Datasets:

Neptune-T
/

DragMesh-3-Scene

#!/usr/bin/env python3
"""Convert PartNet-Mobility dataset to simulation-ready SDF format.
This script processes articulated URDF assets from PartNet-Mobility into
Drake-compatible SDF files with:
- Physics properties (mass, inertia from VLM-estimated masses)
- Canonicalized orientation (front-facing +Y, z-up)
- Corrected scale (VLM-validated dimensions)
- Placement type metadata (floor/wall/ceiling)
Usage:
    # Process single asset or small batch
    python scripts/convert_partnet_mobility.py \\
        --input /path/to/partnet-mobility-v0 \\
        --output /path/to/processed_output
"""

import argparse
import json
import logging
import shutil
import traceback

from dataclasses import asdict, dataclass
from pathlib import Path

import numpy as np

from omegaconf import OmegaConf
from pydrake.multibody.parsing import Parser
from pydrake.multibody.plant import MultibodyPlant
from scipy.spatial.transform import Rotation
from tqdm import tqdm

from scenesmith.agent_utils.articulated_physics_analyzer import (
    PlacementOptions,
    analyze_articulated_physics,
)
from scenesmith.agent_utils.convex_decomposition_server import (
    ConvexDecompositionClient,
    ConvexDecompositionServer,
)
from scenesmith.agent_utils.materials import get_friction
from scenesmith.agent_utils.mesh_utils import convert_objs_to_gltf
from scenesmith.agent_utils.urdf_to_sdf import (
    compute_articulated_bounding_box,
    compute_link_physics_from_meshes,
    compute_sdf_bounding_box,
    convert_urdf_to_sdf,
    extract_link_meshes,
    parse_urdf,
    update_sdf_model_pose,
    validate_urdf_meshes,
)
from scenesmith.agent_utils.vlm_service import VLMService

console_logger = logging.getLogger(__name__)


def get_placement_type(opts: PlacementOptions) -> str:
    """Get canonical placement type from placement options.
    Floor takes precedence if model can be placed on floor or wall/ceiling.
    """
    if opts.on_floor:
        return "floor"
    elif opts.on_wall:
        return "wall"
    elif opts.on_ceiling:
        return "ceiling"
    return "floor"  # Default to floor.


@dataclass
class AnalysisResult:
    """Result of VLM analysis for an articulated asset."""

    # Canonicalization.
    front_axis: str  # "+X", "-X", "+Y", "-Y" (Z is always up).

    # Placement.
    placement_options: PlacementOptions

    # Scale.
    scale_correct: bool
    corrected_dimensions_m: tuple[float, float, float] | None  # (W, D, H)
    scale_factor: float  # Applied scale factor (1.0 if unchanged).

    # Physics per link.
    link_materials: dict[str, str]  # link_name -> material
    link_masses: dict[str, float]  # link_name -> mass_kg

    # Metadata.
    category: str
    model_id: str

    # Total mass (sum of link masses).
    total_mass_kg: float = 0.0

    # Link descriptions (from VLM analysis).
    link_descriptions: dict[str, str] | None = None

    # VLM images directory (if saved).
    vlm_images_dir: Path | None = None

    # Front view image index (for debugging).
    front_view_image_index: int | None = None

    # Whether this object is a manipuland (can be picked up/manipulated).
    is_manipuland: bool = False

    # Overall description of the object (from VLM analysis).
    object_description: str | None = None


def load_asset_metadata(asset_dir: Path) -> dict:
    """Load asset metadata from meta.json.
    Args:
        asset_dir: Path to asset directory.
    Returns:
        Metadata dict with at least 'model_cat' and 'model_id' keys.
    """
    meta_path = asset_dir / "meta.json"
    if not meta_path.exists():
        return {"model_cat": "Unknown", "model_id": asset_dir.name}

    with open(meta_path) as f:
        return json.load(f)


def compute_front_rotation_matrix(front_axis: str) -> np.ndarray:
    """Compute rotation matrix to rotate front axis to +Y.
    Args:
        front_axis: Current front axis ("+X", "-X", "+Y", "-Y").
    Returns:
        3x3 rotation matrix.
    """
    # Rotation around Z axis to align front to +Y.
    rotations = {
        "+Y": 0.0,  # Already facing +Y.
        "-Y": np.pi,  # 180 degrees.
        "+X": -np.pi / 2,  # 90 degrees CCW.
        "-X": np.pi / 2,  # 90 degrees CW.
    }

    angle = rotations.get(front_axis, 0.0)
    return Rotation.from_euler("z", angle).as_matrix()


def compute_canonical_pose(
    placement_type: str,
    front_axis: str,
    min_xyz: np.ndarray,
    max_xyz: np.ndarray,
    center: np.ndarray,
) -> tuple[float, float, float, float, float, float]:
    """Compute canonical pose (x, y, z, roll, pitch, yaw) based on placement type.
    Different placement types require different canonicalization:
    - Floor/on_object: Bottom at Z=0, center XY, front faces +Y
    - Wall: Back at Y=0 (against wall), center X, center Z
    - Ceiling: Top at Z=0, center XY (no flip - objects already modeled correctly)
    Args:
        placement_type: One of "floor", "wall", "ceiling", or default "floor".
        front_axis: Front axis for rotation ("+X", "-X", "+Y", "-Y").
        min_xyz: Minimum bounding box coordinates.
        max_xyz: Maximum bounding box coordinates.
        center: Center of bounding box.
    Returns:
        Pose tuple (tx, ty, tz, roll, pitch, yaw).
    """
    # Rotation around Z to align front to +Y.
    front_rotation = compute_front_rotation_matrix(front_axis)
    yaw = Rotation.from_matrix(front_rotation).as_euler("xyz")[2]

    # SDF pose semantics: p_world = R @ p_model + t
    # Since rotation is applied first, we must compute translation based on the
    # ROTATED bounding box, not the original. Otherwise centering will be wrong.
    # Build all 8 corners of the bounding box.
    corners = np.array(
        [
            [min_xyz[0], min_xyz[1], min_xyz[2]],
            [min_xyz[0], min_xyz[1], max_xyz[2]],
            [min_xyz[0], max_xyz[1], min_xyz[2]],
            [min_xyz[0], max_xyz[1], max_xyz[2]],
            [max_xyz[0], min_xyz[1], min_xyz[2]],
            [max_xyz[0], min_xyz[1], max_xyz[2]],
            [max_xyz[0], max_xyz[1], min_xyz[2]],
            [max_xyz[0], max_xyz[1], max_xyz[2]],
        ]
    )
    # Rotate corners by front_rotation (yaw around Z).
    rotated_corners = corners @ front_rotation.T
    # Compute rotated bounding box.
    rot_min = rotated_corners.min(axis=0)
    rot_max = rotated_corners.max(axis=0)
    rot_center = (rot_min + rot_max) / 2

    if placement_type == "wall":
        # Wall objects: back at Y=0 (against wall), center X, center Z.
        # After rotation, object faces +Y, so back is at rotated min Y.
        tx = -rot_center[0]
        ty = -rot_min[1]  # Put back at Y=0.
        tz = -rot_center[2]  # Center vertically.
    elif placement_type == "ceiling":
        # Ceiling objects: top at Z=0, center XY.
        # No flip needed - ceiling objects (lights, fans) are modeled with
        # their active side facing down already.
        tx = -rot_center[0]
        ty = -rot_center[1]
        tz = -rot_max[2]  # Put top at Z=0.
    else:
        # Floor and on_object: bottom at Z=0, center XY.
        tx = -rot_center[0]
        ty = -rot_center[1]
        tz = -rot_min[2]  # Put bottom at Z=0.

    # No roll/pitch needed - only yaw for front axis alignment.
    roll, pitch = 0.0, 0.0

    return tx, ty, tz, roll, pitch, yaw


def analyze_asset_with_vlm(
    asset_dir: Path,
    metadata: dict,
    vlm_images_output_dir: Path | None = None,
) -> AnalysisResult:
    """Analyze asset using VLM for physics and canonicalization.
    Args:
        asset_dir: Path to asset directory.
        metadata: Asset metadata.
        vlm_images_output_dir: Directory to save VLM analysis images (if provided).
    Returns:
        AnalysisResult with canonicalization and physics info.
    """
    category = metadata.get("model_cat", "Unknown")
    model_id = metadata.get("model_id", asset_dir.name)

    # Parse URDF to get link names.
    urdf_path = asset_dir / "mobility.urdf"
    if not urdf_path.exists():
        raise FileNotFoundError(f"URDF not found: {urdf_path}")

    # Load VLM configuration.
    config_path = (
        Path(__file__).parent.parent
        / "configurations/furniture_agent/base_furniture_agent.yaml"
    )
    cfg = OmegaConf.load(config_path)

    # Create VLM service.
    vlm_service = VLMService()

    # Compute bounding box for VLM analysis.
    bbox_min, bbox_max, _ = compute_articulated_bounding_box(urdf_path)

    # Get link names with visual geometry (excludes empty links like "base").
    link_meshes = extract_link_meshes(urdf_path)
    link_names_with_geometry = [lm.link_name for lm in link_meshes]

    # Analyze with VLM using per-link rendering.
    vlm_analysis = analyze_articulated_physics(
        urdf_path=urdf_path,
        link_names=link_names_with_geometry,
        bounding_box={"min": bbox_min.tolist(), "max": bbox_max.tolist()},
        vlm_service=vlm_service,
        cfg=cfg,
        category=category,
        debug_output_dir=vlm_images_output_dir,
    )

    # Build result from VLM analysis.
    placement = PlacementOptions(
        on_floor=vlm_analysis.placement_options.on_floor,
        on_wall=vlm_analysis.placement_options.on_wall,
        on_ceiling=vlm_analysis.placement_options.on_ceiling,
        on_object=vlm_analysis.placement_options.on_object,
    )

    return AnalysisResult(
        front_axis=vlm_analysis.front_axis,
        placement_options=placement,
        scale_correct=vlm_analysis.scale_correct,
        corrected_dimensions_m=None,
        scale_factor=vlm_analysis.scale_factor,
        link_materials=vlm_analysis.link_materials,
        link_masses=vlm_analysis.link_masses,
        category=category,
        model_id=model_id,
        total_mass_kg=vlm_analysis.total_mass_kg,
        link_descriptions=vlm_analysis.link_descriptions,
        vlm_images_dir=vlm_analysis.vlm_images_dir,
        front_view_image_index=vlm_analysis.front_view_image_index,
        object_description=vlm_analysis.object_description,
    )


def validate_with_drake(model_path: Path) -> bool:
    """Validate URDF or SDF file can be parsed by Drake.
    Args:
        model_path: Path to URDF or SDF file.
    Returns:
        True if validation passes.
    """
    try:
        plant = MultibodyPlant(time_step=0.0)
        parser = Parser(plant)
        parser.AddModels(str(model_path))
        plant.Finalize()
        return True
    except Exception as e:
        console_logger.error(f"Drake validation failed for {model_path}: {e}")
        return False


def process_single_asset(
    asset_dir: Path, output_dir: Path, collision_client: ConvexDecompositionClient
) -> Path | None:
    """Process a single PartNet-Mobility asset.
    Args:
        asset_dir: Path to asset directory.
        output_dir: Path to output directory.
        collision_client: Convex decomposition client for collision geometry generation.
    Returns:
        Path to generated SDF, or None if processing failed.
    """
    try:
        # Load metadata.
        metadata = load_asset_metadata(asset_dir)
        model_id = asset_dir.name
        category = metadata.get("model_cat", "Unknown")

        console_logger.info(f"Processing {model_id} ({category})")

        # Check URDF exists.
        urdf_path = asset_dir / "mobility.urdf"
        if not urdf_path.exists():
            console_logger.warning(f"No URDF found in {asset_dir}")
            return None

        # Validate URDF has valid meshes.
        urdf_result = parse_urdf(urdf_path)
        valid_meshes, _ = validate_urdf_meshes(
            urdf_path=urdf_path, urdf_result=urdf_result
        )
        if not valid_meshes:
            console_logger.warning(f"No valid meshes in {model_id}")
            return None

        # Create output directory for this asset.
        asset_output_dir = output_dir / model_id
        asset_output_dir.mkdir(parents=True, exist_ok=True)

        # Create VLM images directory.
        vlm_images_dir = asset_output_dir / "vlm_images"
        vlm_images_dir.mkdir(parents=True, exist_ok=True)

        # Analyze asset with VLM.
        analysis = analyze_asset_with_vlm(
            asset_dir=asset_dir,
            metadata=metadata,
            vlm_images_output_dir=vlm_images_dir,
        )

        # Compute link physics from meshes.
        link_physics = compute_link_physics_from_meshes(
            urdf_path=urdf_path, link_masses=analysis.link_masses
        )

        # Compute link friction from materials.
        link_friction = {
            link_name: get_friction(material)
            for link_name, material in analysis.link_materials.items()
        }

        # Two-pass conversion to handle GLTF coordinate frame changes:
        # Pass 1: Convert URDF to SDF WITHOUT model pose (exports GLTFs).
        sdf_path = convert_urdf_to_sdf(
            urdf_path=urdf_path,
            output_path=asset_output_dir / "mobility.sdf",
            link_physics=link_physics,
            link_friction=link_friction,
            model_name=f"partnet_{model_id}",
            repair_missing_meshes=True,
            model_pose=None,  # No pose yet - will add after computing from GLTFs.
            generate_collision=True,
            collision_client=collision_client,
            collision_threshold=0.05,
            merge_visuals=True,
            scale_factor=analysis.scale_factor,
        )

        # Pass 2: Compute bounding box from output GLTFs (accounts for Y-up export).
        # This gives us the correct coordinate frame for the rendered geometry.
        min_xyz, max_xyz, center = compute_sdf_bounding_box(
            sdf_path=sdf_path,
            scale_factor=analysis.scale_factor,
        )

        # Compute model-level pose for canonicalization.
        placement_type = get_placement_type(analysis.placement_options)
        model_pose = compute_canonical_pose(
            placement_type=placement_type,
            front_axis=analysis.front_axis,
            min_xyz=min_xyz,
            max_xyz=max_xyz,
            center=center,
        )

        # Update SDF with the correct model pose.
        # Note: model_pose is already in scaled coordinates (from GLTF bounding box).
        update_sdf_model_pose(sdf_path, model_pose)

        # Validate with Drake.
        if not validate_with_drake(sdf_path):
            console_logger.error(f"Drake validation failed for {model_id}")
            return None

        # Save analysis results.
        # Bounding box and model_pose are already in scaled coordinates:
        # - compute_sdf_bounding_box scales vertices and uses SDF link poses (already scaled)
        # - compute_canonical_pose derives model_pose from scaled bounding box
        analysis_path = asset_output_dir / "analysis.json"
        analysis_dict = {
            "front_axis": analysis.front_axis,
            "placement_options": asdict(analysis.placement_options),
            "scale_correct": analysis.scale_correct,
            "scale_factor": analysis.scale_factor,
            "link_materials": analysis.link_materials,
            "link_masses": analysis.link_masses,
            "total_mass_kg": analysis.total_mass_kg,
            "is_manipuland": analysis.is_manipuland,
            "category": analysis.category,
            "model_id": analysis.model_id,
            "bounding_box": {
                "min": min_xyz.tolist(),
                "max": max_xyz.tolist(),
                "center": center.tolist(),
            },
            "placement_type": placement_type,
            "model_pose": list(model_pose),
            "format": "sdf",
        }
        # Add optional fields if present.
        if analysis.link_descriptions:
            analysis_dict["link_descriptions"] = analysis.link_descriptions
        if analysis.front_view_image_index is not None:
            analysis_dict["front_view_image_index"] = analysis.front_view_image_index
        if analysis.object_description:
            analysis_dict["object_description"] = analysis.object_description
        with open(analysis_path, "w") as f:
            json.dump(analysis_dict, f, indent=2)

        # Copy original files.
        for orig_file in ["meta.json", "bounding_box.json"]:
            src = asset_dir / orig_file
            dst = asset_output_dir / orig_file
            if src.exists() and not dst.exists():
                dst.write_text(src.read_text())

        # Check if merge succeeded by looking for merged GLTFs.
        # If merge failed, fall back to copying OBJs and converting them.
        visual_dst = asset_output_dir / "visual"
        merged_gltfs = list(visual_dst.glob("*_visual.gltf"))
        if not merged_gltfs:
            console_logger.info("Merge failed, falling back to individual GLTFs")
            # Copy textured_objs directory from source.
            textured_objs_src = asset_dir / "textured_objs"
            if textured_objs_src.exists():
                if visual_dst.exists():
                    shutil.rmtree(visual_dst)
                shutil.copytree(src=textured_objs_src, dst=visual_dst)
            # Convert OBJ files to GLTF.
            if visual_dst.exists():
                convert_objs_to_gltf(visual_dst)

        console_logger.info(f"Successfully processed {model_id}")
        return sdf_path

    except Exception as e:
        console_logger.error(f"Failed to process {asset_dir.name}: {e}")
        console_logger.debug(traceback.format_exc())
        return None


def find_partnet_assets(input_path: Path) -> list[Path]:
    """Find PartNet-Mobility asset directories.
    Args:
        input_path: Path to PartNet-Mobility dataset root.
    Returns:
        List of asset directory paths.
    """
    assets = []
    for child in sorted(input_path.iterdir()):
        if child.is_dir() and (child / "mobility.urdf").exists():
            assets.append(child)
    return assets


def main():
    parser = argparse.ArgumentParser(
        description="Convert PartNet-Mobility dataset to simulation-ready SDF."
    )
    parser.add_argument(
        "--input",
        "-i",
        type=Path,
        required=True,
        help="Path to PartNet-Mobility dataset root.",
    )
    parser.add_argument(
        "--output",
        "-o",
        type=Path,
        required=True,
        help="Path to output directory.",
    )
    parser.add_argument(
        "--limit",
        "-l",
        type=int,
        default=None,
        help="Limit number of assets to process (for testing).",
    )
    parser.add_argument(
        "--skip-existing",
        action="store_true",
        help="Skip assets that already have output.",
    )
    parser.add_argument(
        "--category",
        type=str,
        default=None,
        help="Only process assets of this category.",
    )
    parser.add_argument(
        "--debug",
        action="store_true",
        help="Enable debug logging.",
    )
    parser.add_argument(
        "--ids",
        type=str,
        default=None,
        help="Comma-separated list of asset IDs to process.",
    )
    args = parser.parse_args()

    # Configure logging.
    logging.basicConfig(
        level=logging.DEBUG if args.debug else logging.INFO, format="%(message)s"
    )

    # Find assets.
    console_logger.info(f"Scanning {args.input} for PartNet-Mobility assets...")
    assets = find_partnet_assets(args.input)
    console_logger.info(f"Found {len(assets)} assets")

    # Filter by specific IDs if specified.
    if args.ids:
        target_ids = set(args.ids.split(","))
        filtered = [a for a in assets if a.name in target_ids]
        assets = filtered
        console_logger.info(f"Filtered to {len(assets)} assets by ID: {args.ids}")

    # Filter by category if specified.
    if args.category:
        filtered = []
        for asset_dir in assets:
            metadata = load_asset_metadata(asset_dir)
            if metadata.get("model_cat") == args.category:
                filtered.append(asset_dir)
        assets = filtered
        console_logger.info(f"Filtered to {len(assets)} {args.category} assets")

    # Skip existing if requested.
    if args.skip_existing:
        filtered = []
        for asset_dir in assets:
            output_sdf = args.output / asset_dir.name / "mobility.sdf"
            if not output_sdf.exists():
                filtered.append(asset_dir)
        skipped = len(assets) - len(filtered)
        assets = filtered
        console_logger.info(
            f"Skipping {skipped} existing assets, {len(assets)} remaining"
        )

    # Apply limit.
    if args.limit:
        assets = assets[: args.limit]
        console_logger.info(f"Limited to {len(assets)} assets")

    if not assets:
        console_logger.info("No assets to process")
        return

    # Create output directory.
    args.output.mkdir(parents=True, exist_ok=True)

    # Start convex decomposition server for collision geometry generation.
    console_logger.info("Starting convex decomposition server")
    collision_server = ConvexDecompositionServer(port_range=(7100, 7150))
    collision_server.start()
    collision_server.wait_until_ready()
    collision_client = collision_server.get_client()

    # Process assets sequentially.
    # For parallel processing, use scripts/convert_partnet_parallel.sh instead.
    success_count = 0
    fail_count = 0
    try:
        for asset_dir in tqdm(assets, desc="Processing"):
            result = process_single_asset(
                asset_dir=asset_dir,
                output_dir=args.output,
                collision_client=collision_client,
            )
            if result:
                success_count += 1
            else:
                fail_count += 1
    finally:
        # Stop convex decomposition server.
        console_logger.info("Stopping convex decomposition server")
        collision_server.stop()

    console_logger.info(f"Done! Success: {success_count}, Failed: {fail_count}")


if __name__ == "__main__":
    main()

	#!/usr/bin/env python3
	"""Convert PartNet-Mobility dataset to simulation-ready SDF format.

	This script processes articulated URDF assets from PartNet-Mobility into
	Drake-compatible SDF files with:
	- Physics properties (mass, inertia from VLM-estimated masses)
	- Canonicalized orientation (front-facing +Y, z-up)
	- Corrected scale (VLM-validated dimensions)
	- Placement type metadata (floor/wall/ceiling)

	Usage:
	# Process single asset or small batch
	python scripts/convert_partnet_mobility.py \\
	--input /path/to/partnet-mobility-v0 \\
	--output /path/to/processed_output
	"""

	import argparse
	import json
	import logging
	import shutil
	import traceback

	from dataclasses import asdict, dataclass
	from pathlib import Path

	import numpy as np

	from omegaconf import OmegaConf
	from pydrake.multibody.parsing import Parser
	from pydrake.multibody.plant import MultibodyPlant
	from scipy.spatial.transform import Rotation
	from tqdm import tqdm

	from scenesmith.agent_utils.articulated_physics_analyzer import (
	PlacementOptions,
	analyze_articulated_physics,
	)
	from scenesmith.agent_utils.convex_decomposition_server import (
	ConvexDecompositionClient,
	ConvexDecompositionServer,
	)
	from scenesmith.agent_utils.materials import get_friction
	from scenesmith.agent_utils.mesh_utils import convert_objs_to_gltf
	from scenesmith.agent_utils.urdf_to_sdf import (
	compute_articulated_bounding_box,
	compute_link_physics_from_meshes,
	compute_sdf_bounding_box,
	convert_urdf_to_sdf,
	extract_link_meshes,
	parse_urdf,
	update_sdf_model_pose,
	validate_urdf_meshes,
	)
	from scenesmith.agent_utils.vlm_service import VLMService

	console_logger = logging.getLogger(__name__)


	def get_placement_type(opts: PlacementOptions) -> str:
	"""Get canonical placement type from placement options.

	Floor takes precedence if model can be placed on floor or wall/ceiling.
	"""
	if opts.on_floor:
	return "floor"
	elif opts.on_wall:
	return "wall"
	elif opts.on_ceiling:
	return "ceiling"
	return "floor" # Default to floor.


	@dataclass
	class AnalysisResult:
	"""Result of VLM analysis for an articulated asset."""

	# Canonicalization.
	front_axis: str # "+X", "-X", "+Y", "-Y" (Z is always up).

	# Placement.
	placement_options: PlacementOptions

	# Scale.
	scale_correct: bool
	corrected_dimensions_m: tuple[float, float, float] \| None # (W, D, H)
	scale_factor: float # Applied scale factor (1.0 if unchanged).

	# Physics per link.
	link_materials: dict[str, str] # link_name -> material
	link_masses: dict[str, float] # link_name -> mass_kg

	# Metadata.
	category: str
	model_id: str

	# Total mass (sum of link masses).
	total_mass_kg: float = 0.0

	# Link descriptions (from VLM analysis).
	link_descriptions: dict[str, str] \| None = None

	# VLM images directory (if saved).
	vlm_images_dir: Path \| None = None

	# Front view image index (for debugging).
	front_view_image_index: int \| None = None

	# Whether this object is a manipuland (can be picked up/manipulated).
	is_manipuland: bool = False

	# Overall description of the object (from VLM analysis).
	object_description: str \| None = None


	def load_asset_metadata(asset_dir: Path) -> dict:
	"""Load asset metadata from meta.json.

	Args:
	asset_dir: Path to asset directory.

	Returns:
	Metadata dict with at least 'model_cat' and 'model_id' keys.
	"""
	meta_path = asset_dir / "meta.json"
	if not meta_path.exists():
	return {"model_cat": "Unknown", "model_id": asset_dir.name}

	with open(meta_path) as f:
	return json.load(f)


	def compute_front_rotation_matrix(front_axis: str) -> np.ndarray:
	"""Compute rotation matrix to rotate front axis to +Y.

	Args:
	front_axis: Current front axis ("+X", "-X", "+Y", "-Y").

	Returns:
	3x3 rotation matrix.
	"""
	# Rotation around Z axis to align front to +Y.
	rotations = {
	"+Y": 0.0, # Already facing +Y.
	"-Y": np.pi, # 180 degrees.
	"+X": -np.pi / 2, # 90 degrees CCW.
	"-X": np.pi / 2, # 90 degrees CW.
	}

	angle = rotations.get(front_axis, 0.0)
	return Rotation.from_euler("z", angle).as_matrix()


	def compute_canonical_pose(
	placement_type: str,
	front_axis: str,
	min_xyz: np.ndarray,
	max_xyz: np.ndarray,
	center: np.ndarray,
	) -> tuple[float, float, float, float, float, float]:
	"""Compute canonical pose (x, y, z, roll, pitch, yaw) based on placement type.

	Different placement types require different canonicalization:
	- Floor/on_object: Bottom at Z=0, center XY, front faces +Y
	- Wall: Back at Y=0 (against wall), center X, center Z
	- Ceiling: Top at Z=0, center XY (no flip - objects already modeled correctly)

	Args:
	placement_type: One of "floor", "wall", "ceiling", or default "floor".
	front_axis: Front axis for rotation ("+X", "-X", "+Y", "-Y").
	min_xyz: Minimum bounding box coordinates.
	max_xyz: Maximum bounding box coordinates.
	center: Center of bounding box.

	Returns:
	Pose tuple (tx, ty, tz, roll, pitch, yaw).
	"""
	# Rotation around Z to align front to +Y.
	front_rotation = compute_front_rotation_matrix(front_axis)
	yaw = Rotation.from_matrix(front_rotation).as_euler("xyz")[2]

	# SDF pose semantics: p_world = R @ p_model + t
	# Since rotation is applied first, we must compute translation based on the
	# ROTATED bounding box, not the original. Otherwise centering will be wrong.
	# Build all 8 corners of the bounding box.
	corners = np.array(
	[
	[min_xyz[0], min_xyz[1], min_xyz[2]],
	[min_xyz[0], min_xyz[1], max_xyz[2]],
	[min_xyz[0], max_xyz[1], min_xyz[2]],
	[min_xyz[0], max_xyz[1], max_xyz[2]],
	[max_xyz[0], min_xyz[1], min_xyz[2]],
	[max_xyz[0], min_xyz[1], max_xyz[2]],
	[max_xyz[0], max_xyz[1], min_xyz[2]],
	[max_xyz[0], max_xyz[1], max_xyz[2]],
	]
	)
	# Rotate corners by front_rotation (yaw around Z).
	rotated_corners = corners @ front_rotation.T
	# Compute rotated bounding box.
	rot_min = rotated_corners.min(axis=0)
	rot_max = rotated_corners.max(axis=0)
	rot_center = (rot_min + rot_max) / 2

	if placement_type == "wall":
	# Wall objects: back at Y=0 (against wall), center X, center Z.
	# After rotation, object faces +Y, so back is at rotated min Y.
	tx = -rot_center[0]
	ty = -rot_min[1] # Put back at Y=0.
	tz = -rot_center[2] # Center vertically.
	elif placement_type == "ceiling":
	# Ceiling objects: top at Z=0, center XY.
	# No flip needed - ceiling objects (lights, fans) are modeled with
	# their active side facing down already.
	tx = -rot_center[0]
	ty = -rot_center[1]
	tz = -rot_max[2] # Put top at Z=0.
	else:
	# Floor and on_object: bottom at Z=0, center XY.
	tx = -rot_center[0]
	ty = -rot_center[1]
	tz = -rot_min[2] # Put bottom at Z=0.

	# No roll/pitch needed - only yaw for front axis alignment.
	roll, pitch = 0.0, 0.0

	return tx, ty, tz, roll, pitch, yaw


	def analyze_asset_with_vlm(
	asset_dir: Path,
	metadata: dict,
	vlm_images_output_dir: Path \| None = None,
	) -> AnalysisResult:
	"""Analyze asset using VLM for physics and canonicalization.

	Args:
	asset_dir: Path to asset directory.
	metadata: Asset metadata.
	vlm_images_output_dir: Directory to save VLM analysis images (if provided).

	Returns:
	AnalysisResult with canonicalization and physics info.
	"""
	category = metadata.get("model_cat", "Unknown")
	model_id = metadata.get("model_id", asset_dir.name)

	# Parse URDF to get link names.
	urdf_path = asset_dir / "mobility.urdf"
	if not urdf_path.exists():
	raise FileNotFoundError(f"URDF not found: {urdf_path}")

	# Load VLM configuration.
	config_path = (
	Path(__file__).parent.parent
	/ "configurations/furniture_agent/base_furniture_agent.yaml"
	)
	cfg = OmegaConf.load(config_path)

	# Create VLM service.
	vlm_service = VLMService()

	# Compute bounding box for VLM analysis.
	bbox_min, bbox_max, _ = compute_articulated_bounding_box(urdf_path)

	# Get link names with visual geometry (excludes empty links like "base").
	link_meshes = extract_link_meshes(urdf_path)
	link_names_with_geometry = [lm.link_name for lm in link_meshes]

	# Analyze with VLM using per-link rendering.
	vlm_analysis = analyze_articulated_physics(
	urdf_path=urdf_path,
	link_names=link_names_with_geometry,
	bounding_box={"min": bbox_min.tolist(), "max": bbox_max.tolist()},
	vlm_service=vlm_service,
	cfg=cfg,
	category=category,
	debug_output_dir=vlm_images_output_dir,
	)

	# Build result from VLM analysis.
	placement = PlacementOptions(
	on_floor=vlm_analysis.placement_options.on_floor,
	on_wall=vlm_analysis.placement_options.on_wall,
	on_ceiling=vlm_analysis.placement_options.on_ceiling,
	on_object=vlm_analysis.placement_options.on_object,
	)

	return AnalysisResult(
	front_axis=vlm_analysis.front_axis,
	placement_options=placement,
	scale_correct=vlm_analysis.scale_correct,
	corrected_dimensions_m=None,
	scale_factor=vlm_analysis.scale_factor,
	link_materials=vlm_analysis.link_materials,
	link_masses=vlm_analysis.link_masses,
	category=category,
	model_id=model_id,
	total_mass_kg=vlm_analysis.total_mass_kg,
	link_descriptions=vlm_analysis.link_descriptions,
	vlm_images_dir=vlm_analysis.vlm_images_dir,
	front_view_image_index=vlm_analysis.front_view_image_index,
	object_description=vlm_analysis.object_description,
	)


	def validate_with_drake(model_path: Path) -> bool:
	"""Validate URDF or SDF file can be parsed by Drake.

	Args:
	model_path: Path to URDF or SDF file.

	Returns:
	True if validation passes.
	"""
	try:
	plant = MultibodyPlant(time_step=0.0)
	parser = Parser(plant)
	parser.AddModels(str(model_path))
	plant.Finalize()
	return True
	except Exception as e:
	console_logger.error(f"Drake validation failed for {model_path}: {e}")
	return False


	def process_single_asset(
	asset_dir: Path, output_dir: Path, collision_client: ConvexDecompositionClient
	) -> Path \| None:
	"""Process a single PartNet-Mobility asset.

	Args:
	asset_dir: Path to asset directory.
	output_dir: Path to output directory.
	collision_client: Convex decomposition client for collision geometry generation.

	Returns:
	Path to generated SDF, or None if processing failed.
	"""
	try:
	# Load metadata.
	metadata = load_asset_metadata(asset_dir)
	model_id = asset_dir.name
	category = metadata.get("model_cat", "Unknown")

	console_logger.info(f"Processing {model_id} ({category})")

	# Check URDF exists.
	urdf_path = asset_dir / "mobility.urdf"
	if not urdf_path.exists():
	console_logger.warning(f"No URDF found in {asset_dir}")
	return None

	# Validate URDF has valid meshes.
	urdf_result = parse_urdf(urdf_path)
	valid_meshes, _ = validate_urdf_meshes(
	urdf_path=urdf_path, urdf_result=urdf_result
	)
	if not valid_meshes:
	console_logger.warning(f"No valid meshes in {model_id}")
	return None

	# Create output directory for this asset.
	asset_output_dir = output_dir / model_id
	asset_output_dir.mkdir(parents=True, exist_ok=True)

	# Create VLM images directory.
	vlm_images_dir = asset_output_dir / "vlm_images"
	vlm_images_dir.mkdir(parents=True, exist_ok=True)

	# Analyze asset with VLM.
	analysis = analyze_asset_with_vlm(
	asset_dir=asset_dir,
	metadata=metadata,
	vlm_images_output_dir=vlm_images_dir,
	)

	# Compute link physics from meshes.
	link_physics = compute_link_physics_from_meshes(
	urdf_path=urdf_path, link_masses=analysis.link_masses
	)

	# Compute link friction from materials.
	link_friction = {
	link_name: get_friction(material)
	for link_name, material in analysis.link_materials.items()
	}

	# Two-pass conversion to handle GLTF coordinate frame changes:
	# Pass 1: Convert URDF to SDF WITHOUT model pose (exports GLTFs).
	sdf_path = convert_urdf_to_sdf(
	urdf_path=urdf_path,
	output_path=asset_output_dir / "mobility.sdf",
	link_physics=link_physics,
	link_friction=link_friction,
	model_name=f"partnet_{model_id}",
	repair_missing_meshes=True,
	model_pose=None, # No pose yet - will add after computing from GLTFs.
	generate_collision=True,
	collision_client=collision_client,
	collision_threshold=0.05,
	merge_visuals=True,
	scale_factor=analysis.scale_factor,
	)

	# Pass 2: Compute bounding box from output GLTFs (accounts for Y-up export).
	# This gives us the correct coordinate frame for the rendered geometry.
	min_xyz, max_xyz, center = compute_sdf_bounding_box(
	sdf_path=sdf_path,
	scale_factor=analysis.scale_factor,
	)

	# Compute model-level pose for canonicalization.
	placement_type = get_placement_type(analysis.placement_options)
	model_pose = compute_canonical_pose(
	placement_type=placement_type,
	front_axis=analysis.front_axis,
	min_xyz=min_xyz,
	max_xyz=max_xyz,
	center=center,
	)

	# Update SDF with the correct model pose.
	# Note: model_pose is already in scaled coordinates (from GLTF bounding box).
	update_sdf_model_pose(sdf_path, model_pose)

	# Validate with Drake.
	if not validate_with_drake(sdf_path):
	console_logger.error(f"Drake validation failed for {model_id}")
	return None

	# Save analysis results.
	# Bounding box and model_pose are already in scaled coordinates:
	# - compute_sdf_bounding_box scales vertices and uses SDF link poses (already scaled)
	# - compute_canonical_pose derives model_pose from scaled bounding box
	analysis_path = asset_output_dir / "analysis.json"
	analysis_dict = {
	"front_axis": analysis.front_axis,
	"placement_options": asdict(analysis.placement_options),
	"scale_correct": analysis.scale_correct,
	"scale_factor": analysis.scale_factor,
	"link_materials": analysis.link_materials,
	"link_masses": analysis.link_masses,
	"total_mass_kg": analysis.total_mass_kg,
	"is_manipuland": analysis.is_manipuland,
	"category": analysis.category,
	"model_id": analysis.model_id,
	"bounding_box": {
	"min": min_xyz.tolist(),
	"max": max_xyz.tolist(),
	"center": center.tolist(),
	},
	"placement_type": placement_type,
	"model_pose": list(model_pose),
	"format": "sdf",
	}
	# Add optional fields if present.
	if analysis.link_descriptions:
	analysis_dict["link_descriptions"] = analysis.link_descriptions
	if analysis.front_view_image_index is not None:
	analysis_dict["front_view_image_index"] = analysis.front_view_image_index
	if analysis.object_description:
	analysis_dict["object_description"] = analysis.object_description
	with open(analysis_path, "w") as f:
	json.dump(analysis_dict, f, indent=2)

	# Copy original files.
	for orig_file in ["meta.json", "bounding_box.json"]:
	src = asset_dir / orig_file
	dst = asset_output_dir / orig_file
	if src.exists() and not dst.exists():
	dst.write_text(src.read_text())

	# Check if merge succeeded by looking for merged GLTFs.
	# If merge failed, fall back to copying OBJs and converting them.
	visual_dst = asset_output_dir / "visual"
	merged_gltfs = list(visual_dst.glob("*_visual.gltf"))
	if not merged_gltfs:
	console_logger.info("Merge failed, falling back to individual GLTFs")
	# Copy textured_objs directory from source.
	textured_objs_src = asset_dir / "textured_objs"
	if textured_objs_src.exists():
	if visual_dst.exists():
	shutil.rmtree(visual_dst)
	shutil.copytree(src=textured_objs_src, dst=visual_dst)
	# Convert OBJ files to GLTF.
	if visual_dst.exists():
	convert_objs_to_gltf(visual_dst)

	console_logger.info(f"Successfully processed {model_id}")
	return sdf_path

	except Exception as e:
	console_logger.error(f"Failed to process {asset_dir.name}: {e}")
	console_logger.debug(traceback.format_exc())
	return None


	def find_partnet_assets(input_path: Path) -> list[Path]:
	"""Find PartNet-Mobility asset directories.

	Args:
	input_path: Path to PartNet-Mobility dataset root.

	Returns:
	List of asset directory paths.
	"""
	assets = []
	for child in sorted(input_path.iterdir()):
	if child.is_dir() and (child / "mobility.urdf").exists():
	assets.append(child)
	return assets


	def main():
	parser = argparse.ArgumentParser(
	description="Convert PartNet-Mobility dataset to simulation-ready SDF."
	)
	parser.add_argument(
	"--input",
	"-i",
	type=Path,
	required=True,
	help="Path to PartNet-Mobility dataset root.",
	)
	parser.add_argument(
	"--output",
	"-o",
	type=Path,
	required=True,
	help="Path to output directory.",
	)
	parser.add_argument(
	"--limit",
	"-l",
	type=int,
	default=None,
	help="Limit number of assets to process (for testing).",
	)
	parser.add_argument(
	"--skip-existing",
	action="store_true",
	help="Skip assets that already have output.",
	)
	parser.add_argument(
	"--category",
	type=str,
	default=None,
	help="Only process assets of this category.",
	)
	parser.add_argument(
	"--debug",
	action="store_true",
	help="Enable debug logging.",
	)
	parser.add_argument(
	"--ids",
	type=str,
	default=None,
	help="Comma-separated list of asset IDs to process.",
	)
	args = parser.parse_args()

	# Configure logging.
	logging.basicConfig(
	level=logging.DEBUG if args.debug else logging.INFO, format="%(message)s"
	)

	# Find assets.
	console_logger.info(f"Scanning {args.input} for PartNet-Mobility assets...")
	assets = find_partnet_assets(args.input)
	console_logger.info(f"Found {len(assets)} assets")

	# Filter by specific IDs if specified.
	if args.ids:
	target_ids = set(args.ids.split(","))
	filtered = [a for a in assets if a.name in target_ids]
	assets = filtered
	console_logger.info(f"Filtered to {len(assets)} assets by ID: {args.ids}")

	# Filter by category if specified.
	if args.category:
	filtered = []
	for asset_dir in assets:
	metadata = load_asset_metadata(asset_dir)
	if metadata.get("model_cat") == args.category:
	filtered.append(asset_dir)
	assets = filtered
	console_logger.info(f"Filtered to {len(assets)} {args.category} assets")

	# Skip existing if requested.
	if args.skip_existing:
	filtered = []
	for asset_dir in assets:
	output_sdf = args.output / asset_dir.name / "mobility.sdf"
	if not output_sdf.exists():
	filtered.append(asset_dir)
	skipped = len(assets) - len(filtered)
	assets = filtered
	console_logger.info(
	f"Skipping {skipped} existing assets, {len(assets)} remaining"
	)

	# Apply limit.
	if args.limit:
	assets = assets[: args.limit]
	console_logger.info(f"Limited to {len(assets)} assets")

	if not assets:
	console_logger.info("No assets to process")
	return

	# Create output directory.
	args.output.mkdir(parents=True, exist_ok=True)

	# Start convex decomposition server for collision geometry generation.
	console_logger.info("Starting convex decomposition server")
	collision_server = ConvexDecompositionServer(port_range=(7100, 7150))
	collision_server.start()
	collision_server.wait_until_ready()
	collision_client = collision_server.get_client()

	# Process assets sequentially.
	# For parallel processing, use scripts/convert_partnet_parallel.sh instead.
	success_count = 0
	fail_count = 0
	try:
	for asset_dir in tqdm(assets, desc="Processing"):
	result = process_single_asset(
	asset_dir=asset_dir,
	output_dir=args.output,
	collision_client=collision_client,
	)
	if result:
	success_count += 1
	else:
	fail_count += 1
	finally:
	# Stop convex decomposition server.
	console_logger.info("Stopping convex decomposition server")
	collision_server.stop()

	console_logger.info(f"Done! Success: {success_count}, Failed: {fail_count}")


	if __name__ == "__main__":
	main()