Physical AI Data Factory - Simulation

Synthetic data generation for physical-AI use cases. Includes Defect Image Generation (DIG) on printed circuit boards.

Container: ${SDG_IMAGE}

Prerequisites

Clone tested repo

cd ~
git clone https://github.com/NVIDIA/paidf-simulation.git
cd paidf-simulation
git checkout main

Download USD assets

Obtain the USD asset bundle from your project channel. Extract anywhere convenient, then point PCB_USD_PATH at the main scene USD:

unzip <usd-assets>.zip -d <path/to/usd-assets>
export PCB_USD_PATH=<path/to/usd-assets>/spark_lighting.usd

The flow run commands below mount $(dirname $PCB_USD_PATH) read-only into the container; USDs don't need to live inside the repo.

Pull image

IMAGE=${SDG_IMAGE}
docker login nvcr.io
docker pull $IMAGE

Run the pipelines

Flow 1: Good Image Pipeline

Prepare the configuration

Config Path: configs/flow1_good_image/good_image.yaml

Key	Description
`scan_grid.x_num`, `y_num`	Number of cells along each axis. Default 10 by 10 produces 100 frames per trigger
`resolution`	Render resolution `[W, H]`. Default `[1920, 1080]`
`pathtracing.total_spp`	Accumulated samples per pixel. Higher values produce cleaner images and longer render times. Default: 32
`lighting.ring_light`	`true`: per-layer RGB ring light (soldering light). `false`: white light only
`writer.{rgb, bounding_box_2d_tight, semantic_segmentation, ...}`	Per-annotator on/off switches
`rename_to_grid_index`	Default `false` keeps `_NNNN.png` naming. Do not change for standard SDG flows

Config Path: configs/pcba_target.yaml

Key	Description
`component_types`	Component scope names under `pcba_root`. Only listed scopes receive semantic labels. The good-image pipeline does not use this file

Run the pipeline

# Flow 1 — Good Image Pipeline
PCB=~/paidf-simulation
IMAGE=${SDG_IMAGE}
OUTPUT=$PCB/sdg_test_output/flow1_good_image

# Pre-flight
mkdir -p $OUTPUT && chmod 777 $OUTPUT
ls /usr/share/nvidia/nvoptix.bin

# Run (Stage 1: Kit render)
docker run --rm --gpus all --network host \
  -v /usr/share/nvidia/nvoptix.bin:/usr/share/nvidia/nvoptix.bin:ro \
  -v $(dirname $PCB_USD_PATH):$(dirname $PCB_USD_PATH):ro \
  -e PCB_USD_PATH=$PCB_USD_PATH -e PAIDF_SIM_ROOT=$PCB \
  -v $PCB:/workspace/paidf-simulation \
  $IMAGE \
  "scripts/sdg/standalone/sdg_pipeline.py \
    --config configs/flow1_good_image/good_image.yaml \
    --pcba-config configs/pcba_target.yaml"

Verify the output

$ ls $OUTPUT/trigger_0000/ | sed 's/.*\.//' | sort | uniq -c
    301 json    # 100 bbox_labels + 100 bbox_prim_paths + 100 semseg_labels + metadata.json
    100 npy     # bbox_2d_tight arrays
    200 png     # 100 rgb + 100 semantic_segmentation
      1 txt     # metadata.txt

$ ls $OUTPUT/trigger_0000/rgb_*.png | head -3
rgb_0000.png  rgb_0001.png  rgb_0002.png    # _NNNN naming, no _x*_y*

$ cat $OUTPUT/trigger_0000/semantic_segmentation_labels_0000.json
{"(0, 0, 0, 0)":   {"class": "BACKGROUND"},
 "(0, 0, 0, 255)": {"class": "UNLABELLED"},
 "(33, 243, 3, 255)": {"class": "capacitor"}}
# RGBA values may change; assignment is randomized per run.

Flow 2: Defect Image Pipeline (Missing, Shift, Sideflip, Tombstone)

Prepare the configuration

Config Path: configs/flow2_defect_image/defect_image.yaml

Key	Description
`defects.shift.{enabled, ratio, translate_range, rotate_z_range}`	XY translation and Z-axis rotation defects
`defects.tombstone.{enabled, ratio, angle_min, angle_max}`	Tilt around Y axis (tombstone)
`defects.sideflip.{enabled, ratio, angle_min, angle_max}`	Flip around X axis
`writer.semantic_types`	Must include `defect` for defect labels to appear in semantic segmentation output
`scan_grid.x_num/y_num`, `resolution`, `lighting`, `pathtracing`	Same as Flow 1

Config Path: configs/flow2_defect_image/missing_image.yaml

Key	Description
`missing.ratio`	Fraction of component pool to hide per trigger (0–1)
`writer.reference.{rgb, semantic_segmentation, ...}`	Pass 1: all components visible. Segmentation labels mark hidden components with `defect=missing`
`writer.defective.rgb`	Pass 2: selected components hidden; RGB output only

Run the pipeline

# Flow 2 — Defect Image Pipeline
PCB=~/paidf-simulation
IMAGE=${SDG_IMAGE}

# Pre-flight
mkdir -p $PCB/sdg_test_output/flow2_defect_image \
         $PCB/sdg_test_output/flow2_missing_image
chmod 777 $PCB/sdg_test_output/flow2_defect_image \
          $PCB/sdg_test_output/flow2_missing_image

# Run (a) pose defects: shift / tombstone / sideflip
docker run --rm --gpus all --network host \
  -v /usr/share/nvidia/nvoptix.bin:/usr/share/nvidia/nvoptix.bin:ro \
  -v $(dirname $PCB_USD_PATH):$(dirname $PCB_USD_PATH):ro \
  -e PCB_USD_PATH=$PCB_USD_PATH -e PAIDF_SIM_ROOT=$PCB \
  -v $PCB:/workspace/paidf-simulation \
  $IMAGE \
  "scripts/sdg/standalone/sdg_pipeline.py \
   --config configs/flow2_defect_image/defect_image.yaml \
   --pcba-config configs/pcba_target.yaml"

# Run (b) missing components
docker run --rm --gpus all --network host \
  -v /usr/share/nvidia/nvoptix.bin:/usr/share/nvidia/nvoptix.bin:ro \
  -v $(dirname $PCB_USD_PATH):$(dirname $PCB_USD_PATH):ro \
  -e PCB_USD_PATH=$PCB_USD_PATH -e PAIDF_SIM_ROOT=$PCB \
  -v $PCB:/workspace/paidf-simulation \
  $IMAGE \
  "scripts/sdg/standalone/sdg_pipeline.py \
   --config configs/flow2_defect_image/missing_image.yaml \
   --pcba-config configs/pcba_target.yaml"

Verify the output

(a) Pose defects: output under flow2_defect_image/trigger_0000/

$ ls $OUTPUT/flow2_defect_image/trigger_0000/ | sed 's/.*\.//' | sort | uniq -c
    301 json
    100 npy
    200 png
      1 txt
# 100 of each: rgb / semseg / semseg_labels / bbox_npy / bbox_labels / bbox_prim_paths

$ cat $OUTPUT/flow2_defect_image/trigger_0000/semantic_segmentation_labels_0000.json
# defect classes appear (cells without that defect won't show all three keys):
{"(0, 0, 0, 0)":   {"class": "BACKGROUND"},
 "(0, 0, 0, 255)": {"class": "UNLABELLED"},
 "(33, 243, 3, 255)":  {"defect": "sideflip"},
 "(240, 4, 111, 255)": {"defect": "shift"},
 "(27, 186, 239, 255)":{"defect": "tombstone"}}

(b) Missing components: output under flow2_missing_image/trigger_0000/

$ ls $OUTPUT/flow2_missing_image/trigger_0000/reference/ | sed 's/.*\.//' | sort | uniq -c
    300 json    # 100 bbox_labels + 100 bbox_prim_paths + 100 semseg_labels
    100 npy     # bbox arrays
    100 png     # 100 colorized semantic_segmentation (rgb off in reference)
      1 txt

$ ls $OUTPUT/flow2_missing_image/trigger_0000/defective/ | sed 's/.*\.//' | sort | uniq -c
    100 png     # rgb only
      1 txt

$ cat $OUTPUT/flow2_missing_image/trigger_0000/reference/semantic_segmentation_labels_0000.json
# Hidden components labeled with defect=missing
{"(0, 0, 0, 0)":   {"class": "BACKGROUND"},
 "(0, 0, 0, 255)": {"class": "UNLABELLED"},
 "(33, 243, 3, 255)": {"defect": "missing"}}

Between Pass 1 and Pass 2, the pipeline log includes [Pipeline] Hiding N components. For example, with missing.ratio: 0.5, about 1138 of ~2276 components in the pool may be hidden.

Flow 3: Good and Defect Pairs Dataset

Paired golden / defect data for ChangeNet-style training is handled by the simulation skill's paired sub-mode of the single-flow track. See skills/simulation/SKILL.md for routing detail; the implementation runs the good flow and the defect flow with the same random_seed and post-processes via scripts/postprocess/build_pair_dataset.py.

Flow 4: USD2ROI Day-1

Prepare the configuration

Config Path: configs/cad2roi/day1/replicator/usd2roi_target.yaml

Key	Description
`scene`	CAD-derived USD path relative to repo root
`real_image`	Real PCB photo path relative to repo root
`semantics: [{match, labels}, ...]`	Prim-path glob to label rules (author per board)
`camera.translate`	`[x, y]` ortho camera center in mm (z fixed at 5000)
`resolution`	Match the real photo aspect ratio so post-MI scale factors `sX` and `sY` are close to 1.0
`registration.sx_range / sy_range / rot_range_deg / shift_range`	MI search ranges. Tighten if you have priors
`registration.min_mi`	Stage 2 exits with code 2 if `mi_after` < this. Default 0.5
`crop.classes`	Class labels to extract ROIs for, e.g. `[capacitor, solder, pad, ic]`
`crop.bridge / bridge_dis / bridge_classes`	Enable bridge crops, pixel distance threshold, and class pairs to bridge
`output.dir`	Container-absolute path, e.g. `/workspace/paidf-simulation/sdg_test_output/flow4_day1_rois`

Prepare the real-image input

Download the sample image (screenshot from an AOI machine): https://drive.google.com/file/d/18rzCtpPgn7paNGv8AtN-xu9ZEcEKyk5c/view?usp=share_link

cd ~/paidf-simulation
mv ~/Downloads/real.png ./scripts/usd2roi/input

Run the scripts

Step 1: Pre-flight

PCB=~/paidf-simulation
IMAGE=${SDG_IMAGE}
OUTPUT=$PCB/sdg_test_output/flow4_day1_rois
YAML=configs/cad2roi/day1/replicator/usd2roi_target.yaml

mkdir -p $OUTPUT && chmod 777 $OUTPUT
ls /usr/share/nvidia/nvoptix.bin

Step 2: Stage 1 render

docker run --rm --gpus all --network host \
  -v /usr/share/nvidia/nvoptix.bin:/usr/share/nvidia/nvoptix.bin:ro \
  -v $(dirname $PCB_USD_PATH):$(dirname $PCB_USD_PATH):ro \
  -e PCB_USD_PATH=$PCB_USD_PATH -e PAIDF_SIM_ROOT=$PCB \
  -v $PCB:/workspace/paidf-simulation \
  $IMAGE \
  "scripts/usd2roi/usd2roi_render.py --config $YAML"

Step 3: Stage 2 register

docker run --rm --gpus all --network host \
  -v $PCB:/workspace/paidf-simulation \
  --entrypoint python3 \
  $IMAGE \
  scripts/usd2roi/usd2roi_register.py --config $YAML

Step 4: Stage 3 crop

docker run --rm \
  -v $PCB:/workspace/paidf-simulation \
  --entrypoint python3 \
  $IMAGE \
  scripts/usd2roi/usd2roi_crop.py --config $YAML

Verify the output

$ ls $OUTPUT/
sdg/  aligned/  crop/

$ ls $OUTPUT/sdg/
rgb_0000.png  semantic_segmentation_0000.png  semantic_segmentation_labels_0000.json
metadata.txt  semantic_stats.json

$ python3 -m json.tool $OUTPUT/aligned/params.json
{
    "scaleX":       n,
    "scaleY":       n,
    "rotation_deg": n,
    "tx":           n,
    "ty":           n,
    "mi_before":    n,
    "mi_after":     n
}
# n values change per run.
# Open $OUTPUT/aligned/blink.gif to visually QA ref ↔ aligned alternation.

$ ls $OUTPUT/aligned/
ref_crop.png  aligned_crop.png  blink.gif  params.json
semantic_segmentation_0000.png  semantic_segmentation_labels_0000.json
sdg_crop_stats.json  metadata.txt

$ echo "ROIs:    $(ls $OUTPUT/crop/component/normal_img/*.png | wc -l)"
$ echo "Bridges: $(ls $OUTPUT/crop/bridge/normal_img/*.png  | wc -l)"
ROIs:    24
Bridges: 2

Flow 5: USD2ROI Day-0

Prepare the configuration

Config Path: configs/cad2roi/day0/sdg/day0_image.yaml

Run the scripts

Step 1: Pre-flight

PCB=~/paidf-simulation
IMAGE=${SDG_IMAGE}
OUTPUT=$PCB/sdg_test_output/flow5_day0_rois

mkdir -p $OUTPUT && chmod 777 $OUTPUT
ls /usr/share/nvidia/nvoptix.bin

Step 2: Stage 1 render

docker run --rm --gpus all --network host \
  -v /usr/share/nvidia/nvoptix.bin:/usr/share/nvidia/nvoptix.bin:ro \
  -v $(dirname $PCB_USD_PATH):$(dirname $PCB_USD_PATH):ro \
  -e PCB_USD_PATH=$PCB_USD_PATH -e PAIDF_SIM_ROOT=$PCB \
  -v $PCB:/workspace/paidf-simulation \
  $IMAGE \
  "scripts/sdg/standalone/sdg_pipeline.py \
    --config configs/cad2roi/day0/sdg/day0_image.yaml \
    --pcba-config configs/pcba_target.yaml"

Step 3: Set permissions between stages

docker run --rm \
  -v $PCB:/workspace/paidf-simulation \
  --entrypoint chmod \
  $IMAGE 777 /workspace/paidf-simulation/sdg_test_output/flow5_day0_rois

Step 4: Anchor crop (Stage 2)

docker run --rm \
  -v $PCB:/workspace/paidf-simulation \
  --entrypoint python3 \
  $IMAGE \
  scripts/usd2roi/usd2roi_crop.py --config configs/cad2roi/day0/usd2roi/day0_crop.yaml

Verify the output

$ ls $OUTPUT/
trigger_0000/  crop/

# Stage 1 — labelled scan_grid render (rename_to_grid_index: true)
$ ls $OUTPUT/trigger_0000/ | head -3
rgb_x0_y0.png  rgb_x0_y1.png  rgb_x0_y2.png   # _x*_y* spatial naming

$ ls $OUTPUT/trigger_0000/ | sed 's/.*\.//' | sort | uniq -c
    101 json    # 100 semseg_labels + 1 metadata.json
    200 png     # 100 rgb_x*_y* + 100 semantic_segmentation_x*_y*
      1 txt     # metadata.txt

# Stage 2 — multi-cell anchor crop
$ ls $OUTPUT/crop/component/ | head -5
x0_y0  x0_y1  x0_y2  x0_y3  x0_y4   # 100 cell directories total

$ echo "Total ROIs: $(find $OUTPUT/crop/component -path '*/normal_img/*.png' | wc -l)"
Total ROIs: 2150

$ ls $OUTPUT/crop/component/x0_y0/
normal_img/  cad_mask/  semantic_segmentation_labels.json

$ cat $OUTPUT/crop/component/x0_y0/semantic_segmentation_labels.json
{"(33, 243, 3, 255)":  {"class": "pad"},
 "(240, 4, 111, 255)": {"class": "capacitor"},
 "(27, 186, 239, 255)":{"class": "solder"}}
# RGBA values may change per run.

Wall-clock timings (one L40 GPU)

Approximate run times on a single NVIDIA L40 GPU:

Flow	scan_grid / capture	Time
1: Good	10×10 = 100 frames	~12 min
2a: Defect (pose)	10×10 = 100 frames	~19 min
2b: Missing	10×10 × 2 pass	~18 min
3: Pairs, Mode A (sequential)	10×10 + post-process	~24 min
4: Day-1 ROIs	1 frame + register + crop	~6 min
5: Day-0 ROIs	10×10 + crop	~10 min

Contributing

External contributions are welcome. All commits must be signed off under the Developer Certificate of Origin (DCO); see CONTRIBUTING.md for details.

License

Source code in this repository is licensed under the Apache License, Version 2.0; see LICENSE. Third-party runtime dependencies and their licenses are documented in third_party/ (see third_party/licenses.txt for the auto-generated inventory and third_party/README.md for notes on specific entries).

Name		Name	Last commit message	Last commit date
.agents		.agents
.claude		.claude
.codex		.codex
.cursor/rules		.cursor/rules
configs		configs
docs		docs
scripts		scripts
skills/simulation		skills/simulation
third_party		third_party
.dockerignore		.dockerignore
.gitignore		.gitignore
CONTRIBUTING.md		CONTRIBUTING.md
Dockerfile		Dockerfile
LICENSE		LICENSE
README.md		README.md
SECURITY.md		SECURITY.md
pyproject.toml		pyproject.toml
requirements.txt		requirements.txt

Navigation Menu

Uh oh!

Repository files navigation

Physical AI Data Factory - Simulation

Prerequisites

Clone tested repo

Download USD assets

Pull image

Run the pipelines

Flow 1: Good Image Pipeline

Prepare the configuration

Run the pipeline

Verify the output

Flow 2: Defect Image Pipeline (Missing, Shift, Sideflip, Tombstone)

Prepare the configuration

Run the pipeline

Verify the output

Flow 3: Good and Defect Pairs Dataset

Flow 4: USD2ROI Day-1

Prepare the configuration

Prepare the real-image input

Run the scripts

Verify the output

Flow 5: USD2ROI Day-0

Prepare the configuration

Run the scripts

Verify the output

Wall-clock timings (one L40 GPU)

Contributing

License

About

Releases

Packages

Contributors

Languages

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

Physical AI Data Factory - Simulation

Prerequisites

Clone tested repo

Download USD assets

Pull image

Run the pipelines

Flow 1: Good Image Pipeline

Prepare the configuration

Run the pipeline

Verify the output

Flow 2: Defect Image Pipeline (Missing, Shift, Sideflip, Tombstone)

Prepare the configuration

Run the pipeline

Verify the output

Flow 3: Good and Defect Pairs Dataset

Flow 4: USD2ROI Day-1

Prepare the configuration

Prepare the real-image input

Run the scripts

Verify the output

Flow 5: USD2ROI Day-0

Prepare the configuration

Run the scripts

Verify the output

Wall-clock timings (one L40 GPU)

Contributing

License

About

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages