zxie701/synthesis-retail-pickplace4tabero

Synthesis Retail Pick-place Assets for Tabero

Private, offline-ready repackaging of 133 user-selected Extwin Synthesis assets: 93 drink + 40 snack. Native USD/USDC layers and textures: 622 files, 244.79 MB (decimal; metadata/tools are additional).

Selection and preservation

• The user-provided asset_list.csv is the exact whitelist, reported by the user as already used on their replay and replacement-augmentation machine.
• Its SHA256 is 465eaffbb863e3939516d5f4296be1cdd576aa3dca915e4ca78e6d490059a34d. No extra assets were selected; none were dropped.
• IDs, names, CSV category labels, source URLs and usd_entry paths are preserved. In particular, existing CSV labels such as cereal001 being in drink are intentionally not reclassified.
• Only the listed entry URLs and their OpenUSD-discovered dependencies were downloaded.
• Native bytes, units, pivots, geometry, materials, collision and physics parameters are unmodified. There is no flattening, rescaling, collider replacement or wrapper conversion.
• Original replay-machine file hashes were not supplied. Therefore this is a re-download from the listed URLs, NOT a claim of byte identity with an older local snapshot. This snapshot is pinned using MANIFEST.json and a Hugging Face commit revision.
• Per-asset meta.json combines the supplied CSV row with download checksums and static validation; it does not pretend to be raw API metadata.

Download and use

Authenticate to Hugging Face with an account authorized for this private dataset. Use the immutable revision in the delivered hf_dataset.lock.json rather than relying on a moving main:

hf download zxie701/synthesis-retail-pickplace4tabero --repo-type dataset \
  --revision <COMMIT_FROM_HF_DATASET_LOCK> --local-dir /path/to/synthesis_assets
python /path/to/synthesis_assets/tools/verify_pool.py /path/to/synthesis_assets

HF_ENDPOINT must target an endpoint that supports this private repository and authentication. Once downloaded and verified, simulation uses local files only; Extwin access is unnecessary.

import json
from pathlib import Path
from pxr import Usd

root = Path('/path/to/synthesis_assets')
index = json.loads((root / 'INDEX.json').read_text())
asset = index['assets'][0]  # Or sample by asset['category'] == 'drink'/'snack'.
entry = root / asset['usd_entry']
stage = Usd.Stage.Open(str(entry))
assert stage is not None
# In an existing scene, reference entry onto the project's object prim as usual.

usd_entry is relative to this dataset root and exactly matches the input CSV. For example: usd/001db7d2-ba63-42da-b7fd-835635959ce6/usd/24fa58e35428a5fa/24fa58e35428a5fa/1/model_drink056.usd. All dependencies retain their native relative directory layout. Keep the whole per-ID directory. The included tools require Python 3.10+; downloading/static USD inspection additionally requires requests and usd-core (this package was checked with OpenUSD 26.8). The checksum verifier is standard-library only and works with the target Python 3.11.

Validation and limits

See download_report.json and validation_report.json. All selected assets were loaded with OpenUSD and independently checked with UsdUtils.ComputeAllDependencies, with no unresolved dependencies or composition errors. Recorded dimensions, rigid bodies, colliders and joints are descriptive, not additional filters. The default composed stages contain one rigid body in 124 assets and two in 8 assets. drink085 (f760fa91-f31a-4cfb-8436-3016e5a03c22) has 4 colliders but no authored PhysicsRigidBodyAPI detected in its default composed stage. It is deliberately retained, unmodified, because the supplied whitelist is authoritative; the task may add a body at spawn. This static observation is not a claim that the user's replay fails. See the per-asset report.

This workstation did not re-run Isaac Sim 5.1/Isaac Lab replay or manipulation. The user's prior task usage and this package's static validation are different evidence levels. Official Synthesis documentation targets Isaac Sim 5.0+ / Kit 107.3+; target-version rendering, PhysX schemas, replay compatibility and dynamics are not proven by a generic OpenUSD load. Native assets have authored physical parameters that should not be assumed empirically calibrated. For movable pick-place objects use scene support such as a ground/table; do not fix a target object to the world merely to prevent falling.

Provenance and license

Creator/source: Extwin / Extwin Synthesis.

• Synthesis asset service
• Original HF dataset
• Official explorer
• Pinned upstream README
• Per-asset attribution

Assets remain CC BY-NC 4.0. The full legal text is in LICENSE. This private repackaging does not grant commercial rights or imply Extwin endorsement. Noncommercial scope must be assessed for the actual project; internal company R&D is not automatically noncommercial. Original notices, including authored copyright/doc metadata, are preserved unchanged. Changes: selection, directory packaging, CSV/index/metadata, checksums and tools only; native asset files were not modified. Private access is not a replacement license and no additional restriction on licensed rights is asserted.

Layout

asset_list.csv                exact source whitelist
INDEX.json                    local sampling index
categories.json               upstream Retail tree + preserved CSV sampling labels
selection_manifest.json       explicit no-additions/no-removals selection provenance
download_report.json          download completeness
validation_report.json        OpenUSD static inspection (not dynamics)
MANIFEST.json / manifest.sha256
LICENSE / ATTRIBUTION.md / source_evidence/
tools/                        downloader, offline verifier and tests
usd/<asset_id>/meta.json
usd/<asset_id>/<native server-relative directory layout>

No credentials, dependency environment, caches, work logs or incomplete downloads are included.