Libero Scenes

LW-BenchHub utilizes USD-based scene environments built on Isaacsim's robust simulation foundation for the Lightwheel-Libero-Tasks, providing diverse manipulation scenarios across different room types.

USD-Based Scene Architecture

Libero scenes are built using Universal Scene Description (USD) format, providing:

• High-fidelity rendering with realistic lighting and materials
• Fully interactive objects including furniture, containers, and manipulable items
• Physics-accurate simulation through IsaacLab's physics engine
• Optimized performance for large-scale parallel training

Scene Distribution System

Remote Scene Loading

LW-BenchHub implements a distributed scene management system that automatically handles scene acquisition:

# Scenes are automatically downloaded and cached
scene_name = "libero-1-1"    # Scene identifier

Caching Architecture

• Local cache: ~/.cache/lightwheel_sdk/floorplan/
• Automatic versioning: Ensures scene consistency across updates
• Incremental download: Only downloads new or updated scenes
• ZIP compression: Efficient scene package distribution

Scene Configuration

Scene Types

Scene ID	Scene Type	Description	Key Features
libero-1-1	Kitchen Scene	Kitchen environment with table top setup	Kitchen-specific objects and furniture layout
libero-2-2	Living Room Scene	Living room environment for manipulation tasks	Living room furniture and object arrangements
libero-4-4	Study Scene	Study room environment with desk and storage	Study-specific furniture and object placement
libero-5-5	Floor Task Scene	Floor-level manipulation environment	Ground-level tasks with floor-based objects
libero-8-8	Kitchen Scene	Different kitchen background for advanced tasks	Alternative kitchen layout with complex scenarios

Technical Implementation

SceneLoader Class

The core scene management is handled by the FloorplanLoader:

from lw_benchhub.core.scenes.loader.floorplan import FloorplanLoader

# Initialize loader with remote server
loader = FloorplanLoader(host="scene_server_host")

# Asynchronous scene acquisition
loader.acquire_usd(scene='libero', layout_id=1, style_id=1)
usd_path = loader.wait_for_result()

Scene Integration

class LiberoEnvCfg(LwScene):
    scene_name: str = "libero-1-1"
    layout_id: int = 1
    style_id: int = 1
    scene_type: str = "libero"
    
    # Enable specific objects for tasks
    enable_objects: List[str] = ["mokapot", "stovetop"]

Object Management

Interactive Elements

Libero scenes include fully functional objects across different environments:

Kitchen Scenes (1-1, 8-8)

• Containers: Baskets, bowls, plates, mugs
• Food Items: Alphabet soup, cream cheese, butter
• Beverages: Wine bottles, various containers
• Utensils: Kitchen tools and accessories

Living Room Scene (2-2)

• Furniture: Sofa, foods, mugs
• Decorative Items: Household objects and accessories
• Containers: Living room appropriate storage items

Study Scene (4-4)

• Desk Accessories: Books, study materials, desk organizers
• Storage: Shelves, cabinets, storage containers
• Study Tools: Writing instruments, reference materials

Floor Task Scene (5-5)

• Ground-level Objects: Floor-based containers and items
• Low-height Furniture: Accessible from ground level
• Floor Storage: Ground-level storage solutions

Task-Scene Matching Examples

task_scene_mapping = {
    # Kitchen tasks
    "L90K9TurnOnTheStove": "libero-1-1",
    "LSPickUpTheBlackBowlFromTableCenterAndPlaceItOnThePlate": "libero-8-8",
    
    # Living room tasks
    "L10L2PutBothTheCreamCheeseBoxAndTheButterInTheBasket": "libero-2-2",
    
    # Study room tasks
    "L90S4PickUpTheBookOnTheLeftAndPlaceItOnTopOfTheShelf": "libero-4-4",
    
    # Floor tasks
    "LOPickUpTheAlphabetSoupAndPlaceItInTheBasket": "libero-5-5",
}

These Libero scenes provide a comprehensive testing environment for evaluating robotic manipulation capabilities across various scenarios and complexity levels.