Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed because of a cast error
Error code:   DatasetGenerationCastError
Exception:    DatasetGenerationCastError
Message:      An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 9 new columns ({'box_yaw_deg', 'lift_end_s', 'row_key', 'source_key', 'entry_sha256_robot', 'is_yaw0', 'entry_sha256_object', 'box_scale', 'duration_s'}) and 3 missing columns ({'yaw_variants_deg', 'row_frames_30fps', 'n_clean_rows'}).

This happened while the csv dataset builder was generating data using

hf://datasets/daibfc/h2r-omomo-largebox/rows.csv (at revision 280aca6cd24e3dc791f4ad0b384341a972f2e2b6), ['hf://datasets/daibfc/h2r-omomo-largebox@280aca6cd24e3dc791f4ad0b384341a972f2e2b6/omomo_largebox_324.csv', 'hf://datasets/daibfc/h2r-omomo-largebox@280aca6cd24e3dc791f4ad0b384341a972f2e2b6/rows.csv']

Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1856, in _prepare_split_single
                  writer.write_table(table)
                  ~~~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 811, in write_table
                  self._write_table(pa_table, writer_batch_size=writer_batch_size)
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 822, in _write_table
                  pa_table = table_cast(pa_table, self._schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              row_key: string
              omomo_sequence: string
              subject: string
              source_key: string
              box_scale: double
              box_yaw_deg: int64
              is_yaw0: int64
              frames_30fps: int64
              duration_s: double
              has_lift: int64
              lift_start_s: double
              lift_end_s: double
              has_release: int64
              release_start_s: double
              entry_sha256_robot: string
              entry_sha256_object: string
              -- schema metadata --
              pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 2229
              to
              {'omomo_sequence': Value('string'), 'frames_30fps': Value('int64'), 'subject': Value('string'), 'n_clean_rows': Value('int64'), 'yaw_variants_deg': Value('string'), 'row_frames_30fps': Value('int64'), 'has_lift': Value('int64'), 'has_release': Value('int64'), 'lift_start_s': Value('float64'), 'release_start_s': Value('float64')}
              because column names don't match
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1379, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 956, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1702, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1858, in _prepare_split_single
                  raise DatasetGenerationCastError.from_cast_error(
                  ...<4 lines>...
                  )
              datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
              
              All the data files must have the same columns, but at some point there are 9 new columns ({'box_yaw_deg', 'lift_end_s', 'row_key', 'source_key', 'entry_sha256_robot', 'is_yaw0', 'entry_sha256_object', 'box_scale', 'duration_s'}) and 3 missing columns ({'yaw_variants_deg', 'row_frames_30fps', 'n_clean_rows'}).
              
              This happened while the csv dataset builder was generating data using
              
              hf://datasets/daibfc/h2r-omomo-largebox/rows.csv (at revision 280aca6cd24e3dc791f4ad0b384341a972f2e2b6), ['hf://datasets/daibfc/h2r-omomo-largebox@280aca6cd24e3dc791f4ad0b384341a972f2e2b6/omomo_largebox_324.csv', 'hf://datasets/daibfc/h2r-omomo-largebox@280aca6cd24e3dc791f4ad0b384341a972f2e2b6/rows.csv']
              
              Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

omomo_sequence
string
frames_30fps
int64
subject
string
n_clean_rows
int64
yaw_variants_deg
string
row_frames_30fps
int64
has_lift
int64
has_release
int64
lift_start_s
float64
release_start_s
float64
sub3_largebox_000
118
sub3
2
0;+35
118
0
0
null
null
sub3_largebox_001
178
sub3
2
0;+35
178
0
0
null
null
sub3_largebox_002
165
sub3
2
0;+25
165
0
0
null
null
sub3_largebox_003
196
sub3
2
0;+25
196
1
1
1.06
5.9
sub3_largebox_004
214
sub3
2
0;+35
214
1
1
1.36
6.24
sub3_largebox_005
193
sub3
2
0;+25
193
1
1
1.04
5.64
sub3_largebox_009
159
sub3
2
0;-25
159
1
1
1.6
4.78
sub3_largebox_010
198
sub3
2
0;-15
198
1
1
1.26
6
sub3_largebox_011
192
sub3
2
0;+25
192
1
1
1.22
5.86
sub3_largebox_012
203
sub3
2
0;+35
203
1
0
3.88
null
sub3_largebox_013
216
sub3
2
0;+25
216
1
0
3.98
null
sub3_largebox_014
253
sub3
2
0;+35
253
1
0
4.8
null
sub3_largebox_015
147
sub3
2
0;-25
147
1
1
1.04
4.3
sub3_largebox_016
165
sub3
2
0;+25
165
1
1
0.68
4.92
sub3_largebox_017
204
sub3
2
0;-35
204
1
1
1.5
6.02
sub3_largebox_018
191
sub3
2
0;+25
191
1
1
0.54
5.86
sub3_largebox_019
124
sub3
2
0;+35
124
0
0
null
null
sub3_largebox_020
185
sub3
2
0;-35
185
0
0
null
null
sub3_largebox_021
145
sub3
2
0;-15
145
0
0
null
null
sub3_largebox_022
117
sub3
2
0;+15
117
0
0
null
null
sub3_largebox_023
126
sub3
2
0;+25
126
0
0
null
null
sub3_largebox_024
177
sub3
2
0;-35
177
0
0
null
null
sub3_largebox_025
172
sub3
2
0;-25
172
0
0
null
null
sub3_largebox_026
193
sub3
2
0;-15
193
0
0
null
null
sub3_largebox_027
165
sub3
2
0;-25
165
0
0
null
null
sub3_largebox_028
193
sub3
2
0;-15
193
1
1
1.06
5.88
sub3_largebox_029
218
sub3
2
0;-35
218
1
1
1.46
6.74
sub3_largebox_030
204
sub3
2
0;+15
204
1
1
1.08
6.3
sub3_largebox_031
178
sub3
2
0;+25
178
0
0
null
null
sub3_largebox_032
181
sub3
2
0;-35
181
0
0
null
null
sub3_largebox_033
193
sub3
2
0;-15
193
0
0
null
null
sub3_largebox_034
214
sub3
2
0;+35
214
1
1
1.64
6.36
sub3_largebox_035
193
sub3
2
0;+35
193
1
1
1
5.66
sub3_largebox_036
178
sub3
2
0;+35
178
1
1
0.9
5.38
sub7_largebox_000
146
sub7
2
0;+35
146
0
0
null
null
sub7_largebox_001
166
sub7
2
0;-25
166
0
0
null
null
sub7_largebox_002
134
sub7
2
0;+35
134
0
0
null
null
sub7_largebox_003
147
sub7
2
0;-15
147
0
0
null
null
sub7_largebox_004
131
sub7
2
0;+25
131
0
0
null
null
sub7_largebox_005
117
sub7
2
0;+35
117
0
0
null
null
sub7_largebox_012
211
sub7
2
0;+25
211
1
1
0.9
6.12
sub7_largebox_013
185
sub7
2
0;-25
185
1
1
0.88
5.48
sub7_largebox_014
157
sub7
2
0;-25
157
1
1
0.64
4.52
sub7_largebox_015
216
sub7
2
0;-35
216
1
1
1.02
6
sub7_largebox_016
194
sub7
2
0;+25
194
1
0
0.82
null
sub7_largebox_018
161
sub7
2
0;-25
161
1
0
2.54
null
sub7_largebox_019
133
sub7
2
0;-15
133
1
0
1.74
null
sub7_largebox_020
140
sub7
2
0;-15
140
1
0
1.76
null
sub7_largebox_021
155
sub7
2
0;+25
155
1
0
2.3
null
sub7_largebox_022
151
sub7
2
0;+25
151
1
0
1.88
null
sub7_largebox_023
151
sub7
2
0;-15
151
1
0
2.54
null
sub7_largebox_024
168
sub7
2
0;+25
168
1
1
0.46
4.54
sub7_largebox_025
215
sub7
2
0;-15
215
1
1
0.84
5.98
sub7_largebox_026
156
sub7
2
0;-15
156
1
1
0.5
4.24
sub7_largebox_027
184
sub7
2
0;-25
184
1
1
1.26
5.64
sub7_largebox_028
200
sub7
2
0;+25
200
1
1
0.84
6.24
sub7_largebox_029
208
sub7
2
0;-15
208
1
1
0.5
5.44
sub7_largebox_030
190
sub7
2
0;+35
190
0
0
null
null
sub7_largebox_031
192
sub7
2
0;-15
192
0
0
null
null
sub7_largebox_032
184
sub7
2
0;+15
184
0
0
null
null
sub7_largebox_033
210
sub7
2
0;-35
210
0
0
null
null
sub7_largebox_034
188
sub7
2
0;-35
188
0
0
null
null
sub7_largebox_035
190
sub7
2
0;-25
190
0
0
null
null
sub7_largebox_036
158
sub7
2
0;-15
158
0
0
null
null
sub7_largebox_037
131
sub7
2
0;-25
131
0
0
null
null
sub7_largebox_038
189
sub7
2
0;+15
189
0
0
null
null
sub7_largebox_039
147
sub7
2
0;+35
147
0
0
null
null
sub7_largebox_040
160
sub7
2
0;+35
160
0
0
null
null
sub7_largebox_041
165
sub7
2
0;-25
165
0
0
null
null
sub7_largebox_042
213
sub7
2
0;-15
213
1
1
1.62
6.14
sub7_largebox_043
224
sub7
2
0;-25
224
1
1
1.58
6.38
sub7_largebox_044
213
sub7
2
0;+35
213
1
1
1
5.86
sub7_largebox_045
232
sub7
2
0;-35
232
1
1
1.98
6.68
sub7_largebox_046
189
sub7
2
0;-35
189
1
1
1.08
5.5
sub7_largebox_047
198
sub7
2
0;+25
198
1
1
1.7
5.88
sub7_largebox_048
264
sub7
2
0;-15
264
0
0
null
null
sub7_largebox_049
146
sub7
2
0;+15
146
0
0
null
null
sub7_largebox_050
177
sub7
2
0;+15
177
0
0
null
null
sub7_largebox_051
157
sub7
2
0;-15
157
0
0
null
null
sub7_largebox_052
137
sub7
2
0;-25
137
0
0
null
null
sub7_largebox_053
142
sub7
2
0;+35
142
0
0
null
null
sub8_largebox_000
129
sub8
2
0;-35
129
0
0
null
null
sub8_largebox_001
151
sub8
2
0;+35
151
0
0
null
null
sub8_largebox_002
174
sub8
2
0;+35
174
0
0
null
null
sub8_largebox_003
136
sub8
2
0;+15
136
0
0
null
null
sub8_largebox_004
132
sub8
2
0;-25
132
0
0
null
null
sub8_largebox_005
190
sub8
2
0;-15
190
0
0
null
null
sub8_largebox_006
216
sub8
2
0;+25
216
1
1
1.58
6.2
sub8_largebox_018
195
sub8
2
0;-25
195
1
0
2.64
null
sub8_largebox_019
219
sub8
2
0;-15
219
1
0
2.98
null
sub8_largebox_020
198
sub8
2
0;-15
198
1
0
2.98
null
sub8_largebox_021
193
sub8
2
0;-35
193
1
0
3.2
null
sub8_largebox_022
210
sub8
2
0;-25
210
1
0
3.26
null
sub8_largebox_023
218
sub8
2
0;-15
218
1
0
3.2
null
sub8_largebox_024
183
sub8
2
0;-35
183
1
1
0.76
4.88
sub8_largebox_025
191
sub8
2
0;+35
191
1
1
0.82
5.4
sub8_largebox_026
198
sub8
2
0;+25
198
1
1
0.68
5.48
sub8_largebox_027
228
sub8
2
0;-35
228
1
1
1.96
6.56
sub8_largebox_028
202
sub8
2
0;-35
202
1
1
0.86
5.4
sub8_largebox_029
165
sub8
2
0;-15
165
1
1
0.8
4.44
End of preview.

OMOMO largebox -> Unitree G1: the clean training motions of H2ROI

The exact clean reference motions our H2ROI box loco-manipulation policies are trained on: 324 OMOMO largebox sequences (6 subjects) retargeted to the Unitree G1 29-DoF humanoid with the box scaled to 0.7601, as 648 rows (each sequence once with the box at its recorded yaw and once with the box yawed by one of +-15 / +-25 / +-35 deg). Each row = one G1 whole-body motion + the matching box (6-DoF) motion, 30 fps.

These are the clean rows of our training corpus (p23_boxmix_place_A_s07601_degen_v2), copied entry by entry from its teacher_robot.pkl / teacher_objects.pkl (the per-entry digests in rows.csv equal the corpus manifest). Not included: the corpus' 2731 corrupted copies (actor-side perturbations for robustness training), the per-clip plan table, any benchmark / evaluation clip, and any human (SMPL-H) motion. No raw OMOMO data is redistributed.

rows 648 = 324 yaw-0 rows + 324 yaw-variant rows (54 each at -35, -25, -15, +15, +25, +35 deg)
OMOMO sequences 324 (sub3 34, sub7 47, sub8 42, sub10 79, sub12 69, sub16 53)
frames 117,130 at 30 fps (65.1 min; per row 72-399 frames, 2.4-13.3 s); 32.5 min of distinct source motion
box OMOMO largebox, scaled by 0.7601 about the centre of its floor face (true size 25.5 x 25.7 x 27.4 cm)
content 330 rows (165 sequences) lift the box, 125 sequences of them also put it down; 318 rows (159 sequences) push / drag it on the floor (no lift)

Files

file content
motions/g1_robot.pkl G1 reference motion per row (joblib pickle, dict keyed by row key)
motions/largebox_object.pkl box motion per row (same keys)
rows.csv one line per row: key, sequence, subject, yaw, is_yaw0, frames, lift / release times, entry digests
omomo_largebox_324.csv one line per OMOMO sequence: frames, subject, rows, yaw variants, lift / release
assets/largebox.obj OMOMO largebox mesh, unscaled, in its asset frame = the frame the box poses refer to
assets/largebox_s0.7601.obj the same mesh scaled by 0.7601 about the floor-face pivot (= the simulated box), asset frame
assets/largebox_obb.obj unscaled mesh re-expressed in its tight box frame (centred, faces axis-aligned)
assets/box_frame.json pivot, scale, true size, asset->box-frame transforms (unscaled and at 0.7601)
SHA256SUMS sha256 of every file

Row key: <sequence>__box_s0.7601_y<signed yaw>__original__clean, e.g. sub10_largebox_000__box_s0.7601_y+0__original__clean (y+0 = the yaw-0 original, is_yaw0=1 in rows.csv; __original__clean = the clean reference, no corruption).

Format (native SONIC / GR00T-WholeBodyControl motion-library format)

Load with joblib.load (needs joblib, numpy). Both files are dict[row_key] -> dict; T = frames of that row.

g1_robot.pkl entry:

field dtype, shape meaning
root_trans_offset float32 (T, 3) pelvis (root body) position in the world, metres
root_rot float32 (T, 4) pelvis orientation, quaternion x, y, z, w (scipy order)
dof float32 (T, 29) joint angles, rad, order below
pose_aa float32 (T, 30, 3) axis-angle: [:, 0] = root rotation vector of root_rot; [:, 1+i] = joint axis of joint i x dof[:, i] (redundant with dof)
fps int 30
box_scale, box_yaw_deg float 0.7601; 0 or the variant yaw
sampling_weight float our trainer's sampling weight (uniform over rows here; ignore)
paired_reference dict pairing metadata of our trainer (pair_key, source_key, family clean, severity 0, role)

largebox_object.pkl entry:

field dtype, shape meaning
root_pos float32 (T, 1, 3) world position of the largebox.obj asset-frame origin, metres (one object)
root_quat float32 (T, 1, 4) box orientation, quaternion w, x, y, z (note: different order from the robot)
fps, box_scale, box_yaw_deg, sampling_weight, paired_reference as above

World frame: right-handed, z up, floor at z = 0, metres; each row keeps the OMOMO world placement (not re-centred or heading-normalised). Frame rate: 30 fps; our trainer interpolates to 50 Hz (slerp for rotations).

Joint order (dof columns = the MJCF g1_29dof_rev_1_0.xml order of the Unitree G1 29-DoF rev 1.0, rubber hands): left_hip_pitch, left_hip_roll, left_hip_yaw, left_knee, left_ankle_pitch, left_ankle_roll, right_hip_pitch, right_hip_roll, right_hip_yaw, right_knee, right_ankle_pitch, right_ankle_roll, waist_yaw, waist_roll, waist_pitch, left_shoulder_pitch, left_shoulder_roll, left_shoulder_yaw, left_elbow, left_wrist_roll, left_wrist_pitch, left_wrist_yaw, right_shoulder_pitch, right_shoulder_roll, right_shoulder_yaw, right_elbow, right_wrist_roll, right_wrist_pitch, right_wrist_yaw (each _joint). The robot model is the G1 MJCF / USD shipped in NVIDIA GR00T-WholeBodyControl (also in daibfc/h2r-benchmarks assets/).

The box (read this before using the mesh)

  • The poses are of the unscaled largebox.obj asset frame. The simulated box is the mesh scaled by 0.7601 about the pivot (-0.0194094, 0.0195478, 0.1840005) (asset frame; centre of the face that rests on the floor): x_world = root_pos + R(root_quat) @ (pivot + 0.7601 (v - pivot)), i.e. place largebox_s0.7601.obj at the pose directly. Only the geometry is scaled, the pose is not changed. Our simulator gives the box an authored mass of 0.5 kg.
  • The mesh is rotated inside its asset frame. Its axis-aligned bounding box (47.1 x 45.9 x 40.8 cm unscaled) is not the box: the true box is 33.5 x 33.8 x 36.0 cm unscaled (25.5 x 25.7 x 27.4 cm at 0.7601), with faces ~26 deg yawed and ~9 deg tilted from the asset axes. Do not take sizes, faces or a "box frame" from the asset axes or its AABB. Use assets/box_frame.json: origin_from_box_* maps tight-box-frame points to asset-frame points (v_asset = R v_box + t); box pose in the world = T_world_asset @ origin_from_box_at_scale. largebox_obb.obj is the unscaled mesh in that box frame. In the resting pose the box-frame +z face is on the floor (box-frame +z points down).
  • If your framework builds a box primitive instead of using the mesh: a cuboid of 0.2547 x 0.2567 x 0.2737 m posed at T_world_asset @ origin_from_box_at_scale approximates the simulated box (the mesh has rounded edges).

Loading example (run on the files of this repository)

import csv, json, joblib, numpy as np
from scipy.spatial.transform import Rotation as R

robot = joblib.load("motions/g1_robot.pkl")          # {row_key: dict}, 648 rows
box = joblib.load("motions/largebox_object.pkl")     # same keys
rows = list(csv.DictReader(open("rows.csv")))
key = rows[0]["row_key"]                             # e.g. sub10_largebox_000__box_s0.7601_y+0__original__clean
r, o = robot[key], box[key]
print(key, "fps", r["fps"], "frames", len(r["dof"]), "yaw", r["box_yaw_deg"], "scale", r["box_scale"])
print("pelvis xyz", r["root_trans_offset"][0], "pelvis quat xyzw", r["root_rot"][0], "dof", r["dof"].shape)
print("box xyz", o["root_pos"][0, 0], "box quat wxyz", o["root_quat"][0, 0])

# box vertices in the world at frame 0 (mesh already scaled by 0.7601 about its floor-face pivot)
v = np.array([l.split()[1:4] for l in open("assets/largebox_s0.7601.obj") if l.startswith("v ")], float)
q = o["root_quat"][0, 0]                              # wxyz -> scipy wants xyzw
world = R.from_quat(q[[1, 2, 3, 0]]).apply(v) + o["root_pos"][0, 0]
print("box bottom z at frame 0: %.4f m" % world[:, 2].min())

# tight box frame (OBB) in the world: T_world_asset @ origin_from_box_at_scale
F = np.array(json.load(open("assets/box_frame.json"))["origin_from_box_at_scale"])
Rw = R.from_quat(q[[1, 2, 3, 0]]).as_matrix()
print("box centre", Rw @ F[:3, 3] + o["root_pos"][0, 0], "box-frame +z in world", Rw @ F[:3, 2])

# yaw-0 originals vs yaw variants
print(sum(x["is_yaw0"] == "1" for x in rows), "yaw-0 rows,", sum(x["is_yaw0"] == "0" for x in rows), "yaw-variant rows")

Output:

sub10_largebox_000__box_s0.7601_y+0__original__clean fps 30 frames 120 yaw 0.0 scale 0.7601
pelvis xyz [-0.38904428  0.4485077   0.7745581 ] pelvis quat xyzw [ 0.1543576   0.15198188 -0.6824666   0.69807917] dof (120, 29)
box xyz [-0.32813075 -0.1919544   0.19075623] box quat wxyz [-0.0753058   0.48218602  0.8724378  -0.02603893]
box bottom z at frame 0: 0.0055 m
box centre [-0.33086358 -0.19204852  0.14156848] box-frame +z in world [ 0.00342368 -0.00641687 -0.99997355]
324 yaw-0 rows, 324 yaw-variant rows

At frame 0 the box's lowest vertex is within 6 mm of the floor in 632 of the 648 rows; 2 sequences start with it 1.3 cm up and 6 sequences (sub10_largebox_014/022/024/026/031, sub12_largebox_020) start mid-carry with the box 0.41-0.52 m up.

CSV columns

rows.csv: row_key, omomo_sequence, subject, source_key, box_scale, box_yaw_deg, is_yaw0, frames_30fps, duration_s, has_lift, lift_start_s, lift_end_s, has_release, release_start_s, entry_sha256_robot, entry_sha256_object. omomo_largebox_324.csv: omomo_sequence, frames_30fps, subject, n_clean_rows (always 2), yaw_variants_deg (e.g. "0;+35"), row_frames_30fps, has_lift, has_release, lift_start_s, release_start_s (lift / release times of the yaw-0 row). has_lift / has_release and their times (seconds from the clip start) come from our training plan table (per-frame lift / release flags at 50 Hz; the table itself is not part of this release): lift rows raise the box a median 0.78 m, no-lift rows keep it within 1 cm of its start height (95th percentile) while moving it 0.5-1.5 m on the floor. The two rows of a sequence always agree. entry_sha256_* are our trainer's per-entry digests (motion arrays + pairing metadata, excluding sampling_weight; gear_sonic/utils/motion_lib/paired_reference.py: entry_sha256).

How it was produced

  1. OMOMO -> G1 (nominal). HoloSOMA retargeting (amazon-far/holosoma, commit bccd4d7), task object_interaction, robot g1 (29-DoF, g1_29dof_w_largebox.xml), data_format smplh, on HoloSOMA's OMOMO preprocessing: interaction-mesh (Laplacian) matching of the human + box, with foot sticking, joint limits and non-penetration, giving <sequence>_original.npz (qpos = robot root, joints, box pose) at 30 fps for the full-size box.
  2. Box scale 0.7601 and yaw (box-mix). Every row here, including yaw 0, is re-solved with HoloSOMA's augmentation mode anchored to the nominal solution (OmniRetarget-style): the robot-side box (contact targets and collision mesh) is scaled by 0.7601 about the floor-face pivot, the human-side box keeps its real size; the box is optionally rotated about world z at the start (held until it first moves, then blended back to the original orientation, tau 25 frames); hand contact targets use rigid_native_transport (the native G1 hand-contact anchors transported to the scaled box, the rigid hand moved as a whole). HoloSOMA's solver itself is unmodified (driver: H2ROI scripts_sonic_render/box_aug/retarget_box_aug.py + box_mix/contact_targets.py, commit 805031d).
  3. Conversion. ProtoMotions convert_holosoma_npz_to_proto.py + H2ROI chain adapters (box_aug/chain, commit e4a2941) to the SONIC motion-library format above; cells merged by box_mix/build_place_cells.py (ad3098d, arm "place-A", seed 20600), subset to box scale 0.7601 by box_mix/subset_scale.py (d15dc99), then the degen_v2 corpus build, which leaves the clean arrays unchanged (the published entries' digests equal the degen_v2 manifest). The "place-A" and "degen" stages only add corrupted actor copies; they do not touch these clean rows.

The corpus' actor-side copy of every clean row (actor_robot.pkl / actor_objects.pkl) was checked against the teacher copy published here: all arrays are bit-identical; the only difference is the metadata label paired_reference.role (actor_* vs teacher_*).

Code: https://github.com/BFCmath/H2ROI (private; collaborator access), docs docs/BOXMIX_CORPUS.md, docs/DEGEN_V2_HYBRID.md. Policies trained on this data: daibfc/h2r-weights. Benchmarks (disjoint from this data): daibfc/h2r-benchmarks.

sha256

file bytes sha256
assets/box_frame.json 1,832 b217601a7342bfc301a7c9c78747cc9ccf5591e7750d5a689c967afd31385919
assets/largebox.obj 1,998,275 dbcc11281f62e9226f49165252375080d2e490e5a3f0ab6ba917acbd8f7abc1c
assets/largebox_obb.obj 1,998,380 72c22de3e9a61677d0ad79476ec061b02892d87df4913c47a1fe51082f7347e1
assets/largebox_s0.7601.obj 1,915,288 5a24f08aa390e3ff98658afd596ff5123e61e886401bfcda430bd32e582f77dd
motions/g1_robot.pkl 59,361,819 870209fa8fd5565abab9d318a249130800fff17499ac14587bd60048b2c21652
motions/largebox_object.pkl 3,527,739 36fc7d2969001a23197caa93c1158573a1f2a3ef2b062a95cb1026d3b3016b97
omomo_largebox_324.csv 16,520 5f3c29611f98ef28ff6e03a40f34ddefea10297def7cb31341a49f08294f1d79
rows.csv 180,763 8387e59bb56f91ac843694d24031db8f701c1a3a6bf80693ea13a0d3fa6f379b

SHA256SUMS lists the same digests plus this README (sha256sum -c SHA256SUMS).

License and attribution

Non-commercial research use only; use is subject to the OMOMO dataset terms.

  • Source human-object motions: OMOMO -- Jiaman Li, Jiajun Wu, C. Karen Liu, "Object Motion Guided Human Motion Synthesis", ACM Transactions on Graphics (SIGGRAPH Asia 2023). Only robot and box motions derived from it are released here; no human motion.
  • largebox mesh: from the OMOMO dataset (OMOMO terms).
  • Retargeting: HoloSOMA (amazon-far/holosoma, its license); robot: Unitree G1 29-DoF description as shipped in NVIDIA GR00T-WholeBodyControl (that repository's and Unitree's licenses).
  • G1 retargeting, box-scale / yaw augmentation and packaging: the H2ROI team.
@article{li2023object,
  title   = {Object Motion Guided Human Motion Synthesis},
  author  = {Li, Jiaman and Wu, Jiajun and Liu, C. Karen},
  journal = {ACM Transactions on Graphics (TOG)},
  volume  = {42}, number = {6}, year = {2023}
}
Downloads last month
62