Arthur12137/SoftVTBench-archive

SoftVTBench — archive

Frozen snapshot of everything that lived in Arthur12137/SoftVTBench before the 2026-08-13 re-release: the evaluation USD assets, the soft-body assets, and the first partial data drops.

This repo is not maintained. The current dataset is at Arthur12137/SoftVTBench.

Contents: eval-assets/, soft-assets/, object-rigid/, object-soft/, spatial-rigid/, spatial-soft/ (9319 files, 2.3 GB).

---

Original dataset card (kept verbatim)

SoftVTBench dataset card

This is the canonical dataset card for both Hugging Face and ModelScope. Retain license: other until the final asset terms and written redistribution permission described below are attached; do not replace it with the code repository's Apache-2.0 license.

Download

Until the asset permissions described below are complete, download only the trajectory/observation data:

pip install -U huggingface_hub
huggingface-cli download Arthur12137/SoftVTBench \
  --repo-type dataset \
  --include 'object-soft/*' --include 'spatial-soft/*' \
  --include 'object-rigid/*' --include 'spatial-rigid/*' \
  --local-dir ./SoftVTBench_data

Mainland-China mirror:

from modelscope import dataset_snapshot_download

dataset_snapshot_download(
    "Arthur12137/SoftVTBench",
    local_dir="./SoftVTBench_data",
    allow_patterns=[
        "object-soft/**",
        "spatial-soft/**",
        "object-rigid/**",
        "spatial-rigid/**",
    ],
)

Selective download for object-soft training data:

huggingface-cli download Arthur12137/SoftVTBench --repo-type dataset \
  --include 'object-soft/*' \
  --local-dir ./SoftVTBench_data

Contents

Folder	Role	Contents
object-soft/	train/eval	10 object-centric deformable tasks × 50 demonstrations
spatial-soft/	train/eval	10 spatial deformable tasks × 50 demonstrations
object-rigid/	train/eval	matched rigid LIBERO-object baseline
spatial-rigid/	train/eval	matched rigid LIBERO-spatial baseline
eval-assets/	eval	USD scene/deformable assets required to reconstruct closed-loop scenes
soft-assets/	authoring	source asset bundle; not needed by the public evaluator

Soft suites use nested task directories:

object-soft/libero_object/libero_object_task0/{replayed_demos,video_datasets}
spatial-soft/libero_spatial/libero_spatial_task0/{replayed_demos,video_datasets}

Rigid suites use top-level replayed_demos/ and video_datasets/. Each soft suite has one canonical 500-row manifest.jsonl whose HDF5/video paths are relative to the suite root. The current rigid folders do not contain manifest files. Detailed schemas live in the code repository under datasets/schemas/.

Supported use

The public code supports π0.5 vision-only and visuo-tactile fine-tuning plus closed-loop evaluation. SoftVTBench v1 reports:

• Goal Success: the simulator task predicate is satisfied for the terminal success horizon.
• Safe Success: Goal Success and peak canonical FEM deformation does not exceed the compression-sweep threshold for that object.

No separate NoDrop number is reported. See configs/benchmark_protocol_v1.json in the code repository for the machine-readable protocol.

Limitations

• Rollouts are simulated in Isaac Sim and do not measure sim-to-real transfer.
• The safety metric uses privileged FEM state and is intended for evaluation, not as a policy input.
• Thresholds are object-specific; results are invalid if thresholds and online rollouts use different deformation metric IDs.
• Rigid suites report Goal Success only.

Licensing and third-party assets

SoftVTBench-authored metadata and code do not relicense third-party scene, texture, tactile, LIBERO, Tabero, or NVIDIA assets. Publication of this metadata does not establish permission to redistribute eval-assets/ or soft-assets/. The maintainers must attach the applicable terms and written redistribution permission for every file in those folders. Until that evidence is present, do not download, mirror, or redistribute those asset folders. See docs/asset_licensing.md and THIRD_PARTY_NOTICES.md in the code release.

Citation

@article{jing2026softvtbench,
  title   = {SoftVTBench: A Safety-Aware Visuo-Tactile Benchmark for Physically Constrained Robotic Manipulation of Deformable Objects},
  author  = {Jing, Bowen and Wang, Mingxin and Hao, Ruiyang and Ge, Chenchen and Shen, Hanwen and He, Junjie and Cui, Yang and Hou, Yiming and Zhou, Weitao and Wang, Jiawei and Li, Minglei and Zhang, Dandan and Zhao, Ding and Liu, Houde and Li, Xiaofan and Liu, Si and Luo, Ping and Yu, Haibao},
  journal = {arXiv preprint arXiv:2607.04234},
  year    = {2026}
}