ArtiMo Benchmark Release

This package contains the 225-case benchmark data. It is meant to be used with the ArtiMo code repository.

Set paths first:

export ARTIMO=/path/to/ArtiMo
export BENCH=/path/to/benchmark_release
cd "$ARTIMO"

Download the source datasets from:

• PartNet-Mobility: https://sapien.ucsd.edu/browse
• ArtVIP: https://huggingface.co/datasets/X-Humanoid/ArtVIP/tree/main
• Lightwheel sim-ready assets: https://github.com/LightwheelAI/Lightwheel-simready-asset?tab=readme-ov-file

Then point the preparation script at those downloaded source roots. PartNet can be passed either as the extracted dataset directory or as the downloaded zip. ArtVIP and Lightwheel can be passed as raw USD datasets; the script converts only the benchmark assets it needs into ArtiMo's URDF format.

The folder name under DATA is the benchmark asset_name, not the original dataset id. The manifest maps each raw dataset asset to the benchmark folder name and split:

asset_name,asset_collection,source_dataset,source_asset,source_asset_dir,source_file
bin1,causal_data,PartNet-Mobility,102186,102186_trashcan,PartNet-Mobility/dataset/102186
microwave1,causal_data,ArtVIP,microwave_door_13,,ArtVIP/Articulated_objects/small_appliances/microwave/microwave_1/model_microwave_1.usd
electric_kettle1,causal_data,Lightwheel,electric_kettle1,,Lightwheel/Lightwheel_OpenSource/Manipulation/ElectricKettle001/ElectricKettle001.usd

For PartNet-Mobility, source_asset is the raw numeric folder id. The optional source_asset_dir column is only for already-renamed local copies such as 102186_trashcan. Choose an output directory for the prepared ArtiMo data, then generate it by providing one source root per downloaded dataset. The script reads asset_collection from the manifest and automatically writes to causal_data/... or not_causal_data/... under that output directory; users do not decide whether an asset is causal or non-causal:

python tools/prepare_benchmark_assets_from_manifest.py \
  --asset_manifest "$BENCH/manifests/asset_source_manifest.csv" \
  --out_data_root /path/to/prepared_artimo_data \
  --source_root PartNet-Mobility=/path/to/partnet-mobility-v0.zip \
  --source_root ArtVIP=/path/to/artvip \
  --source_root Lightwheel=/path/to/Lightwheel_OpenSource.zip \
  --copy_mode symlink

export DATA=/path/to/prepared_artimo_data

Raw ArtVIP/Lightwheel USD conversion uses the current Python environment. Make sure the ArtiMo repository requirements are installed first; the conversion path needs usd-core/pxr, and 2D evaluation needs opencv-python. If you already converted those datasets into folders containing mobility.urdf, pass those converted roots instead.

After preparation, $DATA has this layout:

$DATA/
  causal_data/
    bin1/
      mobility.urdf
  not_causal_data/
    <asset_name>/
      mobility.urdf

Each prepared asset path is:

$DATA/<asset_collection>/<asset_name>/mobility.urdf

For example, the original PartNet asset 102186 becomes $DATA/causal_data/bin1/. If your local PartNet copy was renamed to include the category, 102186_trashcan is also accepted.

Before running ArtiMo's run_agent on one of these prepared assets, build its textured animated GLB from the URDF and meshes:

bash scripts/textured.sh --root "$DATA/causal_data" bin1

To build textured meshes for every prepared benchmark asset, run the batch command on both asset collections:

JOBS=6 bash scripts/textured.sh --root "$DATA/causal_data" --all
JOBS=6 bash scripts/textured.sh --root "$DATA/not_causal_data" --all

This writes animated_textured_<asset_name>.glb and animated_textured_<asset_name>.report.json into each asset folder. For pure benchmark evaluation of existing prediction GLBs/trajectories, this file is not normally required unless the evaluator needs to fall back to asset rendering.

Contents

• annotations/: benchmark annotations. Each JSON has compact metadata with action_prompt, source dataset, and source file/asset.
• gt_animations/: GT plan.json, trajectory.jsonl, trajectory.npz, and plan_animated.glb.
• gt_phase_static_meshes/: per-phase endpoint static mesh GLBs.
• manifests/eval_manifest_225.json: prebuilt case manifest for evaluation.
• manifests/asset_source_manifest.csv: source-dataset asset id/path to benchmark asset_name mapping.
• manifests/phase_static_manifest.csv: GT phase endpoint meshes for 3D eval.

Evaluate A 3D Prediction

Put each prediction under one root, using either plan_animated.glb, trajectory.jsonl, or both:

<pred_root>/<class>/<asset>/<action>/plan_animated.glb
<pred_root>/<class>/<asset>/<action>/trajectory.jsonl
<pred_root>/<asset>/<action>/<asset>/plan_animated.glb
<pred_root>/<asset>/<action>/<asset>/trajectory.jsonl

Run 3D evaluation:

scripts/run_3d_eval.sh \
  --cases_manifest "$BENCH/manifests/eval_manifest_225.json" \
  --gt_phase_static_manifest "$BENCH/manifests/phase_static_manifest.csv" \
  --data_root "$DATA" \
  --prediction_root /path/to/pred_root \
  --prediction_variant full_agent \
  --out_dir outputs/eval_3d_full_agent \
  --variants full_agent \
  --sequence_source auto

PyTorch3D is optional but useful both for ArtiMo's GPU visual rasterization path and for final GPU-accelerated 3D evaluation. Install it from https://github.com/facebookresearch/pytorch3d/blob/main/INSTALL.md. For eval, use --pc_backend pytorch3d --gpu_devices 0; if PyTorch3D is not installed, use --pc_backend torch --gpu_devices 0 for the Torch CUDA backend.

Evaluate A 2D Prediction

For mesh/GLB methods, run the 3D eval first. Then:

scripts/run_2d_eval.sh \
  --manifest "$BENCH/manifests/eval_manifest_225.json" \
  --own_3d_dir outputs/eval_3d_full_agent \
  --out_dir outputs/eval_2d_full_agent \
  --data_root "$DATA" \
  --variants full_agent \
  --project_root "$BENCH"

For a GIF baseline, put GIFs under one of:

<puppet_root>/<asset>/<action>/reference_view_*.gif
<puppet_root>/causal/<asset>/<action>/reference_view_*.gif

Then add:

--variants puppet_master --puppet_root /path/to/puppet_root

Reproduce Batch Results

Generate predictions for all 225 cases, then run:

scripts/run_3d_eval.sh \
  --cases_manifest "$BENCH/manifests/eval_manifest_225.json" \
  --gt_phase_static_manifest "$BENCH/manifests/phase_static_manifest.csv" \
  --data_root "$DATA" \
  --prediction_root /path/to/pred_root \
  --prediction_variant full_agent \
  --out_dir outputs/eval_3d_full_agent \
  --variants full_agent \
  --sequence_source auto \
  --num_points_per_link 2048 \
  --voxel_resolution 64 \
  --pc_backend auto \
  --gpu_devices 0

scripts/run_2d_eval.sh \
  --manifest "$BENCH/manifests/eval_manifest_225.json" \
  --own_3d_dir outputs/eval_3d_full_agent \
  --out_dir outputs/eval_2d_full_agent \
  --data_root "$DATA" \
  --variants full_agent \
  --project_root "$BENCH" \
  --workers 4

More details are in docs/BENCHMARK_AND_EVAL.md in the ArtiMo repository.