Upload folder using huggingface_hub

3903652 verified 4 months ago

17.2 kB

# Dataset Factory
# Simple factory to create different types of datasets

from typing import Dict, Any, List, Optional
from omegaconf import OmegaConf
import torch


def create_dataset(config: OmegaConf, val: bool = False):
    """
    Create dataset based on config.
    
    Args:
        config: Configuration object
        val: Whether to create validation dataset
        
    Returns:
        Dataset instance
    """
    dataset_type = config.dataset.get('type', 'robotwin')  # Default to robotwin
    
    if dataset_type == 'robotwin':
        from .robotwin2.robotwin_agilex_dataset import RobotWinTaskDataset
        
        # Get all parameters from config
        params = {}
        
        # Add common parameters
        if hasattr(config, 'common'):
            params.update({
                'global_downsample_rate': config.common.global_downsample_rate,
                'video_action_freq_ratio': config.common.video_action_freq_ratio,
                'num_video_frames': config.common.num_video_frames,
                'video_size': (config.common.video_height, config.common.video_width),
                'state_dim': config.common.state_dim,
                'action_dim': config.common.action_dim,
            })
        
        # Add dataset-specific parameters
        if hasattr(config.dataset, 'dataset_dir'):
            params['dataset_dir'] = config.dataset.dataset_dir
        if hasattr(config.dataset, 'data_mode'):
            params['data_mode'] = config.dataset.data_mode
        if hasattr(config.dataset, 'task_mode'):
            params['task_mode'] = config.dataset.task_mode
        if hasattr(config.dataset, 'task_name'):
            params['task_name'] = config.dataset.task_name
        if hasattr(config.dataset, 'max_episodes'):
            params['max_episodes'] = config.dataset.max_episodes
        if hasattr(config.dataset, 'image_aug'):
            params['image_aug'] = config.dataset.image_aug and not val  # No aug for validation
        if hasattr(config.dataset, 'randomized_limit_per_task'):
            params['randomized_limit_per_task'] = config.dataset.randomized_limit_per_task
        
        # Add VLM checkpoint path
        if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
            params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path
        
        # Add any additional parameters from dataset.params
        if hasattr(config.dataset, 'params'):
            additional_params = OmegaConf.to_object(config.dataset.params)
            params.update(additional_params)
        
        # Set validation flag
        params['val'] = val
        
        return RobotWinTaskDataset(**params)
    
    elif dataset_type == 'ac_one':
        from .ac_one.ac_one_dataset import ACOneDataset
        
        # Get all parameters from config
        params = {}
        
        # Add common parameters
        if hasattr(config, 'common'):
            params.update({
                'global_downsample_rate': config.common.global_downsample_rate,
                'video_action_freq_ratio': config.common.video_action_freq_ratio,
                'num_video_frames': config.common.num_video_frames,
                'video_size': (config.common.video_height, config.common.video_width),
            })
        
        # Add dataset-specific parameters
        if hasattr(config.dataset, 'dataset_dir'):
            params['dataset_dir'] = config.dataset.dataset_dir
        if hasattr(config.dataset, 'task_mode'):
            params['task_mode'] = config.dataset.task_mode
        if hasattr(config.dataset, 'task_name'):
            params['task_name'] = config.dataset.task_name
        if hasattr(config.dataset, 'max_episodes'):
            params['max_episodes'] = config.dataset.max_episodes
        if hasattr(config.dataset, 'val_episodes'):
            params['val_episodes'] = config.dataset.val_episodes
        if hasattr(config.dataset, 'image_aug'):
            params['image_aug'] = config.dataset.image_aug and not val  # No aug for validation
        
        # Add VLM checkpoint path
        if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
            params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path
        
        # Add any additional parameters from dataset.params
        if hasattr(config.dataset, 'params'):
            additional_params = OmegaConf.to_object(config.dataset.params)
            params.update(additional_params)
        
        # Set validation flag
        params['val'] = val
        
        return ACOneDataset(**params)

    elif dataset_type == 'latent_action':
        from .latent_action.latent_action_dataset import LatentActionDataset

        params = {}

        # Common parameters
        if hasattr(config, 'common'):
            params.update({
                'global_downsample_rate': config.common.global_downsample_rate,
                'num_video_frames': config.common.num_video_frames,
                'video_size': (config.common.video_height, config.common.video_width),
            })

        if hasattr(config.dataset, 'dataset_dir'):
            dataset_dir = list(config.dataset.dataset_dir)
            params['dataset_dir'] = [str(p) for p in dataset_dir]
        if hasattr(config.dataset, 'max_episodes'):
            params['max_episodes'] = config.dataset.max_episodes
        if hasattr(config.dataset, 'image_aug'):
            params['image_aug'] = config.dataset.image_aug and not val

        # Optional VLM checkpoint path
        if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
            params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path

        # Optional additional params
        if hasattr(config.dataset, 'params'):
            additional_params = OmegaConf.to_object(config.dataset.params)
            params.update(additional_params)

        params['val'] = val

        return LatentActionDataset(**params)

    elif dataset_type == 'aloha_agilex_2':
        from .aloha_agilex_2.aloha_agilex2_dataset import AlohaAgilex2Dataset
        
        # Get all parameters from config
        params = {}
        
        # Add common parameters
        if hasattr(config, 'common'):
            params.update({
                'global_downsample_rate': config.common.global_downsample_rate,
                'video_action_freq_ratio': config.common.video_action_freq_ratio,
                'num_video_frames': config.common.num_video_frames,
                'video_size': (config.common.video_height, config.common.video_width),
            })
        
        # Add dataset-specific parameters
        if hasattr(config.dataset, 'dataset_dir'):
            params['dataset_dir'] = config.dataset.dataset_dir
        if hasattr(config.dataset, 'task_mode'):
            params['task_mode'] = config.dataset.task_mode
        if hasattr(config.dataset, 'task_name'):
            params['task_name'] = config.dataset.task_name
        if hasattr(config.dataset, 'max_episodes'):
            params['max_episodes'] = config.dataset.max_episodes
        if hasattr(config.dataset, 'val_episodes'):
            params['val_episodes'] = config.dataset.val_episodes
        if hasattr(config.dataset, 'image_aug'):
            params['image_aug'] = config.dataset.image_aug and not val  # No aug for validation
        
        # Add VLM checkpoint path
        if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
            params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path
        
        # Add any additional parameters from dataset.params
        if hasattr(config.dataset, 'params'):
            additional_params = OmegaConf.to_object(config.dataset.params)
            params.update(additional_params)
        
        # Set validation flag
        params['val'] = val
        
        return AlohaAgilex2Dataset(**params)

    elif dataset_type == "manifeel_zarr":
        from .manifeel.manifeel_zarr_dataset import ManiFeelZarrDataset
        params = {}
        if hasattr(config, "common"):
            params.update({
                "global_downsample_rate": config.common.global_downsample_rate,
                "video_action_freq_ratio": config.common.video_action_freq_ratio,
                "num_video_frames": config.common.num_video_frames,
                "video_size": (config.common.video_height, config.common.video_width),
            })
        if hasattr(config.dataset, "dataset_dir"):
            params["dataset_dir"] = config.dataset.dataset_dir
        if hasattr(config.dataset, "max_episodes"):
            params["max_episodes"] = config.dataset.max_episodes
        if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
            params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path
        params["val"] = val
        return ManiFeelZarrDataset(**params)

    elif dataset_type == 'libero':
        from .libero_dataset import LiberoDataset
        params = {}
        if hasattr(config, 'common'):
            params.update({
                'global_downsample_rate': config.common.global_downsample_rate,
                'video_action_freq_ratio': config.common.video_action_freq_ratio,
                'num_video_frames': config.common.num_video_frames,
                'video_size': (config.common.video_height, config.common.video_width),
                'state_dim': config.common.state_dim,
                'action_dim': config.common.action_dim,
            })
        if hasattr(config.dataset, 'dataset_dir'):
            params['dataset_dir'] = config.dataset.dataset_dir
        elif hasattr(config.dataset, 'params') and hasattr(config.dataset.params, 'root'):
            params['dataset_dir'] = config.dataset.params.root
        if hasattr(config.dataset, 'max_episodes'):
            params['max_episodes'] = config.dataset.max_episodes
        if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
            params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path
        return LiberoDataset(**params)

    elif dataset_type == 'lerobot':
        from .lerobot.lerobot_dataset import LeRobotMotusDataset

        # Get all parameters from config
        params = {}

        # Add common parameters
        if hasattr(config, 'common'):
            params.update({
                'global_downsample_rate': config.common.global_downsample_rate,
                'video_action_freq_ratio': config.common.video_action_freq_ratio,
                'num_video_frames': config.common.num_video_frames,
                'video_size': (config.common.video_height, config.common.video_width),
            })

        # Add dataset-specific parameters
        if hasattr(config.dataset, 'dataset_dir'):
            params['dataset_dir'] = config.dataset.dataset_dir
        if hasattr(config.dataset, 'task_mode'):
            params['task_mode'] = config.dataset.task_mode
        if hasattr(config.dataset, 'task_name'):
            params['task_name'] = config.dataset.task_name
        if hasattr(config.dataset, 'max_episodes'):
            params['max_episodes'] = config.dataset.max_episodes
        if hasattr(config.dataset, 'image_aug'):
            params['image_aug'] = config.dataset.image_aug and not val

        # Add VLM checkpoint path
        if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
            params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path

        # Add any additional parameters from dataset.params
        if hasattr(config.dataset, 'params'):
            additional_params = OmegaConf.to_object(config.dataset.params)
            params.update(additional_params)

        # Set validation flag
        params['val'] = val
        
        return LeRobotMotusDataset(**params)

    # Example: Add more dataset types here
    # elif dataset_type == 'bridge':
    #     from .bridge_dataset import BridgeDataset  
    #     return BridgeDataset(**params)
    
    else:
        raise ValueError(f"Unknown dataset type: {dataset_type}. Available types: robotwin, aloha_agilex_1, ac_one, aloha_agilex_2, table30")


def _process_vlm_inputs_batch(vlm_inputs: List[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
    """Process and batch VLM inputs with padding."""
    # Extract components
    input_ids_list = [vlm_input['input_ids'] for vlm_input in vlm_inputs]
    pixel_values_list = [vlm_input.get('pixel_values') for vlm_input in vlm_inputs]
    image_grid_thw_list = [vlm_input.get('image_grid_thw') for vlm_input in vlm_inputs]
    attention_mask_list = [vlm_input.get('attention_mask') for vlm_input in vlm_inputs]
    
    # Pad input_ids to same length (simplified like model implementation)
    max_seq_len = max(ids.shape[1] for ids in input_ids_list)
    padded_input_ids = []
    padded_attention_masks = []
    
    for ids, mask in zip(input_ids_list, attention_mask_list):
        if ids.shape[1] < max_seq_len:
            padding_size = max_seq_len - ids.shape[1]
            # Pad input_ids
            padding = torch.zeros(ids.shape[0], padding_size, dtype=ids.dtype, device=ids.device)
            padded_ids = torch.cat([ids, padding], dim=1)
            # Pad attention_mask
            if mask is not None:
                mask_padding = torch.zeros(mask.shape[0], padding_size, dtype=mask.dtype, device=mask.device)
                padded_mask = torch.cat([mask, mask_padding], dim=1)
            else:
                padded_mask = None
        else:
            padded_ids = ids
            padded_mask = mask
            
        padded_input_ids.append(padded_ids)
        padded_attention_masks.append(padded_mask)
    
    # Batch everything
    return {
        'input_ids': torch.cat(padded_input_ids, dim=0),
        'pixel_values': torch.cat([pv for pv in pixel_values_list if pv is not None], dim=0) if pixel_values_list and any(pv is not None for pv in pixel_values_list) else None,
        'image_grid_thw': torch.cat([igt for igt in image_grid_thw_list if igt is not None], dim=0) if image_grid_thw_list and any(igt is not None for igt in image_grid_thw_list) else None,
        'attention_mask': torch.cat([mask for mask in padded_attention_masks if mask is not None], dim=0) if any(mask is not None for mask in padded_attention_masks) else None,
    }


def _process_language_embeddings_batch(language_embeddings: List[torch.Tensor], text_len: int = 512) -> torch.Tensor:
    """Process and batch language embeddings with padding."""
    padded_embeddings = []
    
    for emb in language_embeddings:
        if emb.shape[0] <= text_len:
            padded = torch.cat([emb, emb.new_zeros(text_len - emb.shape[0], emb.shape[1])])
        else:
            padded = emb[:text_len]
        padded_embeddings.append(padded)
    
    # Stack to [B, seq_len, dim]
    return torch.stack(padded_embeddings, dim=0)


def collate_fn(batch: List[Optional[Dict[str, Any]]]) -> Optional[Dict[str, Any]]:
    """
    Universal collate function for all datasets.
    
    Args:
        batch: List of sample dictionaries (may contain None)
        
    Returns:
        Batched dictionary or None if all samples are None
    """
    # Filter out None samples
    batch = [sample for sample in batch if sample is not None]
    
    if len(batch) == 0:
        return None
    
    # Stack tensors（支持无 initial_state 的样本）
    first_frames = torch.stack([sample['first_frame'] for sample in batch])
    video_frames = torch.stack([sample['video_frames'] for sample in batch])
    action_sequences = torch.stack([sample['action_sequence'] for sample in batch])
    has_initial_state = all(('initial_state' in sample and sample['initial_state'] is not None) for sample in batch)
    initial_states = torch.stack([sample['initial_state'] for sample in batch]) if has_initial_state else None
    
    # Process VLM inputs with padding in collate_fn
    vlm_inputs = [sample.get('vlm_inputs') for sample in batch]
    processed_vlm_inputs = None
    if vlm_inputs and all(vlm_input is not None for vlm_input in vlm_inputs):
        processed_vlm_inputs = _process_vlm_inputs_batch(vlm_inputs)
    
    # Process language embeddings with padding in collate_fn  
    language_embeddings = [sample.get('language_embedding') for sample in batch if 'language_embedding' in sample]
    processed_language_embeddings = None
    if language_embeddings and any(emb is not None for emb in language_embeddings):
        processed_language_embeddings = _process_language_embeddings_batch(language_embeddings)
    
    result = {
        'first_frame': first_frames,             # [B, C, H, W]
        'video_frames': video_frames,            # [B, F, C, H, W]
        'action_sequence': action_sequences,     # [B, F, D]
        'vlm_inputs': processed_vlm_inputs,
        'language_embedding': processed_language_embeddings,
    }

    if initial_states is not None:
        result['initial_state'] = initial_states
    
    return result# This line is no-op - we'll use sed instead

	# Dataset Factory
	# Simple factory to create different types of datasets

	from typing import Dict, Any, List, Optional
	from omegaconf import OmegaConf
	import torch


	def create_dataset(config: OmegaConf, val: bool = False):
	"""
	Create dataset based on config.

	Args:
	config: Configuration object
	val: Whether to create validation dataset

	Returns:
	Dataset instance
	"""
	dataset_type = config.dataset.get('type', 'robotwin') # Default to robotwin

	if dataset_type == 'robotwin':
	from .robotwin2.robotwin_agilex_dataset import RobotWinTaskDataset

	# Get all parameters from config
	params = {}

	# Add common parameters
	if hasattr(config, 'common'):
	params.update({
	'global_downsample_rate': config.common.global_downsample_rate,
	'video_action_freq_ratio': config.common.video_action_freq_ratio,
	'num_video_frames': config.common.num_video_frames,
	'video_size': (config.common.video_height, config.common.video_width),
	'state_dim': config.common.state_dim,
	'action_dim': config.common.action_dim,
	})

	# Add dataset-specific parameters
	if hasattr(config.dataset, 'dataset_dir'):
	params['dataset_dir'] = config.dataset.dataset_dir
	if hasattr(config.dataset, 'data_mode'):
	params['data_mode'] = config.dataset.data_mode
	if hasattr(config.dataset, 'task_mode'):
	params['task_mode'] = config.dataset.task_mode
	if hasattr(config.dataset, 'task_name'):
	params['task_name'] = config.dataset.task_name
	if hasattr(config.dataset, 'max_episodes'):
	params['max_episodes'] = config.dataset.max_episodes
	if hasattr(config.dataset, 'image_aug'):
	params['image_aug'] = config.dataset.image_aug and not val # No aug for validation
	if hasattr(config.dataset, 'randomized_limit_per_task'):
	params['randomized_limit_per_task'] = config.dataset.randomized_limit_per_task

	# Add VLM checkpoint path
	if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
	params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path

	# Add any additional parameters from dataset.params
	if hasattr(config.dataset, 'params'):
	additional_params = OmegaConf.to_object(config.dataset.params)
	params.update(additional_params)

	# Set validation flag
	params['val'] = val

	return RobotWinTaskDataset(**params)

	elif dataset_type == 'ac_one':
	from .ac_one.ac_one_dataset import ACOneDataset

	# Get all parameters from config
	params = {}

	# Add common parameters
	if hasattr(config, 'common'):
	params.update({
	'global_downsample_rate': config.common.global_downsample_rate,
	'video_action_freq_ratio': config.common.video_action_freq_ratio,
	'num_video_frames': config.common.num_video_frames,
	'video_size': (config.common.video_height, config.common.video_width),
	})

	# Add dataset-specific parameters
	if hasattr(config.dataset, 'dataset_dir'):
	params['dataset_dir'] = config.dataset.dataset_dir
	if hasattr(config.dataset, 'task_mode'):
	params['task_mode'] = config.dataset.task_mode
	if hasattr(config.dataset, 'task_name'):
	params['task_name'] = config.dataset.task_name
	if hasattr(config.dataset, 'max_episodes'):
	params['max_episodes'] = config.dataset.max_episodes
	if hasattr(config.dataset, 'val_episodes'):
	params['val_episodes'] = config.dataset.val_episodes
	if hasattr(config.dataset, 'image_aug'):
	params['image_aug'] = config.dataset.image_aug and not val # No aug for validation

	# Add VLM checkpoint path
	if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
	params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path

	# Add any additional parameters from dataset.params
	if hasattr(config.dataset, 'params'):
	additional_params = OmegaConf.to_object(config.dataset.params)
	params.update(additional_params)

	# Set validation flag
	params['val'] = val

	return ACOneDataset(**params)

	elif dataset_type == 'latent_action':
	from .latent_action.latent_action_dataset import LatentActionDataset

	params = {}

	# Common parameters
	if hasattr(config, 'common'):
	params.update({
	'global_downsample_rate': config.common.global_downsample_rate,
	'num_video_frames': config.common.num_video_frames,
	'video_size': (config.common.video_height, config.common.video_width),
	})

	if hasattr(config.dataset, 'dataset_dir'):
	dataset_dir = list(config.dataset.dataset_dir)
	params['dataset_dir'] = [str(p) for p in dataset_dir]
	if hasattr(config.dataset, 'max_episodes'):
	params['max_episodes'] = config.dataset.max_episodes
	if hasattr(config.dataset, 'image_aug'):
	params['image_aug'] = config.dataset.image_aug and not val

	# Optional VLM checkpoint path
	if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
	params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path

	# Optional additional params
	if hasattr(config.dataset, 'params'):
	additional_params = OmegaConf.to_object(config.dataset.params)
	params.update(additional_params)

	params['val'] = val

	return LatentActionDataset(**params)

	elif dataset_type == 'aloha_agilex_2':
	from .aloha_agilex_2.aloha_agilex2_dataset import AlohaAgilex2Dataset

	# Get all parameters from config
	params = {}

	# Add common parameters
	if hasattr(config, 'common'):
	params.update({
	'global_downsample_rate': config.common.global_downsample_rate,
	'video_action_freq_ratio': config.common.video_action_freq_ratio,
	'num_video_frames': config.common.num_video_frames,
	'video_size': (config.common.video_height, config.common.video_width),
	})

	# Add dataset-specific parameters
	if hasattr(config.dataset, 'dataset_dir'):
	params['dataset_dir'] = config.dataset.dataset_dir
	if hasattr(config.dataset, 'task_mode'):
	params['task_mode'] = config.dataset.task_mode
	if hasattr(config.dataset, 'task_name'):
	params['task_name'] = config.dataset.task_name
	if hasattr(config.dataset, 'max_episodes'):
	params['max_episodes'] = config.dataset.max_episodes
	if hasattr(config.dataset, 'val_episodes'):
	params['val_episodes'] = config.dataset.val_episodes
	if hasattr(config.dataset, 'image_aug'):
	params['image_aug'] = config.dataset.image_aug and not val # No aug for validation

	# Add VLM checkpoint path
	if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
	params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path

	# Add any additional parameters from dataset.params
	if hasattr(config.dataset, 'params'):
	additional_params = OmegaConf.to_object(config.dataset.params)
	params.update(additional_params)

	# Set validation flag
	params['val'] = val

	return AlohaAgilex2Dataset(**params)

	elif dataset_type == "manifeel_zarr":
	from .manifeel.manifeel_zarr_dataset import ManiFeelZarrDataset
	params = {}
	if hasattr(config, "common"):
	params.update({
	"global_downsample_rate": config.common.global_downsample_rate,
	"video_action_freq_ratio": config.common.video_action_freq_ratio,
	"num_video_frames": config.common.num_video_frames,
	"video_size": (config.common.video_height, config.common.video_width),
	})
	if hasattr(config.dataset, "dataset_dir"):
	params["dataset_dir"] = config.dataset.dataset_dir
	if hasattr(config.dataset, "max_episodes"):
	params["max_episodes"] = config.dataset.max_episodes
	if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
	params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path
	params["val"] = val
	return ManiFeelZarrDataset(**params)

	elif dataset_type == 'libero':
	from .libero_dataset import LiberoDataset
	params = {}
	if hasattr(config, 'common'):
	params.update({
	'global_downsample_rate': config.common.global_downsample_rate,
	'video_action_freq_ratio': config.common.video_action_freq_ratio,
	'num_video_frames': config.common.num_video_frames,
	'video_size': (config.common.video_height, config.common.video_width),
	'state_dim': config.common.state_dim,
	'action_dim': config.common.action_dim,
	})
	if hasattr(config.dataset, 'dataset_dir'):
	params['dataset_dir'] = config.dataset.dataset_dir
	elif hasattr(config.dataset, 'params') and hasattr(config.dataset.params, 'root'):
	params['dataset_dir'] = config.dataset.params.root
	if hasattr(config.dataset, 'max_episodes'):
	params['max_episodes'] = config.dataset.max_episodes
	if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
	params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path
	return LiberoDataset(**params)

	elif dataset_type == 'lerobot':
	from .lerobot.lerobot_dataset import LeRobotMotusDataset

	# Get all parameters from config
	params = {}

	# Add common parameters
	if hasattr(config, 'common'):
	params.update({
	'global_downsample_rate': config.common.global_downsample_rate,
	'video_action_freq_ratio': config.common.video_action_freq_ratio,
	'num_video_frames': config.common.num_video_frames,
	'video_size': (config.common.video_height, config.common.video_width),
	})

	# Add dataset-specific parameters
	if hasattr(config.dataset, 'dataset_dir'):
	params['dataset_dir'] = config.dataset.dataset_dir
	if hasattr(config.dataset, 'task_mode'):
	params['task_mode'] = config.dataset.task_mode
	if hasattr(config.dataset, 'task_name'):
	params['task_name'] = config.dataset.task_name
	if hasattr(config.dataset, 'max_episodes'):
	params['max_episodes'] = config.dataset.max_episodes
	if hasattr(config.dataset, 'image_aug'):
	params['image_aug'] = config.dataset.image_aug and not val

	# Add VLM checkpoint path
	if hasattr(config.model, 'vlm') and hasattr(config.model.vlm, 'checkpoint_path'):
	params['vlm_checkpoint_path'] = config.model.vlm.checkpoint_path

	# Add any additional parameters from dataset.params
	if hasattr(config.dataset, 'params'):
	additional_params = OmegaConf.to_object(config.dataset.params)
	params.update(additional_params)

	# Set validation flag
	params['val'] = val

	return LeRobotMotusDataset(**params)

	# Example: Add more dataset types here
	# elif dataset_type == 'bridge':
	# from .bridge_dataset import BridgeDataset
	# return BridgeDataset(**params)

	else:
	raise ValueError(f"Unknown dataset type: {dataset_type}. Available types: robotwin, aloha_agilex_1, ac_one, aloha_agilex_2, table30")


	def _process_vlm_inputs_batch(vlm_inputs: List[Dict[str, Any]]) -> Dict[str, torch.Tensor]:
	"""Process and batch VLM inputs with padding."""
	# Extract components
	input_ids_list = [vlm_input['input_ids'] for vlm_input in vlm_inputs]
	pixel_values_list = [vlm_input.get('pixel_values') for vlm_input in vlm_inputs]
	image_grid_thw_list = [vlm_input.get('image_grid_thw') for vlm_input in vlm_inputs]
	attention_mask_list = [vlm_input.get('attention_mask') for vlm_input in vlm_inputs]

	# Pad input_ids to same length (simplified like model implementation)
	max_seq_len = max(ids.shape[1] for ids in input_ids_list)
	padded_input_ids = []
	padded_attention_masks = []

	for ids, mask in zip(input_ids_list, attention_mask_list):
	if ids.shape[1] < max_seq_len:
	padding_size = max_seq_len - ids.shape[1]
	# Pad input_ids
	padding = torch.zeros(ids.shape[0], padding_size, dtype=ids.dtype, device=ids.device)
	padded_ids = torch.cat([ids, padding], dim=1)
	# Pad attention_mask
	if mask is not None:
	mask_padding = torch.zeros(mask.shape[0], padding_size, dtype=mask.dtype, device=mask.device)
	padded_mask = torch.cat([mask, mask_padding], dim=1)
	else:
	padded_mask = None
	else:
	padded_ids = ids
	padded_mask = mask

	padded_input_ids.append(padded_ids)
	padded_attention_masks.append(padded_mask)

	# Batch everything
	return {
	'input_ids': torch.cat(padded_input_ids, dim=0),
	'pixel_values': torch.cat([pv for pv in pixel_values_list if pv is not None], dim=0) if pixel_values_list and any(pv is not None for pv in pixel_values_list) else None,
	'image_grid_thw': torch.cat([igt for igt in image_grid_thw_list if igt is not None], dim=0) if image_grid_thw_list and any(igt is not None for igt in image_grid_thw_list) else None,
	'attention_mask': torch.cat([mask for mask in padded_attention_masks if mask is not None], dim=0) if any(mask is not None for mask in padded_attention_masks) else None,
	}


	def _process_language_embeddings_batch(language_embeddings: List[torch.Tensor], text_len: int = 512) -> torch.Tensor:
	"""Process and batch language embeddings with padding."""
	padded_embeddings = []

	for emb in language_embeddings:
	if emb.shape[0] <= text_len:
	padded = torch.cat([emb, emb.new_zeros(text_len - emb.shape[0], emb.shape[1])])
	else:
	padded = emb[:text_len]
	padded_embeddings.append(padded)

	# Stack to [B, seq_len, dim]
	return torch.stack(padded_embeddings, dim=0)


	def collate_fn(batch: List[Optional[Dict[str, Any]]]) -> Optional[Dict[str, Any]]:
	"""
	Universal collate function for all datasets.

	Args:
	batch: List of sample dictionaries (may contain None)

	Returns:
	Batched dictionary or None if all samples are None
	"""
	# Filter out None samples
	batch = [sample for sample in batch if sample is not None]

	if len(batch) == 0:
	return None

	# Stack tensors（支持无 initial_state 的样本）
	first_frames = torch.stack([sample['first_frame'] for sample in batch])
	video_frames = torch.stack([sample['video_frames'] for sample in batch])
	action_sequences = torch.stack([sample['action_sequence'] for sample in batch])
	has_initial_state = all(('initial_state' in sample and sample['initial_state'] is not None) for sample in batch)
	initial_states = torch.stack([sample['initial_state'] for sample in batch]) if has_initial_state else None

	# Process VLM inputs with padding in collate_fn
	vlm_inputs = [sample.get('vlm_inputs') for sample in batch]
	processed_vlm_inputs = None
	if vlm_inputs and all(vlm_input is not None for vlm_input in vlm_inputs):
	processed_vlm_inputs = _process_vlm_inputs_batch(vlm_inputs)

	# Process language embeddings with padding in collate_fn
	language_embeddings = [sample.get('language_embedding') for sample in batch if 'language_embedding' in sample]
	processed_language_embeddings = None
	if language_embeddings and any(emb is not None for emb in language_embeddings):
	processed_language_embeddings = _process_language_embeddings_batch(language_embeddings)

	result = {
	'first_frame': first_frames, # [B, C, H, W]
	'video_frames': video_frames, # [B, F, C, H, W]
	'action_sequence': action_sequences, # [B, F, D]
	'vlm_inputs': processed_vlm_inputs,
	'language_embedding': processed_language_embeddings,
	}

	if initial_states is not None:
	result['initial_state'] = initial_states

	return result# This line is no-op - we'll use sed instead