RKrohn/MiTaS

Multi-Resolution Tactile Imitation Learning for Contact-Rich Robotic Manipulation

MiTaS demonstration dataset accompanying the paper.

• Project page: https://mitas-touch.github.io/
• OpenReview: https://openreview.net/forum?id=b2gSINMX3R
• arXiv: https://arxiv.org/abs/2606.06281

@inproceedings{
krohn2026multiresolution,
title={Multi-Resolution Tactile Imitation Learning for Contact-Rich Robotic Manipulation},
author={Rickmer Krohn and Erik Helmut and Niklas Funk and Jan Peters and Vignesh Prasad and Georgia Chalvatzaki},
booktitle={10th Annual Conference on Robot Learning},
year={2026},
url={https://openreview.net/forum?id=b2gSINMX3R}
}

Rickmer Krohn, Erik Helmut, Niklas Funk, Jan Peters, Vignesh Prasad, Georgia Chalvatzaki. Accepted at CoRL 2026.

Dataset

Real-world teleoperated demonstrations for contact-rich manipulation. Each episode is a self-contained HDF5 file with robot actions, proprioception, RGB camera frames, and tactile images (PNG bytes stored inside the file). Some episodes also embed an MP4 of the RGB stream.

Folder	Episodes
board_wiping	62
gear_assembly	64
key_inlock	64
lamp_install	62
lightbulb_connection	62
lamp_install_dual_gelsight	31

The first five tasks use RGB, GelSight Mini, and EveTac. lamp_install_dual_gelsight is an extra lamp-installation set (not in the paper) with two GelSight Mini sensors (gelsight_red, gelsight_yellow) and no EveTac.

345 episodes in total, about 35 GB.

Layout

board_wiping/session_01/episode_000.h5
board_wiping/session_02/episode_000.h5
...
lamp_install_dual_gelsight/episode_000.h5

Episodes are numbered chronologically from the recording time attribute within each session. Session 01 and session 02 are two independent collections of the same five tasks.

Sensors and robot

• 7-DoF arm with gripper (Polymetis)
• RGB camera, history length k=2
• GelSight Mini, history length k=2
• EveTac, history length k=16

Tactile and RGB frames are stored as PNG byte strings in object datasets, not as decoded arrays.

HDF5 schema

Top-level groups: action/, observation/ (proprio / control), observations/ (sensors). First dimension of numeric datasets is episode length T.

Useful fields:

• action/cartesian_position — (T, 6)
• action/cartesian_position_delta_ik — (T, 6)
• action/joint_position — (T, 7)
• action/gripper_position — (T,)
• observations/image_png/<camera_id> — (T, 2) PNG bytes
• observations/tactile_png/gelsight — (T, 2) PNG bytes
• observations/tactile_png/evetac — (T, 16) PNG bytes
• observations/tactile_png/gelsight_red / gelsight_yellow — (T, 2)
• observations/videos/<camera_id> — raw MP4 bytes when present

File attributes include time, success, camera_history_k, and (on later recordings) sensor_mode.

Load an episode

import io
import h5py
from PIL import Image

path = "board_wiping/session_01/episode_000.h5"

with h5py.File(path, "r") as f:
    print(dict(f.attrs))
    print("T =", f["action/joint_position"].shape[0])
    camera_id = next(iter(f["observations/image_png"].keys()))
    image = Image.open(io.BytesIO(f["observations/image_png"][camera_id][0, 0]))
    tactile = Image.open(io.BytesIO(f["observations/tactile_png/gelsight"][0, 0]))

For the dual-GelSight set, read observations/tactile_png/gelsight_red and gelsight_yellow instead of gelsight / evetac.

License

CC-BY-4.0