TouchEvent

1. Overview

The TouchEvent dataset belongs to Haptic Sensing Lab at the School of Advanced Manufacturing and Robotics, Peking University. TouchEvent contains synchronized time-series data from real-robot tactile experiments and finite element method (FEM/IPC) simulations. It covers objects with different shapes, materials, and sizes and is organized into four task families:

• Fixed: pressing, lifting, and slipping at a fixed contact position;
• Free: picking and insertion at different gripper widths;
• Obstacle_avoidance: obstacle-avoidance experiments at different contact positions and depths;
• Integrated_tasks: multi-stage experiments involving a paired ball and cylinder.

Depending on the task, the data include robot end-effector poses, gripper states, force/torque or tactile-channel measurements, simulated motion, FEM node displacements, and nodal contact forces. Vision-based tactile branches also provide camera images associated with the CSV records.

Dataset page: https://huggingface.co/datasets/BAAI/TouchEvent

2. Top-Level Structure

TouchEvent/
├── Fixed/
│   ├── digit/
│   ├── s1/
│   └── s2/
├── Free/
│   ├── digit/                 # Repository name; this branch represents DIGIT
│   ├── s1/
│   └── s2/
├── Integrated_tasks/
└── Obstacle_avoidance/

Sensor identifiers:

Identifier	Description
digit	DIGIT vision-based tactile sensor
s1	Spherical vision-based tactile sensor
s2	Ellipsoidal vision-based tactile sensor
s3	Large-area resistive tactile sensor with 12 raw channels

3. Object Naming

Object directories follow this convention:

<shape>_<material>_<size>

Examples:

ball_0010_10
cylinder_PLA_17.5

3.1 Shapes

• ball: spherical contact object;
• cylinder: cylindrical contact object.

3.2 Materials

Code	Material
0010	Ecoflex 00-10
0030	Ecoflex 00-30
0050	Ecoflex 00-50
15A	Dragon Skin 15
PLA	PLA

3.3 Sizes

The size codes are 10, 12.5, 15, 17.5, and 20. They represent the nominal object radius in millimeters.

A complete single-object branch in Fixed, Free, or Obstacle_avoidance contains 2 × 5 × 5 = 50 shape-material-size combinations. Integrated_tasks instead uses selected pairs of complete object identifiers. Because the repository may continue to receive updates, check the relevant online directory for the combinations currently available.

4. Fixed: Fixed-Position Contact

4.1 Directory Structure

Fixed/
└── <sensor>/
    └── <shape>_<material>_<size>/
        ├── press_depth_01/
        │   ├── up/
        │   │   ├── camera_frames_60hz.tar
        │   │   ├── fem_node.csv
        │   │   ├── franka_data.csv
        │   │   ├── multimodal_aligned_data.csv
        │   │   ├── real_sim_motion_alignment.csv
        │   │   └── simulation_data.csv
        │   └── down/
        │       └── ...
        ├── press_depth_02/
        └── ...

Each object condition contains up to ten press-depth settings, from press_depth_01 to press_depth_10. Every depth is divided into two motion stages:

• up: lifting or upward motion;
• down: pressing, downward motion, or slipping.

The current Fixed data are aligned in the displacement domain. Sixty positions are selected for each stage and reindexed from 0 to 59. The files simulation_data.csv, real_sim_motion_alignment.csv, franka_data.csv, and multimodal_aligned_data.csv contain one row per aligned frame. fem_node.csv contains multiple node records per frame, so it should have 60 unique frames rather than 60 rows.

4.2 File Descriptions

franka_data.csv

Real-robot task and control-cycle data, including:

• ROS, robot, and sensor timestamps;
• task state and cycle index;
• commanded and measured gripper widths;
• actual and commanded Franka end-effector positions and quaternion orientations;
• processed, raw, and online-baseline XJC force/torque measurements;
• Franka-estimated six-axis end-effector force/torque;
• directional forces, force reference, detachment detection, and commanded Z-axis velocity and acceleration.

multimodal_aligned_data.csv

The main camera-centered, real-world multimodal alignment table, including:

• camera sequence number, nanosecond timestamp, image path, and resolution;
• timestamp correspondences among the task, robot, XJC force sensor, and gripper;
• actual and commanded end-effector poses;
• XJC and Franka six-axis force/torque values;
• position, velocity, measured effort, and external torque for the seven arm joints;
• position, velocity, effort, and total width for the two gripper fingers.

simulation_data.csv

Frame-level simulated force/torque curves with the following fields:

frame, time_s, phase, command_travel_m,
Fx_N, Fy_N, Fz_N, Mx_Nm, My_Nm, Mz_Nm,
delta_Fx_N, delta_Fy_N, delta_Fz_N,
delta_Mx_Nm, delta_My_Nm, delta_Mz_Nm,
delta_force_norm_N, delta_moment_norm_Nm

real_sim_motion_alignment.csv

A real-to-simulation motion mapping table containing:

• frame, time_s, phase, and camera_time_ns;
• relative displacement and absolute coordinates of the real end effector;
• relative displacement and absolute coordinates of the Isaac TCP;
• relative displacement and absolute coordinates of the IPC contact-region midpoint.

fem_node.csv

Surface-node data from the FEM simulation:

Field	Description	Unit
frame / aligned_frame	Frame index	—
time_s / aligned_time_s	Simulation or aligned time	s
node_id	Node identifier	—
x_m, y_m, z_m	Node coordinates	m
ux_m, uy_m, uz_m	Node displacements	m
fx_N, fy_N, fz_N	Nodal contact forces	N

camera_frames_60hz.tar

An archive containing the image directory referenced by camera_image_path in the CSV files. Extract it with:

tar -xf camera_frames_60hz.tar

The 60hz suffix is retained from the aligned export naming convention. For time synchronization, use the frame indices and nanosecond timestamps stored in the CSV files.

5. Free: Free Picking and Insertion

5.1 Directory Structure

Free/
└── <sensor>/
    └── <shape>_<material>_<size>/
        ├── width_01/
        │   ├── insert/
        │   │   ├── camera_frames_60hz.tar
        │   │   ├── fem_node.csv
        │   │   ├── franka_data.csv
        │   │   └── simulation_data.csv
        │   └── pick/
        │       └── ...
        ├── width_02/
        └── ...

Each object condition contains ten gripper-width settings, from width_01 to width_10. The measured width is recorded in franka_data.csv as measured_gripper_width_m; do not infer a physical width solely from the directory index.

• pick: lifting or picking the object;
• insert: lateral motion, insertion, holding, return, and final insertion.

Free sequences have variable lengths and should not be assumed to contain 60 frames. The three CSV files are aligned through aligned_frame, aligned_time_s, and their phase fields.

5.2 File Descriptions

franka_data.csv

An aligned real-robot table with 13 fields:

aligned_frame, aligned_time_s, alignment_phase, task_ros_time_ns,
camera_image_path,
robot_ee_x_m, robot_ee_y_m, robot_ee_z_m,
robot_ee_qx, robot_ee_qy, robot_ee_qz, robot_ee_qw,
measured_gripper_width_m

simulation_data.csv

An aligned simulation-motion table with seven fields:

aligned_frame, aligned_time_s, alignment_phase,
gripper_y_m, gripper_z_m,
ipc_midpoint_y_m, ipc_midpoint_z_m

fem_node.csv

Aligned node-level FEM data. In addition to the node coordinates, displacements, and contact forces described in Section 4.2, this file contains a phase column.

Legacy .bak File in the DIGIT Branch

Some action directories under Free/digit also contain:

camera_franka_gripper_aligned_60hz.csv.bak

This is a retained legacy or backup alignment table. Use the standardized franka_data.csv for normal analysis. Read the .bak file only when additional legacy columns are needed for traceability.

6. Obstacle_avoidance: Obstacle Avoidance

6.1 Directory Structure

Obstacle_avoidance/
└── <shape>_<material>_<size>/
    ├── position_01/
    │   ├── depth_01/
    │   │   ├── fem_node.csv
    │   │   ├── franka_data.csv
    │   │   └── simulation_data.csv
    │   ├── depth_02/
    │   └── ...
    ├── position_02/
    └── ...

Each object condition contains 12 contact positions, from position_01 to position_12. Every position contains five depths, from depth_01 to depth_05. These directory names are discrete experimental-condition identifiers. Use the end-effector pose, Isaac/IPC coordinates, and phase fields in the CSV files for the physical trajectory.

6.2 File Descriptions

franka_data.csv

Real-world obstacle-avoidance data with 23 fields:

aligned_frame, aligned_time_s, alignment_phase,
channel_01_raw, ..., channel_12_raw,
ee_x_m, ee_y_m, ee_z_m,
ee_qx, ee_qy, ee_qz, ee_qw,
unix_time_ns

channel_01_raw through channel_12_raw are the 12 raw tactile channels. The ee_* fields describe the Franka end-effector pose.

simulation_data.csv

A simulated motion trajectory with nine fields:

frame, time_s, phase,
isaac_x_m, isaac_y_m, isaac_z_m,
ipc_x_m, ipc_y_m, ipc_z_m

fem_node.csv

Frame- and node-level coordinates, displacements, and three-axis contact forces. Its fields follow the common fem_node.csv definition in Section 4.2.

7. Integrated_tasks: Integrated Multi-Stage Tasks

7.1 Directory Structure

Integrated_tasks/
└── <ball_object>_<cylinder_object>/
    ├── pose1/
    │   ├── width1/
    │   │   ├── camera_frames_60hz.tar
    │   │   ├── fem_data.csv
    │   │   ├── fem_node.csv
    │   │   └── franka_s3_data.csv
    │   ├── width2/
    │   └── ...
    ├── pose2/
    └── ...

Unlike the single-object directory names used by the other task families, an integrated-task directory concatenates two complete object identifiers. The first identifier describes the ball and the second describes the cylinder. Examples currently present in the repository include:

ball_0010_20_cylinder_0010_20
ball_0010_20_cylinder_0030_20
ball_0010_20_cylinder_0050_17.5
ball_0010_20_cylinder_15A_20

Each paired-object condition may contain up to five pose settings, pose1 through pose5, and each pose may contain up to eight width settings, width1 through width8. These are ordinal experimental-condition identifiers; note that they do not use the zero-padded naming convention found in Free.

The integrated sequence combines gripper closing, lifting/probing, retreating, a smooth multi-pose motion, and gripper opening. Real-world phase labels include CLOSE_GRIPPER, POSE1_TO_POSE2, POSE2_TO_POSE3, POSE3_TO_POSE4, SMOOTH_POSE4_TO_POSE8, and OPEN_AT_POSE8. Simulation phases provide more detailed sub-stages such as calibrated delays, stable periods, and fault guards.

Integrated-task sequences have variable lengths and should not be assumed to contain 60 frames. Align their files using frame and time_s.

7.2 File Descriptions

franka_s3_data.csv

A 24-field real-world table combining Franka motion, camera references, and S3 tactile readings:

frame, time_s, camera_ros_time_ns, phase, image_relative_path,
ee_x_m, ee_y_m, ee_z_m,
ee_qx, ee_qy, ee_qz, ee_qw,
arduino_ch00_raw, ..., arduino_ch11_raw

The ee_* fields describe the Franka end-effector pose. The twelve arduino_ch*_raw fields are raw tactile channels from the S3 sensor. image_relative_path identifies the corresponding frame in camera_frames_60hz.tar.

fem_data.csv

A nine-field simulation-motion table:

frame, time_s, phase,
isaac_x_m, isaac_y_m, isaac_z_m,
ipc_midpoint_x_m, ipc_midpoint_y_m, ipc_midpoint_z_m

It records the Isaac reference position and IPC contact-region midpoint for every simulation frame.

fem_node.csv

The integrated-task node table uses a wide format, unlike the long node format used by Fixed, Free, and Obstacle_avoidance. Each row represents one frame and starts with:

frame, time_s, alignment_displacement_progress

The remaining columns are grouped by node identifier. Every selected node contributes nine values:

node_<id>_x_m,  node_<id>_y_m,  node_<id>_z_m,
node_<id>_ux_m, node_<id>_uy_m, node_<id>_uz_m,
node_<id>_fx_N, node_<id>_fy_N, node_<id>_fz_N

Current files contain 2,010 columns: three frame-level fields followed by nine fields for each of 223 selected nodes. Node identifiers are encoded in the column names and are not necessarily contiguous.

camera_frames_60hz.tar

The camera archive contains the frames referenced by image_relative_path in franka_s3_data.csv. Extract it in the corresponding width directory before resolving image paths.

8. Alignment and Unit Conventions

• frame and aligned_frame start at 0;
• fields ending in *_time_s are expressed in seconds;
• fields ending in *_time_ns are expressed in nanoseconds;
• position and displacement fields ending in _m are expressed in meters;
• force fields ending in _N are expressed in newtons;
• moment fields ending in _Nm are expressed in newton-meters;
• quaternions follow the qx, qy, qz, qw order;
• align tables within a sequence by frame index first, then use timestamps for validation;
• in the long-format fem_node.csv files used by Fixed, Free, and Obstacle_avoidance, use the frame index together with node_id as the node-record key;
• in the wide-format Integrated_tasks/fem_node.csv, each row is one frame and node identifiers are embedded in column names.