Data / version /1.1 /README.md
JK2727's picture
Publish v1.1 single-KeyFrame human motion references
ccef179 verified
|
Raw History Blame
7.88 kB
# ManipSyn v1.1: single-KeyFrame human motion references
This package provides 75 optimized HandX skeleton references for downstream
retargeting and physical simulation: Bottle instances 3571, 3616 and 3635,
operations 000–004, seeds 42–46. Each candidate contains its own selected `best`
checkpoint from a 500-step DNO run with DDIM 10. Five seeds are retained for each
operation. No cross-seed winner or physical success label is assigned.
## Read the data
Only Python and NumPy are required. Paths in the consumption manifest and sample
index are relative to this package root. The complete package works without the
ManipSynthesis repository, HandX checkpoint, MANO model files or Torch.
```bash
python read_delivery.py /path/to/version/1.1 --verify
python read_delivery.py /path/to/version/1.1 --sample Bottle_3571_000_seed_42
```
```python
from pathlib import Path
from read_delivery import Delivery
data = Delivery(Path('/path/to/version/1.1'))
sample = data.load('Bottle_3571_000_seed_42')
hands = sample['joints_world_m'] # (60, 2, 21, 3); left, right
used = sample['hand_used'] # (2,), bool
time_s = sample['scene']['time_s'] # arange(60) / 30
roots = sample['scene']['object_root_poses']
link_pose = sample['scene']['link_poses_link_0']
affordance = sample['affordance']
```
`samples.jsonl` is the authoritative sample index. `schema.json` defines the
package manifest and sample records. `manifest.json` inventories all payload
files with SHA-256 and byte sizes; `SHA256SUMS` also binds the manifest itself.
`read_delivery.py --verify` validates the entire package, including asset
dependencies, object FK, KeyFrame preservation and HandX feature reconstruction.
## Layout
```text
manifest.json, schema.json, samples.jsonl, SHA256SUMS
read_delivery.py, requirements.txt
assets/Bottle/<instance>/mobility.urdf, textured_objs/...
Articulated/Bottle/<instance>/<operation>/
input/manifest.json, object_motion.json, resolved_affordance.json
input/condition.npz, scene.npz, canonical.json
candidates/seed_<seed>/motion_best.npz, handx_best.npz
candidates/seed_<seed>/metadata.json, quality.json
provenance/inputs/Bottle/<instance>/<operation>.tar.gz
provenance/candidates/<sample_id>/...
provenance/runtimes/<runtime>/snapshot.tar.gz
provenance/reports/... # historical matrix evidence
reports/export_validation.json # fresh export validation
```
## Coordinate and time contract
- All saved world positions are in metres in a right-handed Z-up frame. There
are 60 frames at 30 Hz. Sample timestamps are 0 through 59/30 seconds; the
nominal 60-frame playback duration is 2 seconds.
- `motion_best.npz` contains `left_joints_world_m` and `right_joints_world_m`,
both float32 arrays of shape `(60,21,3)`, plus scalar boolean
`left_hand_used` and `right_hand_used`. These bytes match the source results.
- HandX joint index 0 is the wrist. Finger chains, from wrist to tip, are:
thumb `[0,13,14,15,16]`, index `[0,1,2,3,17]`, middle `[0,4,5,6,18]`,
ring `[0,10,11,12,19]`, pinky `[0,7,8,9,20]`. This is the HandX joint order.
- Matrices `object_root_poses` and `link_poses_<link>` map local metre
coordinates to world coordinates: `p_world = R @ p_local + t`.
- The supplied URDF and OBJ files retain original asset units. Apply the
operation's `asset_scale_to_meters` to mesh vertices and URDF origin
translations. Revolute coordinates are radians; prismatic coordinates in
`scene.npz` are metres. Root translations and saved link poses are already in
metres and must not be scaled again. URDF visual origins still apply.
- Root quaternions in `object_motion.json` use xyzw order. `scene.npz` also
supplies all joint coordinates, including zero values for unspecified joints.
`input/manifest.json` gives their column order, types and units.
## Inputs, contacts and quality
`input/condition.npz` preserves the original input arrays, including static MANO
pose, beta, wrist transforms, vertices and joints in the target link's local
metre coordinates. Grasp keys and links are listed in `input/manifest.json`.
The MANO pose is a 45-dimensional axis-angle vector; wrist rotation is an
axis-angle vector, with wrist translation in link-local metres. The original
builder's parameters and evidence are preserved in the input provenance archive.
`input/manifest.json` also includes the operation family, slug, overall task
description and separate left/right hand instructions from the original input.
**The input `world_joints_m` is the static grasp reference propagated with the
object. It is not the optimized output. Read generated motion from each
candidate's `motion_best.npz`.** Only the first frame of each grasp interval is
hard-conditioned in this method. Intervals use `[start,end)` frame indices.
`resolved_affordance.json` supplies the frozen fingertip targets, local surface
normals, target link, finger, role and active interval. Transform points with
the saved link pose and normals with its rotation. These are intended contact
targets; they do not measure physical contact or force.
`handx_best.npz` contains normalized `(1,42,4,60)` features, the condition and its
boolean mask. The 42 joints are left then right; feature channels are xyz plus
the HandX scalar rotation feature. The scalar is not a robot joint angle or a
MANO pose parameter. Normalization arrays and the fixed canonical transform are
included for decoding without a model.
`quality.json` retains the source skeleton-qualified flag, hard-condition and
saved-replay errors, and reports frozen-target distances and motion speeds.
Distance diagnostics use frames 1–58. Interior speed uses transitions 1→2
through 57→58; endpoint speed uses 0→1 and 58→59. Source static research,
contact and geometry-certification flags are retained separately. A successful
read or a skeleton-qualified flag is not a physical feasibility label.
This release contains human skeleton references and static MANO input geometry.
Generated per-frame MANO meshes/parameters, robot qpos, dynamics rollouts,
contact forces and physical success certification have not been produced for
these 75 candidates. The current single-KeyFrame results have endpoint motion
jumps and did not outperform the no-KeyFrame control in the historical report.
## Provenance and verification scope
The source matrix has 150 groups: 75 no-KeyFrame controls and 75 single-KeyFrame
candidates. This release exports only the 75 single-KeyFrame candidates: 74
computed in the formal run and one adopted compatible pilot. Historical matrix
reports remain explicitly historical and include the control statistics.
Input archives preserve every original input file and original manifest bytes.
Candidate evidence includes the source completion marker, best state, compressed
best audit, saved replay record and experiment/config metadata. Runtime archives
preserve source snapshots and execution identity. Original JSON evidence may
contain the producer's old absolute paths; those are provenance labels, not
dependencies of the portable reader. Model weights are identified by their
original hashes and are not needed to consume this package.
Export validation freshly checks source hashes, exact coverage, saved arrays,
portable assets, independent NumPy feature-to-world reconstruction, URDF FK and
KeyFrame consistency. It does not rerun DNO, fit MANO or execute a physics engine.
Original object assets retain their upstream PartNet-Mobility identity and
terms; this export does not relicense them.
## Download from Hugging Face
```bash
hf download ManipSyn/Data --repo-type dataset \
--include 'version/1.1/**' --local-dir ./manipsyn-data
python ./manipsyn-data/version/1.1/read_delivery.py \
./manipsyn-data/version/1.1 --verify
```
For reproducible use, add `--revision <published-commit>` to the download command.