Description

This PR introduces a new dataset functor architecture in EmbodiChain, enabling efficient export of environment episodes to the LeRobot dataset format. The design adopts a functor-based manager pattern, supporting multi-modal data and flexible configuration.

Key Features

• Functor Pattern: Dataset export logic is encapsulated in functor classes (e.g., LeRobotRecorder)
• Unified Manager: DatasetManager orchestrates functor calls per environment step, supporting multiple dataset formats
• Multi-Modal Support: Handles RGB images (including stereo), proprioceptive states, and actions
• Flexible Configuration: All parameters (e.g., video/image mode, metadata) are configurable via gym config
• Consistent API: Usage matches event/observation manager patterns

Usage Example

# In gym config (JSON):
"dataset": {
    "lerobot": {
        "func": "LeRobotRecorder",
        "mode": "save",
        "params": {
            "save_path": "/tmp/lerobot_data",
            "robot_meta": {...},
            "extra": {"scene_type": "kitchen"},
            "use_videos": true
        }
    }
}

# In environment loop
obs, info = env.reset()
for step in range(max_steps):
    action = ...
    obs, reward, done, info = env.step(action)
    # DatasetManager automatically records and saves episodes

Technical Details

• Efficient Storage: Data is buffered and saved per episode, supporting parallel environments
• Extensible: New dataset formats can be added by implementing functor classes
• Configurable: All functor parameters are set via config

Dependencies

• Requires lerobot package for LeRobot export: pip install lerobot

Type of Change

• New feature (LeRobot export)

Benefits

• Performance: Efficient episode saving
• Extensibility: Easy to add new dataset formats
• Consistency: Unified manager/functor pattern across events, observations, datasets
• Interoperability: Direct export to LeRobot format for visualization and training