bartduis/pointzero-synthetic

PointZero Sparse Synthetic Pretraining

Sparse, training-ready synthetic trajectories associated with PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics.

Release candidate: uploading and validation in progress. This repository is private staging until its license and the full release checks are finalized. Check release-status.json for committed shard counts. The table below describes the frozen intended source inventory, not a claim that every shard is uploaded.

This release covers synthetic pretraining and its original synthetic evaluation splits only. It contains 290,097 intended training trajectories (2,900,970 frames), and 322,230 trajectories including evaluation. Real-world recordings, robot demonstrations, model predictions, and checkpoints are outside this release.

The default viewer shows 33 representative thumbnails from the validated pilot. These previews are separate from the full training/evaluation shards.

Contents and original splits

Component	Split	Source scene entries	Trajectories	Frames
shorts_v7	train	6,885	68,712	687,120
shorts_v7	test	20	200	2,000
tshirt_v7	train	6,593	65,788	657,880
tshirt_v7	test	20	200	2,000
towel_v7	train	5,426	54,190	541,900
towel_v7	test	20	199	1,990
genesis_partnet_50k	train	5,058	50,576	505,760
genesis_partnet_50k	test_similar	1,600	16,000	160,000
genesis_partnet_50k	test_unseen	1,534	15,339	153,390
kubric_50k_v2	train	5,179	50,831	508,310
kubric_50k_v2	test	20	195	1,950

A scene may contain multiple context trajectories. These are different counting units. Original scene split manifests are preserved under splits/, with SHA256 digests in manifests/source-inventory.json. PartNet's test_similar and test_unseen are separate splits.

Shard format

Each approximately 1 GB uncompressed TAR is a WebDataset shard. Members for one trajectory are adjacent and share a stable key:

data/shorts_v7/train/shard-000000.tar
    shorts_v7__000603__0000.points.npy
    shorts_v7__000603__0000.graspers.npy
    shorts_v7__000603__0000.rgb_000.png
    shorts_v7__000603__0000.segmentation_000.png
    ... remaining RGB and segmentation frames ...
    shorts_v7__000603__0000.json

Field	Meaning
points.npy	float32 array [T<=10, N<=1000, XYZ=3]; point index correspondence is preserved across time
graspers.npy	optional float32 array [G, T<=10, XYZ=3] containing recorded interaction tracks
rgb_000.png ... rgb_009.png	exact original PNG bytes for all ten timesteps
segmentation_000.png ... segmentation_009.png	exact original segmentation PNG bytes
json	component, split, source scene/context IDs, timestep names, image dimensions, sampling method, and hashes

The arrays preserve source training coordinates and float32 precision. The export does not normalize, transform, resample, or regenerate point tracks. This sparse release does not include original dense geometry, depth maps, camera calibration, or per-frame visibility labels. first_step names describe the existing sparse cache convention; they are not newly computed visibility annotations.

Rigid trajectories use the existing first-step FPS cache; the other components use the existing first-step sparse cache. Articulated and rigid trajectories normally have no recorded grasper array: PointZero generates conditioning tracks from moving points during batch preparation.

Some rigid trajectories contain fewer than 1000 points in their existing FPS cache. The export preserves those counts; batching uses the existing PointZero padding behavior.

Short source sequences

Some source cloth trajectories contain only 2–8 frames. Their available points, interaction tracks, and images are preserved exactly and their metadata records quality.default_training_eligible: false. The frames column of each Parquet index records the actual frame count. The supplied adapter excludes sequences shorter than ten frames from default training; setting training=False retains all available sequences for inspection. No missing frames are synthesized.

Until the full audit completes, frame totals in the inventory table are nominal counts at ten frames per trajectory. The final validation report will record actual frame totals and the number of short source sequences.

Download one component and train without extraction

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="bartduis/pointzero-synthetic",
    repo_type="dataset",
    allow_patterns=[
        "data/shorts_v7/train/*.tar",
        "manifests/shards/shorts_v7/train/*.json",
        "pointzero_shards.py",
    ],
    # Set revision to the published immutable release commit for reproducibility.
)

import sys
sys.path.insert(0, root)
from pointzero_shards import PointZeroShardDataset
from torch.utils.data import DataLoader

dataset = PointZeroShardDataset(root, components=["shorts_v7"], epoch_samples=10000)
# Use PointZero's existing utils.collate.collate_pad for mixed grasper counts.
from utils.collate import collate_pad
loader = DataLoader(dataset, batch_size=8, num_workers=4, collate_fn=collate_pad)
for epoch in range(10):
    dataset.set_epoch(epoch)
    for batch in loader:
        points = batch["current_scene"]["points"]
        # Pass the batch through the existing PointZero batch preparation/model.

epoch_samples is per distributed rank. Do not attach a DistributedSampler: the iterable adapter handles distributed/worker iteration. Training intentionally samples with replacement, with independently seeded worker streams; shuffle order differs from the original directory loader. Evaluation partitions shards across ranks and workers and traverses every sample once. Set training=False for exact traversal, including when inspecting a train split.

The adapter returns PointZero's current_scene and context structure. RGB and segmentation conditioning use the first timestep, matching the inspected loader; all other images remain available in the archive. The dataset supports standard WebDataset readers too. Parquet indexes provide each sample's shard, TAR byte offset, and byte length for local random access.

Component sampling

The adapter's default weights reproduce the inspected runnable configuration: shorts/T-shirt/towel each 0.33, articulated 1.0, rigid 0.33, normalized across selected components. These rounded weights differ from the paper's stated deformable/articulated/rigid mixture of 4/9, 4/9, 1/9. To use that stated mixture:

paper_weights = {
    "shorts_v7": 4/27, "tshirt_v7": 4/27, "towel_v7": 4/27,
    "genesis_partnet_50k": 4/9, "kubric_50k_v2": 1/9,
}
dataset = PointZeroShardDataset(root, weights=paper_weights)

Download every selected component's train shards before mixing components.

Provenance, license, and limitations

Deformable interactions use procedural clothing with FleX physics and Blender rendering. Articulated interactions use PartNet-Mobility objects in Genesis. Rigid interactions use modified Kubric scenes. See the paper for generation details. This is simulation data with sparse surface correspondences; it does not represent the full diversity or realism of physical interactions.

The release license is pending. No blanket license is assigned to third-party assets by this staging repository. PartNet-Mobility has noncommercial research and education terms, recipient-agreement requirements, and upstream ShapeNet terms; see the official dataset. See LICENSE.md and ATTRIBUTION.md for the component-specific release record.

Validation and versioning

The exporter checks 1–10 available frames per trajectory, finite float32 coordinates, consistent grasper counts, and matching RGB/segmentation dimensions. Only the explicit source-file allowlist is included. Every released member has a SHA256 digest in its sample metadata, and each shard/index has a digest in manifests/shards/. A pilot validated exact source equivalence, PNG integrity, indexed access, weighted training, and disjoint distributed evaluation.

The final release will record exact exported counts, any exclusions, and an immutable Hugging Face revision. Source files are left unchanged.

Citation

Please cite PointZero and the upstream dataset and generation resources relevant to the components you use. A final BibTeX entry will be included with the completed release.