Update docs + examples

NOTICE.md

@@ -22,13 +22,5 @@ https://github.com/Sirui-Xu/InterAct. The retargeting objective is built on them
 
 Robot models: Unitree Robotics, BSD-3-Clause — full text in `docs/LICENSE-unitree.txt`.
 
-The retargeting code is released separately under the BSD 3-Clause Licence,
-Copyright (c) 2026 ETH Zurich: https://github.com/shinben0327/hoi-retarget
-Nothing in it derives from a restrictively licensed source, which is why its terms
-differ from this corpus's.
-
-## Not included
-
-BEHAVE and FORCE (Max Planck terms forbid redistributing modified data without written
-permission), HUMOTO (Adobe Research License v1.1), CHAIRS, and PA-HOI / InterVLA / ReconMot
-(licence could not be verified).
+The retargeting code is released separately, BSD 3-Clause, Copyright (c) 2026
+ETH Zurich: https://github.com/shinben0327/hoi-retarget

README.md

@@ -43,26 +43,31 @@ configs:
 
 # HOI-Retarget — Humanoid Human–Object Interaction Motion
 
-[Project page](https://shinben0327.github.io/hoi-retarget) · [Code](https://github.com/shinben0327/hoi-retarget) · Paper: *manuscript under review*
-
-> **Reading the data.** Hugging Face's built-in table preview struggles with these
-> rows — they carry deeply nested per-frame arrays, e.g.
-> `fixed_contact_points_per_frame_in_object_frame` is `T × 4 × 3`. Use `datasets`
-> or `pyarrow` directly (see **Quick start**), and to *watch* the motions use the
-> 3D viewer: **[HOI-Retarget Contact Playback](https://huggingface.co/spaces/shinben0327/hoi-retarget-viewer)**.
-
-**Human–object interaction capture, retargeted onto humanoid robots with contact preserved.**
-
-Each row is one **(motion, robot)** pair: a robot trajectory — joint angles, a floating base, and
-the 6-DoF pose of the object being manipulated — together with the **per-link contact flags** that
-say which palms and feet touch the object at each frame.
+[Project page](https://shinben0327.github.io/hoi-retarget) · [Code](https://github.com/shinben0327/hoi-retarget) · [3D viewer](https://huggingface.co/spaces/shinben0327/hoi-retarget-viewer) · Paper: *manuscript under review*
+
+> Learning from demonstration (LfD) has enabled humanoid robots to acquire diverse whole-body
+> skills, but extending this paradigm to human-object interaction (HOI) is limited by the
+> availability of robot-compatible interaction references. We present HOI-Retarget, a
+> contact-centric retargeting method that transfers HOI onto a humanoid robot for large-scale
+> motion-data generation. Its windowed trajectory optimization uses every labeled contact as a
+> target in the object frame, balancing body tracking, foot support and smoothness under the
+> robot's kinematic limits. The method can augment a single demonstration across object sizes,
+> absorb contacts reconstructed from monocular video, and extend to several robots manipulating
+> one object.
+
+This is the corpus that method produces: **13,904 humanoid trajectories** from five human–object
+interaction capture sets, each carrying the per-link contact flags that say which palms and feet
+touch the object at every frame.
 
 Retargeting a person carrying a chair onto a humanoid 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**: every labelled contact is expressed as a target in the object frame and
-recovered by a windowed trajectory optimization that balances it against body tracking, foot
-support and smoothness under the robot's kinematic limits.
+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 usable is that the contact
+is right: every labelled contact becomes a target in the object frame, recovered by a windowed
+optimization that trades it against body tracking, foot support and smoothness inside the robot's
+joint limits.
+
+Each row is one **(motion, robot)** pair — joint angles, a floating base, the object's 6-DoF pose,
+and the contact flags.
 
 ---
 
@@ -73,9 +78,9 @@ support and smoothness under the robot's kinematic limits.
 | Source datasets | **5** — OMOMO, ParaHome, NeuralDome, CoRoleHOI, IMHD² |
 | Distinct motions | **6,952** |
 | Robot–motion rows | **13,904** (6,952 G1 + 6,952 H2) |
-| Objects | **75 distinct meshes** (~48 semantic categories) |
+| Objects | **75 distinct meshes** |
 | Duration | **828.1 min · 13.80 h** per robot pass |
-| Collaborative | 912 rows — every CoRoleHOI motion is half of a pair (228 pairs per robot); see `docs/SCHEMA.md` |
+| Collaborative | 912 rows; every CoRoleHOI motion is half of a pair (228 pairs per robot) |
 | Robots | Unitree **G1** (29 DoF, object scale 0.83) · Unitree **H2** (31 DoF, object scale 1.0) |
 | Passing QC | **6,463 G1 (93.0 %)** · **6,098 H2 (87.7 %)** |
 
@@ -88,13 +93,12 @@ support and smoothness under the robot's kinematic limits.
 | NeuralDome | 915 | 20 | 127.8 | 812 | 88.7 |
 | CoRoleHOI | 456 | 23 | 49.1 | 429 | 94.1 |
 | IMHD² | 180 | 7 | 24.4 | 178 | 98.9 |
-| **total** | **6,952** | **75**\* | **828.1** | **6,463** | **93.0** |
+| **total** | **6,952** | **75** | **828.1** | **6,463** | **93.0** |
+
+Object names repeat across datasets but the meshes do not. `chair` in NeuralDome is a different
+scan from `chair` in IMHD², so objects are keyed `<dataset>__<object>` everywhere.
 
-\* Union across datasets. The per-dataset counts sum to 75 and **no mesh is shared between
-datasets** — every one of the 75 was hashed and measured, with zero byte-identical and zero
-geometrically identical pairs. Eight *names* recur (`pan`, `chair`, `suitcase` in three datasets
-each; `book`, `trashcan`, `monitor`, `baseball`, `table` in two), so there are 64 distinct names
-but 75 distinct objects. Merging chair/table/box/case/sofa families gives ~48 semantic categories.
+![The 75 object meshes, grouped by source dataset](docs/objects.jpg)
 
 ### Per robot
 
@@ -103,8 +107,8 @@ but 75 distinct objects. Merging chair/table/box/case/sofa families gives ~48 se
 | `unitree_g1` | 29 | 0.83 | all 5 | 6,952 | 6,463 (93.0 %) |
 | `unitree_h2` | 31 | 1.0 | all 5 | 6,952 | 6,098 (87.7 %) |
 
-Both robots now cover all five datasets. Per-dataset H2 keep rates: OMOMO 86.9 %, ParaHome 95.1 %,
-NeuralDome 86.4 %, CoRoleHOI 79.6 %, IMHD² 94.4 %.
+Both robots cover all five datasets. The H2 is taller and keeps less: 86.9 % on OMOMO, 95.1 % on
+ParaHome, 86.4 % on NeuralDome, 79.6 % on CoRoleHOI, 94.4 % on IMHD².
 
 ---
 
@@ -114,29 +118,31 @@ NeuralDome 86.4 %, CoRoleHOI 79.6 %, IMHD² 94.4 %.
 from datasets import load_dataset
 import numpy as np
 
-ds = load_dataset("shinben0327/hoi-retarget", split="train")          # everything
+ds = load_dataset("shinben0327/hoi-retarget", split="train")           # everything
 ds = load_dataset("shinben0327/hoi-retarget", "omomo", split="train")  # one source dataset
 
 clean = ds.filter(lambda r: r["qc_pass"])          # the curated subset
 g1    = ds.filter(lambda r: r["robot"] == "unitree_g1")
 
 r = clean[0]
-dof = np.array([np.asarray(x) for x in r["dof_pos"]])   # (T, 29) joint angles, radians
-obj = np.array([np.asarray(x) for x in r["object_pos"]])# (T, 3) object position, metres
+dof = np.array([np.asarray(x) for x in r["dof_pos"]])    # (T, 29) joint angles, radians
+obj = np.array([np.asarray(x) for x in r["object_pos"]]) # (T, 3) object position, metres
 print(r["clip_id"], dof.shape, r["object"], r["qc_flags"])
 ```
 
-`examples/to_pkl.py` converts a row back to the `contact_window.pkl` layout the
-pipeline consumes — the same file `hoi-retarget --mode contact` writes, so a converted
-row drops straight into [the code repository](https://github.com/shinben0327/hoi-retarget) for re-solving, rendering or
-contact editing.
+Every column is described in [`docs/SCHEMA.md`](docs/SCHEMA.md).
+
+`examples/to_pkl.py` converts a row back into the `contact_window.pkl` layout the pipeline
+consumes — the same file `hoi-retarget --mode contact` writes, so a converted row drops straight
+into [the code repository](https://github.com/shinben0327/hoi-retarget) for re-solving, rendering
+or contact editing.
 
 ---
 
 ## Quality control
 
-Every row carries a `qc_pass` flag, evaluated per **(motion, robot)** — the two robots have
-different joint limits and fail differently, so this is a property of the pair, not of the motion.
+`qc_pass` is evaluated per **(motion, robot)**. The two robots have different joint limits and fail
+differently, so it is a property of the pair, not of the motion.
 
 ```
 qc_pass = False  if  dj_max_deg    > 60      # peak per-frame joint displacement, deg/frame
@@ -147,24 +153,25 @@ qc_pass = False  if  dj_max_deg    > 60      # peak per-frame joint displacement
 ```
 
 `palm_contact` is the fraction of that excursion window in which the matching palm is flagged in
-contact. A wrist saturated **with** contact is a hard but legitimate grasp — a deep floor reach or
-an overhead carry. Saturated **without** contact is a retargeting artefact.
+contact. A wrist saturated **with** contact is a hard but legitimate grasp: a deep floor reach, an
+overhead carry. Saturated **without** contact is a retargeting artefact.
 
-**What QC does and does not cover.** These flags mark motion that is *physically implausible* —
-a joint parked against its stop, a discontinuity no body could produce. They do **not** judge the
-interaction itself. A clip where the robot kicks a box, pushes a suitcase or never uses its hands
-is unusual, not defective, and passes QC. Filter on the descriptive columns for that, not `qc_pass`.
+These flags catch motion that is physically implausible — a joint parked against its stop, a
+discontinuity no body could produce. They say nothing about the interaction. A clip where the robot
+kicks a box, pushes a suitcase or never uses its hands is unusual, not defective, and passes. Filter
+on the descriptive columns for that.
 
-Failing rows **are included**, flagged `qc_pass = false`, with the raw metrics
-(`wrist_runfrac`, `dj_max_deg`, `palm_contact`) kept as columns so you can apply your own
-thresholds instead of ours.
+Failing rows ship too, flagged `qc_pass = false`, with `wrist_runfrac`, `dj_max_deg` and
+`palm_contact` kept as columns so you can set your own thresholds instead of ours.
+
+---
 
 ## Seeing the motions
 
 Two ways, neither needing a download.
 
-**Inline previews.** Every row carries a 320 x 320 render of that exact trajectory in the `video`
-column — around 50 kB — so the table preview plays each clip in place.
+**Inline previews.** Every row carries a 320 × 320 render of that exact trajectory in the `video`
+column, around 50 kB, so the table preview plays each clip in place.
 
 ```python
 from datasets import load_dataset, Video
@@ -174,22 +181,21 @@ ds.cast_column("video", Video(decode=False))[0]["video"]["bytes"]   # raw mp4
 ```
 
 **The 3D viewer.** [HOI-Retarget Contact Playback](https://huggingface.co/spaces/shinben0327/hoi-retarget-viewer)
-plays any clip in the corpus in the browser — scrub it, orbit it, and watch which of the robot's
-four contact links is touching the object at each frame. Search and filter by dataset, robot, QC
-result, or collaborative pairs.
+plays any clip in the corpus in the browser. Scrub it, orbit it, and watch which of the robot's four
+contact links is on the object at each frame. Search and filter by dataset, robot, QC result or
+collaborative pair.
 
-It reads its geometry from this repository's `viewer/` directory, which holds the same
-trajectories in a compact playback format: per-body world poses and per-frame contact flags as
-little-endian float16, one file per (motion, robot).
+It reads its geometry from this repository's `viewer/` directory, which holds the same trajectories
+in a compact playback format: per-body world poses and per-frame contact flags as little-endian
+float16, one file per (motion, robot).
 
 ```
-viewer/index.json              every clip, with its agents and metadata (0.3 MB gzipped)
+viewer/index.json                  every clip, with its agents and metadata (0.25 MB gzipped)
 viewer/shards/<ds>__<robot>.json   the same records grouped, for lazy loading
-viewer/bin/<ds>__<robot>__<subject>.bin
+viewer/bin/<ds>__<robot>/<subject>.bin
 ```
 
-Each `.bin` is written **block-major** — each array whole, one after another, not interleaved per
-frame:
+Each `.bin` is **block-major** — each array whole, one after another, not interleaved per frame:
 
 ```
 body_pos   (T, NB, 3)   then  body_quat  (T, NB, 4)  wxyz  then
@@ -198,42 +204,38 @@ contact    (T, NC)      0.0 / 1.0
 ```
 
 so `object_pos` begins at half-offset `T*NB*3 + T*NB*4`. `NB` is 42 for the G1 and 36 for the H2;
-`NC` is 4. Slice by block, not by a per-frame stride — the file is the same size either way, so a
+`NC` is 4. Slice by block, not by a per-frame stride: the file is the same size either way, so a
 length check cannot tell the two readings apart.
 
-The 227 collaborative CoRoleHOI clips carry two agents sharing one object. Both agents store their
-own copy of the object track; they agree to 0.35 cm once each agent's `z_offset_m` is applied, and
-a viewer should draw the object once.
+Collaborative clips carry two agents sharing one object. Both agents store their own copy of the
+object track; they agree once each agent's `z_offset_m` is applied, and a viewer should draw the
+object once.
 
-## Licensing
+---
 
-This corpus is **CC BY-NC-SA 4.0**: non-commercial, attribution, and derivatives must carry the
-same licence. That is the most restrictive term among the sources, and it propagates.
+## Licensing
 
-| source dataset | its licence | note |
-|---|---|---|
-| OMOMO | MIT | the only permissively licensed source |
-| ParaHome | CC BY-NC-SA 4.0 | |
-| NeuralDome / HODome | CC BY-NC-SA 4.0 | |
-| CoRoleHOI | CC BY 4.0 | |
-| IMHD² | CC BY-NC-SA 4.0 | |
+**CC BY-NC-SA 4.0**: non-commercial, attribution, and derivatives carry the same licence. That is
+the most restrictive term among the sources, and it propagates.
 
-Retargeted trajectories are **derivative works** of the source motion capture, and the contact
-annotations they were optimised against originate with
-[InterAct](https://github.com/Sirui-Xu/InterAct) under CC BY-NC-SA 4.0. Cite the source dataset
-for any clip you use — the `dataset` column names it, and `NOTICE.md` gives the reference.
+| source dataset | its licence |
+|---|---|
+| OMOMO | MIT |
+| ParaHome | CC BY-NC-SA 4.0 |
+| NeuralDome / HODome | CC BY-NC-SA 4.0 |
+| CoRoleHOI | CC BY 4.0 |
+| IMHD² | CC BY-NC-SA 4.0 |
 
-**Not included**, for licensing reasons: BEHAVE and FORCE (MPI terms forbid redistribution of
-modified data), HUMOTO (Adobe Research License), CHAIRS, and PA-HOI / InterVLA / ReconMot
-(licence unverified).
+These trajectories are derivative works of the source motion capture, and the contact annotations
+they were optimised against originate with [InterAct](https://github.com/Sirui-Xu/InterAct) under
+CC BY-NC-SA 4.0. Cite the source dataset for any clip you use: the `dataset` column names it, and
+`NOTICE.md` gives the reference.
 
-Robot models are Unitree's, BSD-3-Clause — see `docs/LICENSE-unitree.txt`.
+Robot models are Unitree's, BSD-3-Clause — [`docs/LICENSE-unitree.txt`](docs/LICENSE-unitree.txt).
 
-**Why this is more restrictive than the code.** The [retargeting code](https://github.com/shinben0327/hoi-retarget) is
-BSD-3-Clause, because nothing in it derives from a restrictively licensed source. These
-trajectories do: they are optimised against source motion capture and its contact
-annotations, so the sources' terms propagate to them. The two licences are independent
-and both are correct.
+The [retargeting code](https://github.com/shinben0327/hoi-retarget) is BSD-3-Clause. Its terms
+differ from this corpus's because the code derives from nothing restrictively licensed and these
+trajectories do.
 
 ---
 
@@ -250,5 +252,5 @@ and both are correct.
 }
 ```
 
-Also cite the source dataset your clips come from — the `dataset` column names it,
-and `NOTICE.md` lists every reference.
+Cite the source dataset your clips come from as well — the `dataset` column names it, and
+`NOTICE.md` lists every reference.

docs/SCHEMA.md

@@ -91,7 +91,7 @@ All are lists of per-frame values, length `T`. Reconstruct with
 | `object_ground_contact_sequence` | (T,) int8 | object resting on the ground |
 | `object_near_floor_flags` | (T,) bool | object within a floor-proximity threshold |
 
-## Gotchas
+## 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.