lerobot

/

smolvla_base

like 272

Follow

LeRobot 2.49k

🤖 Automated Policy Migration to PolicyProcessorPipeline

This PR migrates your model to the new LeRobot policy format using the modern PolicyProcessorPipeline architecture.

What Changed

✨ New Architecture - PolicyProcessorPipeline

Your model now uses external PolicyProcessorPipeline components for data processing instead of built-in normalization layers. This provides:

• Modularity: Separate preprocessing and postprocessing pipelines
• Flexibility: Easy to swap, configure, and debug processing steps
• Compatibility: Works with the latest LeRobot ecosystem

🔧 Normalization Extraction

We've extracted normalization statistics from your model's state_dict and removed the built-in normalization layers:

• Extracted patterns: normalize_inputs.*, unnormalize_outputs.*, normalize.*, unnormalize.*, input_normalizer.*, output_normalizer.*
• Statistics preserved: Mean, std, min, max values for all features
• Clean model: State dict now contains only core model weights

📦 Files Added

• preprocessor_config.json: Configuration for input preprocessing pipeline
• postprocessor_config.json: Configuration for output postprocessing pipeline
• model.safetensors: Clean model weights without normalization layers
• config.json: Updated model configuration
• train_config.json: Training configuration
• README.md: Updated model card with migration information

🚀 Benefits

• Backward Compatible: Your model behavior remains identical
• Future Ready: Compatible with latest LeRobot features and updates
• Debuggable: Easy to inspect and modify processing steps
• Portable: Processors can be shared and reused across models

💻 Usage

# Load your migrated model
from lerobot.policies import get_policy_class
from lerobot.processor import PolicyProcessorPipeline

# The preprocessor and postprocessor are now external
preprocessor = PolicyProcessorPipeline.from_pretrained("your-model-repo", config_filename="preprocessor_config.json")
postprocessor = PolicyProcessorPipeline.from_pretrained("your-model-repo", config_filename="postprocessor_config.json")
policy = get_policy_class("your-policy-type").from_pretrained("your-model-repo")

# Process data through the pipeline
processed_batch = preprocessor(raw_batch)
action = policy(processed_batch)
final_action = postprocessor(action)

Generated automatically by the LeRobot policy migration script