wzaielamri/vibrosense

Vibro-Sense: Robust Vibration-based Impulse Response Localization and Trajectory Tracking for Robotic Hands

Paper page can be found here.

Github Repository can be found here.

Vibrosense.

This repository provides the full dataset for the Vibro-Sense project, enabling robust, affordable, and scalable contact perception for robotic hands using vibro-acoustic sensing. Our approach leverages seven low-cost piezoelectric microphones and an Audio Spectrogram Transformer to decode vibrational signatures generated during physical interaction. The system achieves high-accuracy whole-body touch localization, robust trajectory tracking, and material-aware perception, and is resilient to robot motion.

Note: The dataset is generated and stored using the webdataset format for efficient large-scale data handling and streaming.

Dataset Structure

The dataset is organized into several main directories:

• localisation/: Contains data for touch localization and material detection tasks, including filtered and processed signals for different materials (soft-plastic (cap), metal, plastic, wood).
• quickdraw/: Contains data for trajectory following task, with both fixed and moving hand conditions, and for different materials.
• quickdraw_testset/: Test set for trajectory following, organized similarly to quickdraw/.
• objects/: Contains object classification data collected with three microphones.

Tasks

This dataset supports four main tasks:

1. Touch Localization

• Predict the contact location on the robotic hand using vibrational signals.
• Data: localisation/filtered/ and localisation/processed_* directories.

2. Trajectory Following (Quickdraw)

• Track and reconstruct the trajectory of contact points during dynamic interactions.
• Data: quickdraw/ and quickdraw_testset/ directories.

3. Object Classification

• Classify the object being touched based on the vibrational response.
• Data: objects/dataset/.

4. Material Detection (using Localisation Dataset)

• Identify the material of the contact surface (e.g., metal, plastic, wood, soft-plastic (cap)) using the localisation dataset.
• Data: localisation/filtered/ and localisation/processed_* directories.

Getting Started

All datasets are provided in processed form, ready for training and evaluation. Please refer to the paper and code repository for details on model training and evaluation.

Citation

@InProceedings{ZaiElAmri2026VibroSense,
  author = {Zai El Amri, Wadhah and {Navarro-Guerrero}, Nicol{\'a}s},
  title = {"Vibro-Sense: Robust Vibration-based Impulse Response Localization and Trajectory Tracking for Robotic Hands"},
  booktitle = {"ArXiv Preprint arXiv:2601.20555, Submitted to Autonomous Robots Springer Journal"},
  year={2026},
}