Relo02's picture
License: CC BY 4.0
8a7e7e8 verified
|
Raw History Blame Contribute Delete
4.81 kB
metadata
license: cc-by-4.0
pretty_name: Loco-manipulation data
tags:
  - robotics
  - teleoperation
  - humanoid
  - unitree-g1
  - tactile

Loco-manipulation data

Whole-body teleoperation demonstrations recorded on a humanoid robot. Each top-level folder is one task; each task holds the takes exactly as recorded (teleop_raw/) and the same takes converted to the HTD raw-episode format (htd/).

Tasks

Folder Prompt Takes Frames Robot
white_becker_from_sliding_ruller/ "pick up the white becker from the black sliding ruller and place it into the blue box" 99 111,719 (62.1 min at 30 Hz) Unitree G1, Inspire RH56DFTP hands, head D435i

Download

Install the Hugging Face CLI (pip install -U huggingface_hub, or curl -LsSf https://hf.co/cli/install.sh | bash). The dataset is public: no login is needed. Every command below writes into ./loco-manipulation-data, keeping the repo's folder layout, and can be re-run to resume an interrupted download.

Everything (~34 GB):

hf download Relo02/loco-manipulation-data --repo-type dataset --local-dir loco-manipulation-data

HTD episodes only (.hdf5, ~15.5 GB):

hf download Relo02/loco-manipulation-data --repo-type dataset --local-dir loco-manipulation-data \
  --include "*/htd/*"

Raw teleop takes only (ep.json, frames.parquet, main.mp4, depth.tar, ~18.4 GB):

hf download Relo02/loco-manipulation-data --repo-type dataset --local-dir loco-manipulation-data \
  --include "*/teleop_raw/*"

Raw takes without the depth images (~0.5 GB): add --exclude "*/depth.tar" to the command above. To unpack the depth images of every downloaded take:

find loco-manipulation-data -name depth.tar -execdir tar xf depth.tar \;

One task only: replace the * at the start of the pattern with the task's folder, e.g. --include "white_becker_from_sliding_ruller/htd/*".

One take only, e.g. raw take 0 and its HTD episode:

hf download Relo02/loco-manipulation-data --repo-type dataset --local-dir loco-manipulation-data \
  --include "white_becker_from_sliding_ruller/teleop_raw/manipulation/pick_up_the_white_becker/episode_000000/*" \
  --include "white_becker_from_sliding_ruller/htd/g1_inspire_wb/white_becker_from_sliding_ruller__manipulation/episode_1.hdf5"

Add --dry-run to any command to list the files and total size without downloading.

From Python:

from huggingface_hub import snapshot_download
snapshot_download("Relo02/loco-manipulation-data", repo_type="dataset",
                  local_dir="loco-manipulation-data", allow_patterns=["*/htd/*"])

Layout of a task

<task>/
  teleop_raw/manipulation/<subtask>/
    metadata/metadata.json      title, category, the prompt (description)
    episode_000000/
      ep.json                   metadata: prompt, fps, length, success, camera pose and
                                intrinsics, and the names/layout of every column below
      frames.parquet            one row per frame (30 Hz): joint positions and velocities,
                                IMU, odometry, hand touch (1062 taxels per hand) and finger
                                forces, walk state, every command, timestamps
      main.mp4                  head camera colour, H.264 640x480 30 fps
      depth.tar                 one 16-bit 640x480 PNG per frame (depth/NNNNNN.png);
                                `tar xf depth.tar` restores the depth/ folder
  htd/g1_inspire_wb/<task>__manipulation/
    episode_1.hdf5 ...          HTD format (schema htd_v1): observations (head camera
                                JPEGs, proprioception, touch), actions, camera poses,
                                masks; the prompt is in provenance.attrs["task"]

The raw takes are numbered from 0 and the HTD episodes from 1: teleop_raw/.../episode_0000NN is htd/.../episode_{NN+1}.hdf5.

Takes were reviewed after recording and failed ones removed; the remaining takes are renumbered contiguously, and episode_index / index inside ep.json and frames.parquet match the folder names.

Recording

  • Operator: Meta Quest 3, controllers (control_mode: controller), so action.hand_landmarks is zero and the HTD fingertip targets are built from the commanded fingers.
  • Lower body: the walking policy on the robot's onboard computer.
  • The HTD episodes carry the head camera only (no wrist cameras) and a placeholder gripper channel, as HTD's own collector writes.

License

Creative Commons Attribution 4.0 International (CC BY 4.0). You may share and adapt the data for any purpose, including commercially, provided you give appropriate credit, link to the license, and indicate if changes were made.