charlesH777/WeldGrind-Dataset

WeldGrind Dataset

2800 synthetic rendered workpiece states from 190 base geometries for surface-contact grinding trajectory research. The six workpiece families are butt plates, T joints, L joints, boxes, cylinders and spheres.

Download

Download weldgrind_dataset_20261110_public_v1.zip (10.60 GB), together with its SHA256 checksum. Extraction requires approximately 10.7 GB of additional space. The archive contains the original prepared point clouds and 1024-point trajectory labels, original RGB-D, calibration and referenced metadata, actual SAM predictions, the selected SAM delta, a Chinese guide and six illustrated examples.

See the Chinese dataset guide, dataset metadata, and verification receipt.

Splits

Split	Samples	Composition
Train	2260	1000 previously seen replay states + 1260 new states
Validation	360	New states from 18 held-out base geometries
Test	180	New states from 9 held-out base geometries

The new training cohort uses 63 base geometries; each new base has 20 states. These new base geometries are disjoint across the three splits. The historical validation/test labels of the old 1000-state cohort are provenance records: those states have been seen and belong to training replay.

Use

Model and loader source: WeldGrind-V1. Python 3.10 or later is required.

sha256sum -c weldgrind_dataset_20261110_public_v1.zip.sha256
python -m zipfile -e weldgrind_dataset_20261110_public_v1.zip .
python -m pip install 'git+https://github.com/CharlesH777/WeldGrind-V1.git'
cd weldgrind_dataset_20261110_public_v1
python tools/verify_dataset.py

from weldgrind_v1.program_data import ProgramDataset

data = ProgramDataset('program_inputs/train.jsonl', seed=20261109,
                      max_points=2048, trajectory_points=1024)
sample = data[0]
print(len(data), sample['sample_id'], sample['trajectory'].shape)

The archive README describes the layout. tools/relocate.py prepares local absolute-path manifests when needed. The original observations, sample order, split definitions and original-GT labels are preserved byte for byte from the local dataset. Later canonical-supervision and process-depth augmentation caches are separate from this release.

World XYZ uses millimetres. Observation clouds are back-projected from noisy RGB-D with actual SAM regions. Target paths, phases, pass indices and source geometry metadata are supervision or provenance, not autonomous inference inputs. The XYZ paths are surface-contact references; execution infeed is a separate control quantity.

Scope

The data is synthetic and does not demonstrate real machining performance. Original missing initial approaches and final retracts have not been invented. Only source assets referenced by each prepared sample are included, rather than complete rendering projects. The selected SAM delta is included; SAM base weights and action-model checkpoints are separate. The PDF describes the earlier local package; this card and the archive README describe the public layout and code installation.