perceptpick / ASSET_BUNDLE.md
varunburde's picture
Initial dataset: PerceptPick benchmark assets in canonical tree layout
af1dd73

PerceptPick — pre-prepared assets

This bundle holds the URDF / VHACD / mesh assets for the YCB-V dataset across nine mesh sources (oracle CAD plus eight reconstruction methods), together with FoundationPose and MegaPose pose-estimator CSVs.

Drop into a perceptpick clone to skip Stage A (01_prepare_assets.py) and the FoundationPose / MegaPose pipelines:

git clone <perceptpick>
cd perceptpick

# 1. download the BOP YCBV test split (scenes 48-59 + models)
#    see the README's "Get the YCB-Video dataset" section.

# 2. unpack this bundle next to the repo
unzip perceptpick_assets.zip

# 3. wire the bundle into the expected path
mkdir -p assets
mv perceptpick_assets/ycbv assets/ycbv

After that, jump straight to Stage B / C — no Stage A re-prep needed.

Layout

ycbv/
├── GT/                                          # oracle CAD (BOP YCBV models)
│   ├── meshes/obj_NNNNNN.{obj,mtl,png}
│   ├── vhacd/obj_NNNNNN_vhacd.obj
│   ├── urdf/obj_NNNNNN.urdf
│   └── pose_estimates/
│       ├── FoundationPose.csv                   # FoundationPose on GT meshes
│       └── MegaPose.csv                         # MegaPose on GT meshes
├── BakedSDF/                                    # 8 reconstruction methods
│   ├── meshes/, vhacd/, urdf/
│   └── pose_estimates/
│       ├── FoundationPose.csv                   # FoundationPose on BakedSDF
│       └── MegaPose.csv                         # MegaPose on BakedSDF
├── MonoSDF/, Nerfacto/, Neuralangelo/
├── NGP/, RealCAP/, UniSurf/, VolSDF/

Each method folder is fully self-contained: the meshes the simulator loads, the URDFs and VHACDs the physics layer needs, and the pose CSVs that were generated using that mesh as the pose-estimator's reference model. The CSVs are tiny; the meshes / VHACDs make up almost all of the disk footprint.

URDF paths

URDFs reference the sibling collision mesh with a relative path: <mesh filename="../vhacd/obj_NNNNNN_vhacd.obj"/>. No absolute paths, no system-specific roots — the bundle is portable.

Running the benchmark

# Stage B — sample antipodal grasps + simulate, per (object, gripper) on the GT meshes
pixi run python scripts/02_grasp_sweep.py --dataset ycbv --mesh-source GT --n-grasps 5000

# Stage C, Condition 1 — Oracle / Oracle (ideal baseline)
pixi run python scripts/04_evaluate.py --dataset ycbv \
    --gt-mesh GT --est-mesh GT \
    --pose-csv FoundationPose.csv --gripper auto --workers 4 --resume --headless

# Stage C, Condition 3 — End-to-end realistic (BakedSDF mesh + BakedSDF-conditioned pose)
pixi run python scripts/04_evaluate.py --dataset ycbv \
    --gt-mesh BakedSDF --est-mesh BakedSDF \
    --pose-csv FoundationPose.csv --gripper auto --workers 4 --resume --headless

If you'd rather regenerate the assets from scratch (e.g. to verify VHACD parameters), ignore this bundle and run scripts/01_prepare_assets.py --dataset ycbv --all-mesh-sources.