SGA-GSN Model Weights

This staging repository contains inference-only checkpoints for SGA-GSN grasp-stability prediction and representative baselines. It is intended to accompany the public SGA-GSN code repository and the 3DA-VTG dataset release.

License: MIT. This follows the upstream AdaPoinTr/PoinTr license observed in /AdaPoinTr/LICENSE (Copyright (c) 2021 Xumin Yu). The staged model weights are released under the same MIT terms as the public SGA-GSN code. Dataset and simulation assets are separate releases with their own license terms.

Checkpoints

Release file	Role	Epoch	Main metrics
checkpoints/shape_completion/ap_ps55.pth	AdaPoinTr shape completion checkpoint required by 3D SGA-GSN and RL perception	512	F-Score 0.7315248407, CDL1 8.8670780700
checkpoints/sga_gsn/ppct_bce_vtencdrop_8dep_vds_featex_best.pth	Main PPCT/SGA-GSN grasp-stability checkpoint used by the RL stack	22	AvgAcc 0.7915571520, F1 0.8032546122
checkpoints/baselines/dgcnn_14m_best.pth	Smaller 3D DGCNN baseline	28	AvgAcc 0.7879940271, F1 0.8019056932
checkpoints/baselines/dgcnn_45m_best.pth	Larger 3D DGCNN baseline	29	AvgAcc 0.7860487511, F1 0.7978581551
checkpoints/baselines/cnn_best.pth	2D CNN baseline	10	AvgAcc 0.7031855615, F1 0.6832639230
checkpoints/baselines/cnnmca_best.pth	2D CNNMCA / MCACNN baseline	22	AvgAcc 0.7680102085, F1 0.7633521055

All checkpoints are inference_only files. They keep base_model, epoch, metrics, and best_metrics; optimizer states are intentionally omitted.

Configs

The configs/ directory contains release snapshots for the staged checkpoints:

configs/VTG_PPCT.yaml
configs/VTG_DGCNN_14M.yaml
configs/VTG_DGCNN_45M.yaml
configs/VTG_CNN.yaml
configs/VTG_CNNMCA.yaml
configs/dataset_configs/

VTG_PPCT.yaml is aligned to the source experiment config used by the released PPCT checkpoint, because the shorter public training config does not load this checkpoint exactly. The other release configs are copied from the public SGA-GSN code tree. For the 2D CNN and CNNMCA release configs, pretrained is set to False to avoid torchvision downloads before loading the released checkpoint. Loading the checkpoint restores the trained backbone weights.

Usage Boundary

This model repo does not include code, datasets, mesh assets, or simulation assets. A working setup requires:

• SGA-GSN code: https://github.com/Yiju1213/SGA-GSN.git
• 3DA-VTG dataset restored to the path expected by the selected config
• VT-Grasp simulation assets when the checkpoints are used through the RL stack
• PyTorch/CUDA environment compatible with the public SGA-GSN Dockerfile

The RL stack currently expects the PPCT checkpoint and shape checkpoint together:

shape_checkpoint: checkpoints/shape_completion/ap_ps55.pth
grasp_checkpoint: checkpoints/sga_gsn/ppct_bce_vtencdrop_8dep_vds_featex_best.pth

Verification

From the repository root:

sha256sum -c checksums.sha256

The staged checkpoints were also loaded through the public /SGA-GSN model builders in the original vt-grasp-0 container.

Metadata

• metadata/checkpoint_manifest.json records release paths, source paths, file sizes, SHA256 values, source metrics, and omitted keys.
• metadata/metrics.json provides a compact metrics table for the staged checkpoints.
• metadata/training_provenance.yaml and metadata/dependencies.yaml document source code, data, and runtime dependencies.