PUB - Publikationen an der Universität Bielefeld

# Tactile Glove In-Hand-Manipulation Dataset *Detailed pictorial overview in `Info-InHandOrientation.pdf`.* - This dataset aims to harness the power of machine learning to decode in-hand object rotations based on human demonstrations. Initially, our emphasis is on deepening the machine's understanding by classifying haptic cues on an object such as contacting, moving in specific directions, and guiding for each finger and the palm. Ultimately, the intention is to deduce also the current orientation of an object. The long-term goal is to apply this knowledge, enabling a robotic hand to replicate in-hand rotations of analogous objects — a prime example of skill transfer in action. ## CONTENTS: 1. Dataset Summary 2. Experiment Details 3. Equipment Specs 4. Data Playback ### Objectives: - Classify six rotation directions for cube and rubber-duck. - Predict 3D angle of rotation, object orientation, and haptic cues. - Forecast sensory cues and finger movements based on 3D rotation. --- ## Dataset Summary ``` | Modality | Specs | Rate | | ---------------------------- | ------------------------------------ | -------- | | Tactile (TactileGlove) | 60 sensors (of 64 total), 12 bit | 1000 Hz | | Posture (Joint_states) | 15 sensors, 12 bit | 1000 Hz | | IMU Orientation (BNO08x) | 9 dimensions | 100 Hz | | Camera Feed (usb_cam) | 640x480 pixels, RGB, 8 bit compressed| 30 Hz | ``` ``` | In-Hand rotations | Repetitions | Duration | | ------------------------- | ------------------- | --------- | | Cube (6 directions) | 50 reps/direction | ~10 sec | | Rubber-Duck (6 directions)| 30 reps/direction | ~10 sec | | Free Rotation | 6 reps each object | Varies | ``` Datasets saved as ROS bag files, organized by date/time. Rotations have separate subfolders. --- ## Experiment Details For each rotation: 1. Grasp top, roll hand to expose palm (~180°). 2. Rotate object in one pre-defined direction (~10 revs). 3. Return hand to start and place object down. - **Cube**: 50 reps/direction, ~10 sec each. Link: [Cube-Rotation](https://pub.uni-bielefeld.de/download/2966879/2967344.zip) - **Rubber-Duck**: 30 reps/direction, ~10 sec each. Link: [Rubber-Duck-Rotation](https://pub.uni-bielefeld.de/download/2966879/2967345.zip) - **Free Rotation**: 6 reps each for cube/rubber-duck, varied directions. Link: [Free-Rotation](https://pub.uni-bielefeld.de/download/2966879/2967345.zip) --- ## Equipment Specs ### Cube - Hand-sized, 3D-printed with rounded edges. - 51mm³, Weight: 137g, ABS, 24 ArUco markers (17.25mm each). ### Rubber-Duck - Hand-sized, flexible plastic. - 100mm length, 70mm width & height, Weight: 46g. ### Camera: Logitech HD webcam C615 - 30 Hz framerate. - Calibration details [here](http://wiki.ros.org/camera_calibration). ### µGlove - Captures hand movements/tactile contacts. - Tactile: 60 cells at 1000 Hz. - Posture: 15 sensors at 1000 Hz. - Orientation: 9 DoF IMU ~100 Hz. --- ## Data Playback For 3D visualization, install ROS and AGNI-Tools. Instructions [here](https://github.com/ubi-agni/TactileGlove). Help: [gbuescher@techfak.uni-bielefeld.de](mailto:gbuescher@techfak.uni-bielefeld.de) or [buescher@robomantic.de](mailto:buescher@robomantic.de).