ( $T$ ): Distribution of intensity between the two motors (controlling "texture").
Rhythm ( $R$ ): Frequency of vibration pulses.
Grain ( $G$ ): Duty cycle of pulses (affecting sharpness).
Signals are rendered as 3-second sequences with hardware constraints on pulse duration and gaps.

B. Probabilistic Modeling

The system models user preference as a latent utility function $f: \mathcal{X} \to \mathbb{R}$ using a Gaussian Process (GP) prior.

Input: Normalized parameter vectors $x \in [0, 1]^4$ .
Likelihood: The system uses a confidence-aware probit likelihood. In each round $t$ $t$ , the user compares two signals ( $x_A, x_B$ $x_{A}, x_{B}$ ) and provides:
1. A binary preference $y_t \in \{+1, -1\}$ .
2. A confidence score $c_t \in \{1, \dots, 5\}$ .
Uncertainty Mapping: The confidence score is mapped to a noise scale $\sigma_t$ . Lower confidence results in higher noise in the likelihood function, effectively down-weighting uncertain judgments during model updates.
Posterior Estimation: The system estimates the posterior mode via Maximum A Posteriori (MAP) estimation with a Laplace approximation to obtain the mean ( $\mu$ ) and covariance ( $\Sigma$ ) of the utility function.

C. Active Query Selection

To maximize data efficiency, the system employs an Expected Information Gain (IG) acquisition strategy.

Instead of random sampling or simple exploitation, the system selects the pair $(x_A, x_B)$ that maximizes the expected reduction in uncertainty about the global preference landscape.
This allows the system to explore regions where the model is uncertain or where user judgments are inconsistent, rather than just refining a local optimum immediately.

3. Key Contributions

Confidence-Aware Preference Model: A novel GP-based framework that integrates user-reported confidence scores to modulate the trust in each pairwise comparison, treating the user as a "variable-noise oracle."
Efficient Active Learning Strategy: The use of Information Gain for query selection enables the system to learn a global preference structure within a strict budget (40 comparisons), balancing exploration and refinement.
Empirical Validation: A user study ( $N=13$ ) demonstrating that VPL achieves high predictive accuracy and user satisfaction with low cognitive workload.
Design Insights: Analysis of sampling behaviors (InfoGain vs. EUBO) and practical implications for deploying preference-based personalization in haptics.

4. Experimental Results

The system was evaluated in a user study using an Xbox controller over 40 rounds of pairwise comparisons.

A. Objective Metrics (RQ1: Accuracy)

Validation Accuracy: The learned model achieved a 92.3% mean accuracy on held-out validation pairs (range: 85%–100%).
Consistency with User Optima: Signals manually tuned by users as their "best" (in a free-adjustment phase) were ranked by the VPL model in the top 18.8% (81.2th percentile) of the utility space.
Convergence: Information Gain declined steadily, and Spearman correlation with the ground truth stabilized, indicating effective learning.

B. Subjective Metrics (RQ2 & RQ3: Satisfaction & Workload)

Satisfaction: Users reported high satisfaction with the recommended signals (Mean $\approx$ 6.5/7). The "Expectation Match" score was high (6.23/7), indicating the algorithm aligned well with users' mental models.
Workload: Using the NASA-TLX scale, the overall workload was low (Mean = 3.71/20).
- Mental Demand was the highest component (5.15), reflecting the cognitive effort of distinguishing subtle haptics.
- Physical Demand and Frustration were very low.
- Performance scores were low (2.23), which on the NASA-TLX scale indicates high self-perceived success.
Time: The entire 40-round session took approximately 12 minutes and 43 seconds.

5. Significance and Future Directions

Significance:

Scalable Personalization: VPL demonstrates that high-quality personalization is possible without requiring users to be experts or endure long, fatiguing sessions.
Robustness to Noise: By explicitly modeling user uncertainty, the system avoids overfitting to noisy or inconsistent judgments, a common failure point in standard preference learning.
Application Domains: The framework is applicable to diverse fields including Automotive HMI (safety warnings), Therapy/Rehabilitation (pain management), Mobile/Wearables (discreet alerts), and Immersive Gaming (VR/AR).

Limitations & Future Work:

Budget Constraints: With only 40 interactions, the learned landscape is a low-resolution approximation; very similar signals may remain indistinguishable.
Parameterization: The current 4-parameter model excludes complex spatiotemporal patterns. Future work aims to extend VPL to higher-dimensional spaces.
Confidence Calibration: The mapping from confidence scores to noise levels is currently fixed. Future iterations could adapt this mapping online to better suit individual user behaviors.

In conclusion, VPL offers a robust, data-efficient, and user-friendly approach to personalizing vibrotactile feedback, bridging the gap between complex haptic design spaces and individual user needs.