shinben0327/v2i-test

HOI Retarget — OMOMO → Unitree G1

Human–object interaction capture, retargeted onto a Unitree G1 with contact preserved.

Each clip is a G1 trajectory — 29 joint angles, a floating base, and the 6-DoF pose of the object it is manipulating — together with the per-link contact flags that say which of the robot's palms and feet are touching the object at each frame.

Retargeting a person carrying a chair onto a G1 is not a pose-matching problem. The robot is shorter, its hands are a rubber block rather than fingers, and a trajectory that merely looks right drops the object the moment physics is switched on. What makes this data useful is that the contact is right: each active contact link stays on the object through the motion.

⚠️ This is a 26-clip smoke subset

The full corpus is 4,421 clips / 805,891 frames / 7.46 hours across the same 13 objects. Published here are 2 clips per object — one lift and one carry — as a format and tooling preview, not a training set. All numbers below describe the subset.

26 clips · 4,144 frames · 138 s · 3.2–8.8 s per clip · object clearance 0.00–1.59 m · object path length 0.79–4.96 m

---

Quick start

from datasets import load_dataset, Video

ds = (load_dataset("shinben0327/v2i-test", split="train")   # config "clips" is the default
        .cast_column("video", Video(decode=False))          # see the note below
        .with_format("numpy"))
c = ds[0]

c["clip_id"]                 # 'largebox/sub16_largebox_007'
c["dof_pos"]                 # (T, 29)  joint angles, radians
c["root_pos"], c["root_rot"] # (T, 3) metres, (T, 4) quaternion wxyz
c["object_pos"], c["object_rot"]
c["per_link_contact_flags"]  # (T, 4)  bool — the point of the dataset
c["video"]["bytes"]          # the rendered mp4, embedded in the parquet

Two things that snippet is doing deliberately:

• .with_format("numpy") returns real ndarrays with the shapes documented below. Without it you get nested Python lists.
• .cast_column("video", Video(decode=False)) is not optional unless you have pip install torchcodec. Reading any row of this config decodes the video column, so ds[0] raises ImportError: To support decoding videos, please install 'torchcodec' even if you only wanted dof_pos. With decoding off you get the raw mp4 bytes and no extra dependency: open("clip.mp4", "wb").write(c["video"]["bytes"]).

Browse without downloading: the Dataset Viewer above plays every clip inline, and every column is queryable with SQL from the Hub.

---

Configs

Config	Rows	Size	What it holds
clips (default)	26	9.9 MB	Metadata, interaction statistics, the trajectory, the contact flags, and the rendered mp4
fk	26	10.4 MB	Forward-kinematics cache — world and pelvis-local pose of all 42 robot bodies

fk is a convenience cache, not new information: it is what you get by running FK on clips' dof_pos + root_pos + root_rot with the G1 URDF. It is split out so the default config stays light enough to browse. Join on clip_id.

fk = load_dataset("shinben0327/v2i-test", "fk", split="train")
fk[0]["world_body_pos"]      # (T, 42, 3)

The repository also carries the same renders as plain files under videos/<object>/<subject>.mp4, and a compact float16 form of the trajectories under viewer/ that powers the 3D viewer.

---

Schema

T is the clip's frame count; every array's first axis is time at 30 fps. Full details are in docs/SCHEMA.md.

Identity

Field	Type	Notes
clip_id	string	<object>/<subject>, e.g. tripod/sub14_tripod_035
object	string	One of the 13 object classes
subject	string	OMOMO's sub<N>_<object>_<ID> — the capture subject and take
source_file	string	The upstream clip this was retargeted from
robot	string	unitree_g1
schema_version	int32	1

Robot trajectory

Field	Shape	Notes
root_pos	(T, 3)	Pelvis position, metres, Z-up world
root_rot	(T, 4)	Pelvis orientation, quaternion wxyz
dof_pos	(T, 29)	Joint angles, radians. Order is given by dof_names
dof_names	(29,) str	left_hip_pitch_joint … right_wrist_yaw_joint
robot_foot_ground_contact_flags	(T, 2) bool	[left, right]

Object trajectory

Field	Shape	Notes
object_pos	(T, 3)	Object frame origin
object_rot	(T, 4)	Quaternion wxyz
object_vel	(T, 3)	Finite-difference linear velocity
object_min_height_per_frame	(T,)	Drop from the object origin down to its lowest mesh point at that orientation
object_near_floor_flags	(T,) bool
object_pos_realsize_ref	(T, 3)	The same trajectory before rescaling into the robot's reach
object_scale	float	0.83 throughout this release
object_model_path	string	e.g. assets/objects/tripod.urdf

Contact

Field	Shape	Notes
per_link_contact_flags	(T, 4) bool	Which contact links are on the object
contact_link_names	(4,) str	left_palm_pad, right_palm_pad, left_ankle_roll_link, right_ankle_roll_link
fixed_contact_points_per_frame_in_object_frame	(T, 4, 3)	Where on the object each link is held. NaN where that link is not in contact
object_contact_sequence	(T,) int8	Object touched by anything
object_ground_contact_sequence	(T,) int8	Object resting on the floor

Precomputed interaction statistics

Per-clip scalars, so you can filter without loading trajectories — obj_contact_frac, obj_airborne_frac, max_obj_clearance_m, mean_obj_clearance_contact_m, obj_travel_m, obj_net_disp_m, root_travel_m, both_feet_ground_frac, flight_frac, n_active_contact_links_mean, and contact_frac__<link> for each of the four contact links.

lifts = ds.filter(lambda r: r["obj_airborne_frac"] > 0.3)   # object leaves the floor

---

Conventions and three things that will bite you

1. Quaternions are wxyz, not xyzw — root_rot, object_rot and world_body_orient alike. SciPy's Rotation.from_quat expects xyzw; reorder before you use it.

2. Object floor clearance is object_pos[:, 2] − object_min_height_per_frame — a subtraction, not an addition. object_min_height_per_frame is the drop from the object's frame origin down to its lowest mesh point at that orientation. Using + produces a plausible-looking table in which a chair is "lifted to 1.7 m". Calibrate against object_ground_contact_sequence: the correct expression reads ≈0.002 m while that flag is set and ≈0.84 m while it is not.

3. The object mesh is scaled by object_scale (0.83), and so is its trajectory (scale_object_mesh and scale_object_trajectory are both True). The object is rescaled into the G1's reach — it is shorter than the humans in OMOMO. Render the mesh at 0.83 or it will sit visibly wrong against the robot. object_pos_realsize_ref keeps the unscaled trajectory.

Also worth knowing: fixed_contact_points_… is NaN-filled rather than zero-filled for out-of-contact frames, so np.nanmean and friends are what you want; and to find the moment of handling in a push or kick clip, restrict to frames where a contact flag is set and take peak object speed there — the frame of greatest object displacement lands after the object has come to rest metres away, untouched.

---

Method

Placeholder — to be filled in on publication.

The retargeting method, its evaluation, and how this data compares to alternatives are described in the accompanying paper. This card documents the released data only.

Compatibility with the source format

Values in the parquet are byte-identical to the pickles the producing pipeline writes — verified array-by-array across every clip, NaN masks included. examples/to_pkl.py reconstructs that directory layout and key set exactly, so a consumer expecting the original format can read this dataset with no changes.

python examples/to_pkl.py --out ./data          # -> data/<object>/<subject>/contact_window.pkl

---

Interactive 3D viewer

The videos are a fixed camera. For orbit, scrub, and per-frame contact highlighting on the real G1 and object geometry, there is a companion Space:

→ shinben0327/v2i-viewer

It reads viewer/manifest.json and the float16 clip binaries under viewer/bin/, documented in docs/SCHEMA.md so you can point your own tooling at them.

---

Licence and provenance

This dataset is released under CC BY-NC-SA 4.0.

That is not the licence of every input — it is the strictest one, which the combined work inherits. The chain was audited rather than assumed; the full record is in NOTICE.md.

Component	Upstream	Terms	Consequence here
Human motion, object motion, object meshes	OMOMO (Li et al., SIGGRAPH Asia 2023)	MIT, © 2023 Jiaman Li	Redistributable. Independently confirmed by InterAct's own licence audit, which singles OMOMO out as MIT while placing GRAB / BEHAVE / INTERCAP behind registration
Per-body contact annotations	InterAct (Xu et al., CVPR 2025)	CC BY-NC-SA 4.0	object_contact_sequence and per_link_contact_flags derive from these. NonCommercial and ShareAlike therefore govern this release
Packaging of the clips consumed here	InterMimic (Xu et al., CVPR 2025)	see repository	Cited below
Robot description used for rendering	Unitree G1	BSD-3-Clause	Notice travels with the renders — docs/LICENSE-unitree-g1.txt
SMPL-X body model	MPI	non-commercial, no redistribution	Not present. This dataset carries G1 joint angles, never a body model or any derivative of one

So: non-commercial research use, share alike, and cite all three upstream datasets. InterAct asks explicitly that the specific sub-dataset be cited, and OMOMO is that sub-dataset.

If you believe something here is misattributed or should not be redistributed, open a discussion and it will be corrected or removed.

Citation

@article{li2023omomo,
  title   = {Object Motion Guided Human Motion Synthesis},
  author  = {Li, Jiaman and Wu, Jiajun and Liu, C. Karen},
  journal = {ACM Transactions on Graphics (SIGGRAPH Asia)},
  volume  = {42}, number = {6}, year = {2023}
}

@inproceedings{xu2025interact,
  title     = {{InterAct}: Advancing Large-Scale Versatile 3D Human-Object Interaction Generation},
  booktitle = {CVPR}, year = {2025}
}

@inproceedings{xu2025intermimic,
  title     = {{InterMimic}: Towards Universal Whole-Body Control for
               Physics-Based Human-Object Interactions},
  author    = {Xu, Sirui and Ling, Hung Yu and Wang, Yu-Xiong and Gui, Liang-Yan},
  booktitle = {CVPR}, year = {2025}
}

A citation for this dataset itself will be added on publication.

Known limitations

• Kinematic, not dynamic. These trajectories have not been through a physics simulator. Torque feasibility, friction and momentum are not guaranteed.
• Not collision-checked. Self-collision and robot–object interpenetration are not constrained in this release.
• A single object scale. Every clip here is object_scale = 0.83.
• Contact links are hands and feet only. The contact map covers the two palm pads and the two ankle roll links. Torso and head contact is not modelled, so clips whose human leans an object against their chest lose that contact.
• Object meshes are not in this repository. They are MIT and could be, but the trajectories do not need them; get them from InterMimic's object-asset download if you want to render your own. The Space embeds decimated copies with the MIT notice attached.
• A smoke subset. 26 of 4,421 clips, chosen to span objects and interaction types rather than sampled at random. Do not use it to estimate corpus statistics.