alessandrolapizza/digit-force-dataset

DIGIT normal-force dataset

2,940 tactile images from a DIGIT vision-based tactile sensor (the original DIGIT by Meta AI, Lambeta et al., 2020, commercial unit from GelSight; not the newer DIGIT 360). Each image is paired with the normal contact force F_z measured by a load cell. The images cover four 3D-printed indenter geometries, 35 contact positions and 21 force levels between 0 and 10 N. They were collected automatically by a Franka FR3 arm under closed-loop force control.

The dataset was used to train the force-estimation head alessandrolapizza/digit-force-mlp (validation MAE 0.227 N). It is part of the EPFL CREATE Lab semester project Development of Fingertip Haptic Device and Optimal Mapping Method of High Dimensional Tactile Data to Lower Dimensional Haptic Feedback. Code, report and CAD are on GitHub.

Contents

Path	Description
tactile/NNNN.png	Raw DIGIT frames (QVGA, 240 × 320 px portrait, RGB)
tactile_nobg/NNNN.png	Background-subtracted frames, clip(I − I_bg + 127, 0, 255). These are the model inputs.
metadata.csv	One row per sample (see below)
calibration_all.csv	Label file in the format used by the collection and training scripts: tactile = /digit-force-dataset/tactile/NNNN.png, FT = "Fx Fy Fz" (Fx and Fy are always 0)
background.png	Reference frame at rest: the last background captured in the session. tactile_nobg/ was computed online with a background re-captured before each indenter, so use it directly rather than recomputing it from tactile/.

metadata.csv columns

Column	Description
file_name	Background-subtracted image (tactile_nobg/NNNN.png)
raw_file_name	Raw image (tactile/NNNN.png)
indenter	sphere, cube_face, pyramid_tip or cube_edge
point_idx	Contact position on the 5 × 7 grid (0–34), numbered row by row: x increases first, then y
grid_x_mm, grid_y_mm	Indenter offset from the gel centre: x ∈ {−7.5 … +7.5} mm, y ∈ {−5 … +5} mm, 2.5 mm pitch
target_force_N	Force set-point of the closed-loop controller (0, 0.5, …, 10 N)
fz_N	Normal force measured by the load cell at capture time (the label)

The dataset is balanced by design: 735 samples per indenter, 140 per force level and 84 per grid position. The measured force is always within 0.21 N of the set-point.

Collection protocol

1. The DIGIT sensor is mounted on a rigid fixture on top of a uniaxial load cell (normal axis only), read over USB serial.
2. A Franka FR3 arm carries one of the four indenters (CAD files in the GitHub repo, cad/indenters/). For every grid position it:
   • takes a 0 N capture with no contact;
   • descends under closed-loop position control until the load cell reaches each force set-point;
   • waits for the force to stabilise, then captures the image and the force synchronously.
3. The load cell is tared and a new background frame is captured before each indenter.

The collection program is scripts/dataset_collection/collect_dataset.cpp (libfranka 0.18.2), together with io_server.py.

License and citation

CC BY 4.0. If you use this dataset, please cite:

@misc{dezen2026fingertip,
  author       = {De Zen, Alessandro},
  title        = {Development of Fingertip Haptic Device and Optimal Mapping Method of High Dimensional Tactile Data to Lower Dimensional Haptic Feedback},
  year         = {2026},
  howpublished = {Semester project report, CREATE Lab, EPFL},
  url          = {https://go.epfl.ch/Fingertip-Haptic-Device-and-Optimal-Mapping-Method}
}