AMASSDataset/convert_motion_data.py

#!/usr/bin/env python3

import numpy as np
import torch
from pathlib import Path
import isaaclab.utils.math as math_utils

def convert_motion_data(input_path: str, output_path: str):
    """Convert motion data from AMASS format to our desired format.
    
    Args:
        input_path: Path to input npz file
        output_path: Path to save the converted data
    """
    # Load the source data
    data = np.load(input_path, allow_pickle=True)
    
    # Extract data with proper slicing
    qpos = data['qpos']  # shape: [T, n_dofs]
    qvel = data['qvel']  # shape: [T, n_dofs]
    
    # Convert quaternions to rotation matrices
    # Move to GPU if available
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    
    # Extract quaternions and reorder from [x,y,z,w] to [w,x,y,z] format
    quats_raw = torch.from_numpy(qpos[:, 3:7]).float()  # [x,y,z,w]
    quats = torch.zeros_like(quats_raw)
    quats[..., 0] = quats_raw[..., 3]  # w
    quats[..., 1:] = quats_raw[..., :3]  # xyz
    quats = quats.to(device)
    
    # Create basis vectors
    num_frames = quats.shape[0]
    basis_vectors = torch.eye(3, device=device).unsqueeze(0).repeat(num_frames, 1, 1)  # [num_frames, 3, 3]
    
    # Rotate each basis vector
    rotations = torch.zeros((num_frames, 3, 3), device=device)
    for i in range(3):
        rotations[..., i] = math_utils.quat_rotate(quats, basis_vectors[..., i])
    
    rotations = rotations.cpu().numpy()
    
    # Create the output dictionary
    output_data = {
        'dof_names': data['joint_names'],
        'body_names': data['body_names'],
        'dof_positions': qpos[:, 7:],           # Remove root state (7 DoFs)
        'dof_velocities': qvel[:, 6:],          # Remove root velocities (6 DoFs)
        'body_positions': qpos[:, :3],           # Root position
        'body_rotations': rotations,             # Root rotation as 3x3 matrix
        'body_linear_velocities': qvel[:, :3],    # Root linear velocity
        'body_angular_velocities': qvel[:, 3:6],  # Root angular velocity
        'fps': 50,                              # Fixed at 50Hz
    }
    
    # Print input data shapes for verification
    print("\nInput data shapes:")
    print(f"qpos shape: {qpos.shape}")
    print(f"qvel shape: {qvel.shape}")
    print(f"quaternions shape: {quats.shape}")
    print(f"rotations shape: {rotations.shape}")
    
    # Save the converted data
    np.savez(output_path, **output_data)
    
    print(f"\nConverted data saved to {output_path}")
    print("\nOutput data contains:")
    for key, value in output_data.items():
        if isinstance(value, np.ndarray):
            print(f"- {key}: shape {value.shape} (dtype: {value.dtype})")
        else:
            print(f"- {key}: {value}")

if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser(description="Convert AMASS motion data to our format")
    parser.add_argument("--input", type=str, required=True, help="Input npz file path")
    parser.add_argument("--output", type=str, help="Output npz file path. If not provided, will use input path with _converted suffix")
    
    args = parser.parse_args()
    
    input_path = args.input
    if args.output is None:
        # Auto-generate output path by adding _converted before .npz
        input_path_obj = Path(input_path)
        output_path = str(input_path_obj.parent / f"{input_path_obj.stem}_converted.npz")
    else:
        output_path = args.output
        
    convert_motion_data(input_path, output_path)