LIBERO — Rerun RRD conversion

Rerun recordings of the LIBERO benchmark demos: 130 simulated tabletop manipulation tasks across five suites, 6,500 teleoperated demos in all. Converted from the full yifengzhu-hf/LIBERO-datasets release.

Each demo becomes several .rrd files, one per layer, sharing a single recording_id. Open them in the Rerun Viewer, or register them as a catalog dataset. Layers live in their own directories, so one demo is spread across the bucket like this:

base/<suite>/<task>__demo_<N>.rrd     # the demo as recorded ('base' layer)
properties/…                          # per-demo catalog columns: suite, task, scene, language, …
urdf/…                                # the posed Franka arm: mounting pose + forward kinematics
cameras/…                             # the two cameras placed in the scene
assets/urdf-model.rrd                 # the arm meshes, once for the whole dataset
blueprints/default.rbl                # default viewer layout

<suite> is one of libero_10, libero_90, libero_goal, libero_object, libero_spatial. <task> is the source file name without its _demo.hdf5 suffix, and <N> runs 0…49.

Provenance and attribution

This dataset is a converted derivative of yifengzhu-hf/LIBERO-datasets, the demonstration data of the LIBERO benchmark. The original authors released it under the Apache License 2.0, at source revision f13aa24a, dated 2025-05-18.

All credit for the underlying data collection belongs to the original authors. The original authors do not endorse this conversion.

Changes from the source

• Each task file (<suite>/<task>_demo.hdf5) holds 50 demos, and every demo becomes its own Rerun recording. The task file's own attributes are copied into each demo's base layer.
• The base layer keeps every HDF5 dataset and attribute of the demo group under its source name, dtype and array width, so a demo can be written back to HDF5 value for value.
• Camera frames are flipped upright. The source stores them bottom-up, in the OpenGL convention.
• Three layers are added beyond the base: urdf poses the arm, cameras places the two scene cameras, and properties carries the per-demo catalog columns.

Size and scale

The full dataset is 6,500 demos: about 45 GB of .rrd files converted from 100 GB of source. A single demo's base layer is 3–11 MB. The shared arm model is 3.8 MB, once for the whole dataset.

Source HDF5 files are uncompressed and .rrd files are compressed, so the base layer lands at 40–56% of the source size. The other three layers add about 1% on top.

One task file per suite, all 50 demos of each:

Sizes are in MB (10⁶ bytes), measured on these source files:

libero_10      : libero_10/KITCHEN_SCENE3_turn_on_the_stove_and_put_the_moka_pot_on_it_demo.hdf5
libero_90      : libero_90/KITCHEN_SCENE3_turn_on_the_stove_demo.hdf5
libero_goal    : libero_goal/turn_on_the_stove_demo.hdf5
libero_object  : libero_object/pick_up_the_salad_dressing_and_place_it_in_the_basket_demo.hdf5
libero_spatial : libero_spatial/pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate_demo.hdf5

How to use

View one demo

pip install "huggingface_hub[cli]" rerun-sdk

# Every layer of one demo, plus the default blueprint and the shared arm model
hf buckets sync hf://buckets/rerun/libero . \
  --include "*/<suite>/<task>__demo_<N>.rrd" \
  --include "assets/urdf-model.rrd" \
  --include "blueprints/default.rbl"

rerun */<suite>/<task>__demo_<N>.rrd blueprints/default.rbl

Keep the leading * so every layer loads.

These recordings were written with Rerun 0.37.0. A newer viewer reads them as they are in most cases; rerun rrd migrate updates them if it cannot.

The arm meshes do not show up in the 3D view when you run the command above. You have to register assets/urdf-model.rrd as the dataset's asset on a catalog to view them. See Local Catalog on rerun-io/rrd-datasets.

Copy the whole bucket

The whole bucket fits on a local disk: hf buckets sync hf://buckets/rerun/libero . copies it. It is also reachable over an S3-compatible API, so rclone can stream it straight into your own S3-compatible storage.

Scope

Conversion code is published at rerun-io/rrd-datasets. The code and these recordings demonstrate the conversion; they are not a versioned dataset release. Files may be regenerated or replaced, and converting the same demo twice does not give identical bytes.

License

Apache License 2.0, the same license as the source dataset — the full text ships as the LICENSE file at the root of this bucket. If you redistribute or build on this data, keep the attribution above and comply with the license terms.

The shared arm model assets/urdf-model.rrd is the one file not derived from the source dataset. It carries geometry from Franka Robotics' franka_description fer model, also Apache License 2.0, Copyright 2023 Franka Robotics GmbH. Its NOTICE ships beside the LICENSE at the bucket root, and applies to those meshes.