Mesh-to-USD Setup
SkillCloud & infraPrepares your repo's mesh-to-USD environment by checking inputs, Docker images, GPU access, and model weights.
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the Mesh-to-USD Setup skill
About this capability
Prepare this repository's mesh-to-USD generation, structural validation, and Isaac Sim drop-test environment. Use when a user asks to build the mesh-to-USD images, verify Docker/GPU access, prepare FoundationPose support calibration, check model weights, validate a standalone GLB or exported HOI seq
What this skill tells your AI
The instructions your AI receives, as published by nvidia-isaac/video_to_data in .claude/skills/mesh-to-usd-setup/SKILL.md and read by ahel’s review.
Prepare the workflow and prove that the requested input mode is runnable. Work
from reconstruction/; keep Isaac Sim, OpenUSD, CoACD, and FoundationPose
dependencies inside their existing containers.
Act before asking
- Inspect the checkout, Python, Docker, GPU, disk, images, and supplied input immediately.
- Default to GPU 0 when the user has not constrained GPU use.
- Use
modules/v2d_hoi_object_reconstruction/mesh_to_usd/tests/data/toy_airplane/einstar/output_aligned.glbfor a non-destructive input check when no asset was supplied. - Build only missing images. Add FoundationPose only for cross-mesh recorded support, not for a standalone mesh or an exact recorded mesh with poses.
- Do not ask about mass, friction, output paths, or drop-test tuning during
setup; those belong to
mesh-to-usd-run. - Never accept the NVIDIA Isaac Sim EULA on the user's behalf. Existing
ACCEPT_EULA=Yor an explicit user instruction is sufficient for a later generation or drop-test run. - Ask only when credentials, EULA acceptance, missing external data, or another genuinely user-only action blocks progress.
Establish the checkout
REPO_ROOT="$(git rev-parse --show-toplevel)"
cd "$REPO_ROOT/reconstruction"
git status --short
python --version
docker version
nvidia-smi
df -h .
Treat
modules/v2d_hoi_object_reconstruction/mesh_to_usd/README.md and the current
CLI help as authoritative:
python modules/v2d_hoi_object_reconstruction/mesh_to_usd/build.py --help
python modules/v2d_hoi_object_reconstruction/mesh_to_usd/run_mesh_to_usd_workflow.py --help
python modules/v2d_hoi_object_reconstruction/mesh_to_usd/run_drop_test.py --help
Preserve unrelated local changes and existing results.
Determine the input contract
Choose the contract from the supplied paths without interviewing the user:
- Standalone mesh: one nonempty USD/USDA/USDC, OBJ, FBX, glTF/GLB, or STL asset. Its physical scale must already be correct.
- Exact exported sequence:
object_mesh/output_aligned.glb,poses.npy, andground_plane.json. - Catalog or cross-mesh support: a target mesh beside
output_symmetry.json, plus a sequence containingedex, per-camera RGB, depth and mask H5 sources, andground_plane.json. - Batch: catalog directories joined to exported sequences by exact
object.idinhoi_metadata.yaml.
Fail on a missing required file. Do not silently substitute geometry-derived support for missing recording data.
Build only missing images
Normal generation plus validation uses:
v2d_hoi_mesh_to_usd
v2d_hoi_mesh_to_usd_validator
Inspect them with docker image inspect. Build both when either is missing:
python modules/v2d_hoi_object_reconstruction/mesh_to_usd/build.py
Use --target generator or --target validator only for a deliberately scoped
repair. Use --dev at runtime only when validating local source changes.
Cross-mesh support additionally needs FoundationPose:
python modules/v2d_foundation_pose/docker/build.py
python modules/v2d_foundation_pose/docker/run_download_weights.py \
--output_dir data/weights/foundationpose
Do not download FoundationPose weights for modes that do not invoke it.
Prove readiness
Before declaring setup complete, show PASS, FAIL, or BLOCKED for each applicable gate:
-
All host
--helpcommands above succeed. -
The selected input contract exists and every required file is nonempty.
-
v2d_hoi_mesh_to_usdandv2d_hoi_mesh_to_usd_validatorexist. -
GPU 0 is visible inside the generator image:
docker run --rm --gpus '"device=0"' --entrypoint nvidia-smi \ v2d_hoi_mesh_to_usd -
Cross-mesh mode, when selected, has the FoundationPose image, populated weights, and complete per-camera inputs.
-
The user has explicitly accepted the EULA or will provide
ACCEPT_EULA=Ybefore an Isaac Sim command. Treat this as a run gate, not a reason to skip other setup checks.
Continue with mesh-to-usd-run when the request also includes generation or a
drop test. Do not stop after setup merely because the full simulation is long.
Signals
- GitHub stars
- 649
- Forks
- 62
- Last commit
- Sep 2026
Advanced
- Catalog kind
- skill
- Gateway key
mesh-to-usd-setup- Source
- github.com/nvidia-isaac/video_to_data
github.com/nvidia-isaac/video_to_data