eternpoler/tactile_teleoperation_dataset

Tactile Teleoperation Fruit Grasping Dataset

Dataset of magnetometer (MAG) streams recorded during teleoperated robotic grasping of soft fruits, collected to study tactile signatures of grasp damage.

Data Collection

A robotic gripper was teleoperated to grasp fruits. Two IMU/magnetometer sensors mounted on the two sides of the gripper jaws recorded 3-axis data. Each recorded trial is stored as one CSV file.

Labels (8 classes)

Fruit type × grasp outcome (intact / damaged):

Class folder	Fruit	Outcome
blueberry_intact	Blueberry	intact
blueberry_damaged	Blueberry	damaged
cherry_intact	Cherry	intact
cherry_damaged	Cherry	damaged
grape_intact	Grape	intact
grape_damaged	Grape	damaged
shine_muscat_intact	Shine Muscat (green grape)	intact
shine_muscat_damaged	Shine Muscat (green grape)	damaged

10 trials per class, 80 CSV files in total. The *_allplot folders contain per-trial waveform plots (PNG) for visualization / quality checking (60 PNG files).

CSV Format

Each CSV has the header:

timestamp,sensor_index,data_type,data_type_name,raw_x,raw_y,raw_z,si_x,si_y,si_z

• timestamp — UNIX timestamp (seconds)
• sensor_index — sensor id (0 or 1, the two sensors on the two gripper jaws)
• data_type / data_type_name — sensor data type; use rows where data_type_name == "MAG" (magnetometer)
• raw_x, raw_y, raw_z — raw ADC values
• si_x, si_y, si_z — values in SI units (µT)

Each sensor provides a 3-channel (si_x, si_y, si_z) time series per trial.

Suggested Usage

• Time-series classification of grasp outcome (intact vs. damaged) per fruit type
• 8-class classification across fruit types and outcomes
• Cross-sensor fusion of the dual (sensor_index 0/1) streams