wzaielamri/vibrosense

Vibrosense数据集是一个用于机器人手振动感知研究的完整数据集，旨在实现鲁棒、经济高效且可扩展的接触感知。该数据集利用七个低成本压电麦克风和音频频谱变换器，解码物理交互过程中产生的振动信号，支持高精度的全身触摸定位、鲁棒的轨迹跟踪和材料感知，并对机器人运动具有适应性。数据集结构包括localisation/（用于触摸定位和材料检测任务，包含不同材料如软塑料、金属、塑料、木材的过滤和处理信号）、quickdraw/和quickdraw_testset/（用于轨迹跟踪任务，包括固定和移动手条件以及不同材料的数据）以及objects/（用于物体分类任务，包含三个麦克风收集的数据）。该数据集支持四个主要任务：触摸定位（预测机器人手上的接触位置）、轨迹跟踪（动态交互中接触点的轨迹重建）、物体分类（基于振动响应分类被触摸的物体）和材料检测（使用定位数据集识别接触表面的材料，如金属、塑料、木材、软塑料）。数据集以webdataset格式存储，便于高效的大规模数据处理和流式传输。Vibrosense Dataset is a full dataset for the Vibro-Sense project, enabling robust, affordable, and scalable contact perception for robotic hands using vibro-acoustic sensing. The approach leverages seven low-cost piezoelectric microphones and an Audio Spectrogram Transformer to decode vibrational signatures generated during physical interaction, achieving high-accuracy whole-body touch localization, robust trajectory tracking, and material-aware perception, with resilience to robot motion. The dataset is organized into directories: localisation/ (for touch localization and material detection tasks, including filtered and processed signals for materials like soft-plastic, metal, plastic, wood), quickdraw/ and quickdraw_testset/ (for trajectory following tasks, with fixed and moving hand conditions and different materials), and objects/ (for object classification data collected with three microphones). It supports four main tasks: Touch Localization (predicting contact location on the robotic hand), Trajectory Following (tracking and reconstructing contact point trajectories during dynamic interactions), Object Classification (classifying the object being touched based on vibrational response), and Material Detection (identifying contact surface materials using the localisation dataset). The dataset is stored in webdataset format for efficient large-scale data handling and streaming.