Schema
One row = one (motion, robot) pair. T is the row's n_frames.
Identity and provenance
| column |
type |
meaning |
clip_id |
string |
<dataset>/<robot>/<subject>, unique |
dataset |
string |
omomo | parahome | neuraldome | corolehoi | imhd2 |
robot |
string |
unitree_g1 (29 DoF) | unitree_h2 (31 DoF) |
subject |
string |
source clip stem, verbatim from the source dataset |
object |
string |
object mesh name; not unique across datasets — see README |
object_model_path |
string |
the URDF this row was solved against, dataset-scoped and relative (e.g. neuraldome/_assets/monitor.urdf). OMOMO's 13 meshes ship here under assets/objects/ (InterMimic, MIT); the other four datasets' are not ours to redistribute. docs/DATA.md in the code repo says how to obtain each one, and hoi-retarget-stage-object writes the URDF and the sample_points.npy the floor-height check needs |
projection_mode |
string |
pipeline metadata (contact) |
schema_version |
int64 |
2: each clip is retargeted from its subject's own recorded skeleton |
Timing and scale
| column |
type |
meaning |
n_frames |
int64 |
T |
fps |
float |
30.0 throughout |
duration_s |
float |
n_frames / fps |
object_scale |
float |
object mesh (geometry) scale: 0.83 on the G1, 1.0 on the H2. The object trajectory is scaled separately, by the robot's root scale (G1 0.83, H2 1.08444) |
object_aug_scale |
float |
extra augmentation factor, 1.0 unless augmented |
scale_object_mesh, scale_object_trajectory |
bool |
whether the mesh (G1 true, H2 false) and the trajectory (true for both) were scaled |
video |
Video(decode=False) |
mp4 preview of this row's motion, 320x320; CoRoleHOI rows show both robots (this row's robot in the normal colour, the partner tinted). Undecoded, so a cell is {bytes, path} and reading rows needs no video decoder; cast_column("video", Video()) (from datasets import Video) with torchcodec installed gives frames instead |
Quality control
| column |
type |
meaning |
qc_pass |
bool |
see the rule in the README; false when qc_flags is non-empty |
qc_flags |
list[string] |
wrist_sustained (runfrac > 0.50), body_folded (trunk folds >= 50 deg further from vertical than the human's, or >= 15 deg alongside >= 15 % joint saturation), object_floating (set by hand: the IMHD² skateboard floats ~25 cm above the floor in the source) — empty when passing |
wrist_runfrac |
float |
longest run with a wrist joint beyond 80 % of its own half-range, as a fraction of the clip |
dj_max_deg |
float |
peak per-frame joint displacement, deg/frame. Descriptive only; no QC rule uses it |
palm_contact |
float |
fraction of that excursion window in which the matching palm is flagged in contact. NaN when the clip has no excursion at all. Descriptive only; no QC rule uses it |
Sequence and collaboration
Every row answers these, whether or not it is collaborative — a solo clip is a sequence of one,
not a row with holes, so groupby("sequence_id") needs no special case.
| column |
type |
meaning |
sequence_id |
string |
the captured event. Robot-agnostic, so grouping by it returns every row from that capture — for a two-actor sequence that is 4 rows (2 actors x 2 robots) |
n_actors |
int32 |
1 or 2 |
actor_role |
string |
solo, or female / male |
partner_clip_id |
string |
the same-robot partner's clip_id; empty string (not null) when there is none |
partner_z_offset_cm |
float32 |
signed: this row's object z minus the partner's. Subtract it from this row's object_pos z to land in the partner's frame. 0.0 when solo |
pair_composable |
bool |
abs(partner_z_offset_cm) <= 3.0; true when solo |
Why the offset exists. CoRoleHOI sequences were captured with two actors and solved one at a
time. object_scale is identical within every pair, but the two halves disagree about where the
floor is: the source preprocessing estimates each actor's floor from that actor's own first frames.
The disagreement is a constant rigid translation, not a scale error and not a drift: horizontally
a fixed 0.23 cm (G1) / 0.30 cm (H2) in every pair, and vertically 0 to 37.6 cm on the G1 and 0 to
49.1 cm on the H2. It is stored per row because it passes through each robot's trajectory scale
(G1 0.83, H2 1.08444); the H2 offset is typically about 1.3x the G1 offset of the same pair, but the
per-pair ratio varies (0.5 to 2.1), so read each robot's own value.
pair = ds.filter(lambda r: r["sequence_id"] == "corolehoi/100_camera__0_191"
and r["robot"] == "unitree_g1")
912 of 13,904 rows are collaborative (228 CoRoleHOI pairs x 2 robots). The other four datasets are
single-actor.
Trajectories
All are lists of per-frame values, length T. Reconstruct with
np.array([np.asarray(x) for x in row[col]]).
| column |
shape |
meaning |
root_pos |
(T, 3) |
floating-base position, metres, world frame |
root_rot |
(T, 4) |
base orientation, quaternion wxyz |
dof_pos |
(T, ndof) |
joint angles, radians; order given by dof_names |
object_pos |
(T, 3) |
object position, metres |
object_rot |
(T, 4) |
object orientation, quaternion wxyz |
object_vel |
(T, 3) |
object linear velocity, m/s |
object_pos_realsize_ref |
(T, 3) |
object position before scaling to the robot's reach |
object_min_height_per_frame |
(T,) |
lowest point of the object above the floor, metres |
Contact
| column |
shape |
meaning |
contact_link_names |
list[string] |
the links contact is tracked on (4 for the G1: two palm pads, two ankles) |
per_link_contact_flags |
(T, n_links) bool |
which of those links touch the object each frame |
fixed_contact_points_per_frame_in_object_frame |
(T, n_links, 3) |
the point on the object each link is holding, in the object frame. NaN where the link is not in contact — this is the no-contact sentinel, not corruption |
robot_foot_ground_contact_flags |
(T, 2) bool |
left/right foot on the ground |
object_contact_sequence |
(T,) int8 |
object in contact with the human/robot |
object_ground_contact_sequence |
(T,) int8 |
object resting on the ground |
object_near_floor_flags |
(T,) bool |
object within a floor-proximity threshold |
Easy to get wrong
object is not a key. Three different chairs share the name chair; join on object_model_path or (dataset, object).- Quaternions are wxyz, not xyzw.
dof_names differs between robots (29 vs 31 entries). Never index dof_pos by a hard-coded position — look the joint up by name.- The entries named
baseball are baseball bats.
