junbin666/isaac-blue-vial-sdg-contract

Isaac Blue Vial SDG Contract

A small, deterministic synthetic-data contract for a blue-vial localization demo. It exercises the dataset schema, domain-randomization fields, validation, training, and sim-to-real benchmark plumbing used by the Isaac Sim-to-Real VLA Lab.

This dataset is an offline procedural preview, not photorealistic NVIDIA Isaac Sim or Isaac Replicator output. It is intended for CI, education, and pipeline integration—not for measuring real-world robot performance.

Dataset structure

The dataset contains 100 images at 64×64 resolution:

Split	Frames	Positive	Negative
train	81	61	20
validation	10	10	0
test	9	9	0

Each JSONL row contains:

• image: browser-friendly RGB PNG (from image_file_name metadata)
• mask: single-channel semantic PNG (from mask_file_name metadata); empty for negative samples
• object_present: whether the blue vial is present
• bbox_2d_tight: [x_min, y_min, x_max, y_max], or null for negatives
• label: blue_vial, or null for negatives
• randomization: lighting, camera, object pose, and physics parameters
• frame_id and deterministic generator seed

Negative samples contain an amber distractor and no blue vial. They are kept in the training split so validation and test remain focused on positive-object localization. A future detection benchmark should add held-out negatives and report false-positive metrics separately.

Split integrity and quality

Exact positive mask/bounding-box groups are assigned as a unit: an identical annotation geometry never appears in more than one split. This removes the 13 cross-split groups found in the first audit. The packaged quality report is available in quality_report.json.

Current checks:

• 99 unique RGB images; the only duplicate pair is two training negatives whose quantized lighting produced the same procedural preview
• 51 unique positive annotation groups
• 0 positive annotation groups crossing split boundaries
• 80 positive and 20 negative frames
• deterministic generation with seed 7

Intended uses

• Validate object-detection and segmentation data loaders
• Exercise deterministic synthetic-data and CI pipelines
• Train a tiny bounding-box regression baseline
• Demonstrate mock-to-real boundaries in a robotics portfolio

Out-of-scope uses

Do not use this dataset as evidence of performance on real cameras, Isaac Replicator imagery, multiple objects, occlusion, safety-critical manipulation, or production robotics. The renderer is deliberately simple and the dataset is too small and visually narrow for real-world generalization.

Reproduction

Generator source: nemoclaw-arduino-robot-arm

.venv/bin/synthetic-data generate \
  --output artifacts/synthetic_dataset --frames 100 --seed 7

.venv/bin/python scripts/build_hf_dataset.py \
  artifacts/synthetic_dataset \
  huggingface-datasets/isaac-blue-vial-sdg-contract

Licensing

The generated dataset is released under CC BY 4.0. The source generator is licensed separately under the MIT License. No external images, camera captures, personal data, or third-party datasets are included.