OpenRIF: Open Robot Interaction Force Dataset

Description

OpenRIF (Open Robot Interaction Force Dataset) is an open-source dataset designed to support large-scale learning for sensorless robot interaction force estimation and robotic haptics research.

The dataset contains approximately 15 hours of real-world and simulated robot interactions, comprising over 8.6 million time steps collected across multiple objects, tasks, spatial configurations, and robot embodiments.

Repository Structure

OpenRIF/
├── asset/
│   └── Videos and images for task illustration
├── data/
│   └── Raw data collected from each task
└── example/
    └── Python example scripts for reading and visualizing the data

Note: The videos and images in the asset/ folder are provided for task illustration only and are not part of the training data.

Data Contents

The basic data include real-time joint angles, joint velocities, joint currents, and joint torques from the robot, together with the corresponding interaction force/torque measurements from F/T sensors and timestamps.

Additional information may also be recorded depending on the task. Details of the experimental and simulation data formats are provided below.

Experimental Data

Experimental data are stored in MATLAB .mat files.

The general file naming convention is:

<Object>_Robot_<TaskIndex>.mat
<Object>_Sensor_<TaskIndex>.mat
<Object>_Task_Time_<TaskIndex>.mat

where <Object> identifies the interaction object and <TaskIndex> identifies the task recording.

Robot State Data

<Object>_Robot_<TaskIndex>.mat

Contains robot state measurements recorded during the experiment.

Example:

===== BP_Robot_A.mat =====
angle:     shape=(69109, 7), dtype=float64
velocity:  shape=(69109, 7), dtype=float64
current:   shape=(69109, 7), dtype=float64
torque:    shape=(69109, 7), dtype=float64
localtime: shape=(1, 69109), dtype=int64

Force/Torque Sensor Data

<Object>_Sensor_<TaskIndex>.mat

Contains force/torque sensor measurements and the corresponding timestamps.

Example:

===== BP_Sensor_A.mat =====
Fx:      shape=(1, 27561), dtype=float64
Fy:      shape=(1, 27561), dtype=float64
Fz:      shape=(1, 27561), dtype=float64
Tx:      shape=(1, 27561), dtype=float64
Ty:      shape=(1, 27561), dtype=float64
Tz:      shape=(1, 27561), dtype=float64
Time:    shape=(1, 27561), dtype=int64
ZeroRef: shape=(1, 6), dtype=float64

ZeroRef contains the zero-reference values of the F/T sensor for the corresponding recording.

Task Timing Data

<Object>_Task_Time_<TaskIndex>.mat

Contains the start and end timestamps of the task. Note that data recording may begin before the actual task starts.

Example:

===== BP_Task_Time_A.mat =====
task_start_time: shape=(1, 1), dtype=int64
task_end_time:   shape=(1, 1), dtype=int64
random_params:   shape=(10, 6), dtype=float64

random_params records the randomized trajectory parameters used for each task.

For more details about the variables and their dimensions, please refer to the Python scripts provided in the example/ folder.

Simulation Data

Simulation data are stored as NumPy .npy files using the following naming convention:

simulation_<Robot>_<Object>_<TaskIndex>.npy

Each file has the shape:

(batch_size, 22)

The 22 columns are organized as follows:

Columns	Data
1–6	Joint positions
7–12	Joint velocities
13–18	Joint torques
19–21	Interaction forces
22	Timestamp

Each row represents one simulation time step containing the robot state, interaction force, and corresponding timestamp.

Examples

Python example scripts for loading, inspecting, and visualizing the dataset are provided in the example/ folder.

License

OpenRIF is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.