3DA-VTG

3DA-VTG is a visuo-tactile grasp-stability dataset prepared for the public SGA-GSN release. It contains paired visual and tactile observations, object-level metadata, and binary grasp-stability labels.

License: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). The dataset is derived from GraspNet-1Billion assets and follows the GraspNet non-commercial redistribution terms. Commercial use requires permission from the GraspNet team.

Contents

• Full dataset shards: data/000.zip to data/087.zip.
• Split files: data/train.csv, data/test.csv, data/train-ids.txt, and data/test-ids.txt.
• Background image for 2D VTG loaders: data/bg_sim.jpg.
• Metadata and format documentation: metadata/.
• Mini dataset for smoke tests: samples/3da_vtg_mini.zip.

Current staging statistics:

Split	Objects	Samples
Train	68	318,532
Test	19	95,000
Total	87 objects with data / 88 shards	413,532

Object 046 is kept as an empty shard for release completeness but is not part of the train or test split.

Download And Restore

For full SGA-GSN use, download the data/ directory from this dataset repo. Then restore it with the SGA-GSN helper script:

bash /SGA-GSN/install/extract_3da_vtg.sh <downloaded_repo>/data /SGA-GSN/data

This creates:

/SGA-GSN/data/3DA-VTG

The SGA-GSN dataset configs expect data/3DA-VTG relative to the SGA-GSN repo root.

For a quick local smoke test, download and unzip:

unzip samples/3da_vtg_mini.zip -d /SGA-GSN/data

The mini dataset restores the same top-level 3DA-VTG/ directory structure.

File Format

Each object shard expands to one zero-padded object directory, for example 006/. Each object directory contains _metadata.json and sensor folders: tac_rgb, tac_dep, vis_rgb, vis_dep, and vis_seg.

Split CSV files use columns:

id,global_index

See metadata/file_format.md and metadata/sample_schema.json for details.

Utilities

The read-only helper script tools/dataset_utils.py can load restored samples and reconstruct visual/tactile point clouds from the released depth images and metadata:

from tools.dataset_utils import sample_to_pointclouds

pcs = sample_to_pointclouds("data/3DA-VTG", "006", "23476304")

The helper requires NumPy and OpenCV for data loading. Open3D is optional and is only imported for interactive 3D visualization. Complete object mesh point clouds still require the separate graspnet-vhacd asset package.

External Assets

This dataset repository does not include object mesh assets or model weights. SGA-GSN still requires:

• object meshes from the separate graspnet-vhacd asset package;
• the AdaPoinTr shape checkpoint ckpts/ap_ps55.pth from the model release.

License and Attribution

3DA-VTG uses CC BY-NC-SA 4.0 because it contains GraspNet-derived data, labels, scores, and object relationships. The GraspNet terms provided for this release state that all data, labels, code, and models belong to the GraspNet team, MVIG, SJTU, are available for free non-commercial use, and may be redistributed under the same non-commercial conditions.

Please cite and attribute both 3DA-VTG and the GraspNet source assets when using this dataset. For commercial use, follow the GraspNet commercial contact path: fhaoshu at gmail.com and CC lucewu at sjtu.edu.cn.

Integrity

Use the published checksums after download:

sha256sum --check checksums.sha256