Hebero-Dataset

Anonymous review release. Author identities are withheld during review.

Project website · Code

Expert demonstrations for Hebero, a GPU-parallel multi-task manipulation benchmark that compiles the 40 LIBERO tasks into a single Isaac Lab vectorised training loop.

The 2000 LIBERO demonstrations are re-packaged here into an Isaac Lab-native state format: alongside the original mujoco actions and proprioception, every file carries the full per-frame physics state of the robot and of every object in the scene. That is what makes the demos usable as more than imitation targets — the simulator can be hard-reset to any step of any demonstration, which is what demonstration-indexed resets, reverse curricula, and demonstration-guided reward tracking all need.

Tasks	40 (4 suites × 10)
Demonstrations	2000 (50 per task)
Transitions	336,575 (≈4.7 h at 20 Hz)
Robot	Franka Panda, 7-DoF arm + parallel gripper
Action space	7-D robosuite OSC_POSE (6-D EE delta + 1 gripper bit)
Scene assets	32 SimReady assets (31 MB of USD layers)
Format	40 HDF5 files (251 MB) + USD tree, approximately 404 MB total
Cameras	none — the demonstrations are state-only

Contents

demos/    40 HDF5 files, one per task — trajectories and per-frame physics state
USD/      32 SimReady assets — the objects and fixtures those scenes contain

The two halves are meant to be used together: the HDF5 files record poses of objects by name, and USD/<name>/ is the geometry that name refers to.

Download

hf download hebero/Hebero-Dataset --repo-type dataset --local-dir ./Hebero-Dataset

from huggingface_hub import snapshot_download

# everything
snapshot_download("hebero/Hebero-Dataset", repo_type="dataset",
                  local_dir="Hebero-Dataset")

# demonstrations only
snapshot_download("hebero/Hebero-Dataset", repo_type="dataset",
                  local_dir="Hebero-Dataset", allow_patterns="demos/*")

Demo files are named demos/{suite}_task{id}_{task_language}_demo.hdf5, so a suite is one glob: demos/libero_spatial_task*.hdf5.

Structure

demos/<task_file>.hdf5
└── data/                                  @env_args = {"env_name": ..., "type": 2}
    └── demo_{0..49}/
        ├── actions                (T, 7)  float32   OSC_POSE command
        ├── obs/                            LIBERO-native proprioception
        │   ├── ee_states          (T, 7)  float64   base-frame EE pos + quat(wxyz)
        │   ├── gripper_states     (T, 2)  float64
        │   ├── joint_states       (T, 7)  float64
        │   └── joint_velocities   (T, 7)  float64
        └── initial_state/                  full per-frame physics state
            ├── articulation/<name>/        robot + cabinets, stoves, microwaves …
            │   ├── joint_position (T, D)  float32
            │   ├── joint_velocity (T, D)  float32
            │   ├── root_pose      (T, 7)  float32   world frame, quat wxyz
            │   └── root_velocity  (T, 6)  float32
            └── rigid_object/<name>/
                ├── root_pose      (T, 7)  float32
                └── root_velocity  (T, 6)  float32

T is per-demo and ranges from 74 to 516 steps. There is no padding and no length attribute — read actions.shape[0].

Two things that will bite you if unread:

• initial_state/ holds the whole trajectory, not just the first frame — the name is historical.
• obs/ and initial_state/ are the same trajectory offset by two control steps: obs/joint_states[t] == initial_state/…/robot/joint_position[t+2, :7]. actions[t] pairs with initial_state[t].

Suites

suite	HF/LIBERO name	what varies	demos	transitions	mean length
Spatial	libero_spatial	spatial relations between identical objects	500	61,750	124
Object	libero_object	the object to be manipulated	500	74,007	148
Goal	libero_goal	the goal predicate, fixed scene	500	63,228	127
Long	libero_10	multi-stage, long-horizon tasks	500	137,590	275

All 40 tasks (click to expand)

suite	id	env_name	demos	frames	T mean	T min	T max	artic.	rigid
libero_spatial	0	libero_spatial_0_pick_up_the_black_bowl_between_the_plate_and_the_ramekin_and_place_it_on_the_plate	50	5,018	100	74	167	3	5
libero_spatial	1	libero_spatial_1_pick_up_the_black_bowl_next_to_the_ramekin_and_place_it_on_the_plate	50	6,657	133	108	188	3	5
libero_spatial	2	libero_spatial_2_pick_up_the_black_bowl_from_table_center_and_place_it_on_the_plate	50	5,832	117	92	163	3	5
libero_spatial	3	libero_spatial_3_pick_up_the_black_bowl_on_the_cookie_box_and_place_it_on_the_plate	50	5,002	100	85	170	3	5
libero_spatial	4	libero_spatial_4_pick_up_the_black_bowl_in_the_top_drawer_of_the_wooden_cabinet_and_place_it_on_the_plate	50	7,429	149	126	189	3	5
libero_spatial	5	libero_spatial_5_pick_up_the_black_bowl_on_the_ramekin_and_place_it_on_the_plate	50	5,746	115	85	153	3	5
libero_spatial	6	libero_spatial_6_pick_up_the_black_bowl_next_to_the_cookie_box_and_place_it_on_the_plate	50	6,262	125	105	158	3	5
libero_spatial	7	libero_spatial_7_pick_up_the_black_bowl_on_the_stove_and_place_it_on_the_plate	50	7,061	141	119	196	3	5
libero_spatial	8	libero_spatial_8_pick_up_the_black_bowl_next_to_the_plate_and_place_it_on_the_plate	50	5,913	118	94	172	3	5
libero_spatial	9	libero_spatial_9_pick_up_the_black_bowl_on_the_wooden_cabinet_and_place_it_on_the_plate	50	6,830	137	116	192	3	5
libero_object	0	libero_object_0_pick_up_the_alphabet_soup_and_place_it_in_the_basket	50	7,758	155	135	195	1	7
libero_object	1	libero_object_1_pick_up_the_cream_cheese_and_place_it_in_the_basket	50	7,144	143	118	186	1	7
libero_object	2	libero_object_2_pick_up_the_salad_dressing_and_place_it_in_the_basket	50	6,591	132	114	215	1	7
libero_object	3	libero_object_3_pick_up_the_bbq_sauce_and_place_it_in_the_basket	50	7,298	146	122	221	1	7
libero_object	4	libero_object_4_pick_up_the_ketchup_and_place_it_in_the_basket	50	8,006	160	134	250	1	7
libero_object	5	libero_object_5_pick_up_the_tomato_sauce_and_place_it_in_the_basket	50	7,314	146	125	183	1	7
libero_object	6	libero_object_6_pick_up_the_butter_and_place_it_in_the_basket	50	7,815	156	143	206	1	7
libero_object	7	libero_object_7_pick_up_the_milk_and_place_it_in_the_basket	50	7,231	145	125	195	1	7
libero_object	8	libero_object_8_pick_up_the_chocolate_pudding_and_place_it_in_the_basket	50	7,932	159	144	223	1	7
libero_object	9	libero_object_9_pick_up_the_orange_juice_and_place_it_in_the_basket	50	6,918	138	113	253	1	7
libero_goal	0	libero_goal_0_open_the_middle_drawer_of_the_cabinet	50	6,977	140	115	195	3	5
libero_goal	1	libero_goal_1_put_the_bowl_on_the_stove	50	5,031	101	89	122	3	5
libero_goal	2	libero_goal_2_put_the_wine_bottle_on_top_of_the_cabinet	50	5,344	107	86	145	3	5
libero_goal	3	libero_goal_3_open_the_top_drawer_and_put_the_bowl_inside	50	10,158	203	169	298	3	5
libero_goal	4	libero_goal_4_put_the_bowl_on_top_of_the_cabinet	50	5,044	101	85	147	3	5
libero_goal	5	libero_goal_5_push_the_plate_to_the_front_of_the_stove	50	7,588	152	114	219	3	5
libero_goal	6	libero_goal_6_put_the_cream_cheese_in_the_bowl	50	5,299	106	84	168	3	5
libero_goal	7	libero_goal_7_turn_on_the_stove	50	4,410	88	74	118	3	5
libero_goal	8	libero_goal_8_put_the_bowl_on_the_plate	50	4,619	92	78	125	3	5
libero_goal	9	libero_goal_9_put_the_wine_bottle_on_the_rack	50	8,758	175	137	346	3	5
libero_10	0	libero_10_0_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket	50	14,650	293	229	387	1	8
libero_10	1	libero_10_1_put_both_the_cream_cheese_box_and_the_butter_in_the_basket	50	12,971	259	233	321	1	8
libero_10	2	libero_10_2_turn_on_the_stove_and_put_the_moka_pot_on_it	50	13,248	265	218	339	2	2
libero_10	3	libero_10_3_put_the_black_bowl_in_the_bottom_drawer_of_the_cabinet_and_close_it	50	12,384	248	198	316	2	3
libero_10	4	libero_10_4_put_the_white_mug_on_the_left_plate_and_put_the_yellow_and_white_mug_on_the_right_plate	50	12,859	257	215	330	1	5
libero_10	5	libero_10_5_pick_up_the_book_and_place_it_in_the_back_compartment_of_the_caddy	50	9,420	188	149	258	1	3
libero_10	6	libero_10_6_put_the_white_mug_on_the_plate_and_put_the_chocolate_pudding_to_the_right_of_the_plate	50	12,706	254	202	341	1	4
libero_10	7	libero_10_7_put_both_the_alphabet_soup_and_the_cream_cheese_box_in_the_basket	50	13,426	269	218	333	1	5
libero_10	8	libero_10_8_put_both_moka_pots_on_the_stove	50	20,744	415	340	516	2	2
libero_10	9	libero_10_9_put_the_yellow_and_white_mug_in_the_microwave_and_close_it	50	15,182	304	223	448	2	2

Conventions

• Units — metres, radians, seconds. Control runs at 20 Hz.
• Quaternions — (w, x, y, z), Isaac Lab convention, throughout.
• Frames — everything under initial_state/ is world frame; obs/ee_states is in the robot base frame and must be composed with initial_state/articulation/robot/root_pose to reach world frame.
• Gripper — actions[:, 6] is +1 to close, the robosuite convention. Isaac Lab's gripper action term uses the opposite sign, so a loader targeting it must invert.
• Robot joints — robot/joint_position has 9 columns: 7 arm, then 2 fingers.

USD assets

USD/ holds the 32 SimReady assets these scenes are built from, converted from LIBERO's original MuJoCo meshes. One directory per asset:

USD/
├── akita_black_bowl/
│   ├── akita_black_bowl.usd          ← reference this
│   └── textures/texture.png
├── wooden_cabinet/
│   ├── wooden_cabinet.usd
│   ├── wooden_cabinet_cube.usd       sub-part, referenced by the entry point
│   └── textures/…
└── …

Always reference USD/<name>/<name>.usd. A few assets ship extra .usd files next to it — flat_stove has knob.usd and burnerplate.usd, wine_rack has five — which are sub-parts composed by the entry point, not alternatives to it.

Matching an asset to the demonstrations. Object keys in the HDF5 files carry a scene-instance suffix that the directory names do not: initial_state/rigid_object/akita_black_bowl_1 → USD/akita_black_bowl/. Strip the trailing _<n>.

Of the 32, 28 are the objects the demonstrations track by name:

	assets
Articulated (4)	flat_stove, microwave, white_cabinet, wooden_cabinet
Rigid (24)	akita_black_bowl, alphabet_soup, basket, bbq_sauce, black_book, butter, chefmate_8_frypan, chocolate_pudding, cookies, cream_cheese, desk_caddy, glazed_rim_porcelain_ramekin, ketchup, milk, moka_pot, orange_juice, plate, porcelain_mug, red_coffee_mug, salad_dressing, tomato_sauce, white_yellow_mug, wine_bottle, wine_rack

The remaining 4 — floor, kitchen_table, living_room_table, study_table — are static fixtures. They are part of the scene but are not tracked entities, so no pose stream exists for them in the HDF5 files.

The Franka Panda itself is not included; it comes from Isaac Lab's own robot assets.

Provenance and known limitations

These are the original LIBERO demonstrations, collected by human teleoperation in MuJoCo, converted to the Isaac Lab state layout. They are not re-recorded under Isaac Lab, and the conversion is a re-packaging, not a re-simulation.

Citation

If you use this dataset, please cite LIBERO, whose demonstrations it repackages:

@inproceedings{liu2023libero,
  title     = {{LIBERO}: Benchmarking Knowledge Transfer for Lifelong Robot Learning},
  author    = {Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and
               Liu, Qiang and Zhu, Yuke and Stone, Peter},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2023}
}

License

The demonstrations originate from LIBERO, released under the MIT License, and this repackaging is distributed under the same terms. LIBERO assets and task definitions remain subject to their original licences.