InternDataEngine

Appearance

YAML Configuration ​

Task configurations in InternDataEngine are defined using YAML files. This page explains the structure and available options.

Overview ​

A task YAML file defines all components needed for a simulation episode: environment, robots, objects, cameras, and skills. The workflow loads this configuration to set up the simulation.

World Settings ​

Global simulation settings are defined in the YAML files under configs/simbox:

yaml
simulator:
  physics_dt: 1/30                   # Physics update rate
  rendering_dt: 1/30                 # Render update rate
  stage_units_in_meters: 1.0         # Stage unit scale
  headless: True                     # Run without GUI; set to False for visual debugging
  anti_aliasing: 0                   # Anti-aliasing level

World Settings ​

  • physics_dt (float): Physics simulation time step (in seconds).
  • rendering_dt (float): Rendering time step (in seconds).
  • stage_units_in_meters (float): Unit scale used for the USD stage.
  • headless (bool): Run without GUI; set to False for visual debugging.
  • anti_aliasing (int): Anti-aliasing level (0 = disabled).

Task Basic Configuration ​

Each task begins with basic metadata and settings:

yaml
tasks:
  -
    name: banana_base_task      # Task identifier
    asset_root: workflows/simbox/example_assets  # Root path for all assets
    task: BananaBaseTask        # Task class name
    task_id: 0                  # Task instance ID

    offset: null                # Optional position offset
    render: True                # Enable rendering

    neglect_collision_names: ["table"]  # Collision filter names

Task Basic Configuration ​

  • name (str): Unique task identifier.
  • asset_root (str): Root directory for all USD assets.
  • task (str): Python class that implements task logic.
  • task_id (int): Instance ID for multi-task scenarios.
  • offset (list): Optional world offset [x, y, z].
  • render (bool): Enable/disable visual rendering.
  • neglect_collision_names (list): Object names to exclude from gripper collision checking.

Arena ​

Arena configuration defines static fixtures in the environment, such as tables, floors, and backgrounds. Each fixture specifies its USD asset path, position, orientation, scale, and optionally texture randomization settings.

Arena configs are stored in workflows/simbox/core/configs/arenas/ and referenced via the arena_file field.

yaml
name: example
fixtures:
  -
    name: table
    path: table0/instance.usd
    target_class: GeometryObject
    translation: [0.0, 0.0, 0.375]
    scale: [0.000525, 0.001053, 0.001056]
  -
    name: floor
    target_class: PlaneObject
    size: [5.0, 5.0]
    translation: [0, 0, 0]

Environment Map ​

Environment map controls scene lighting and HDR backgrounds. It supports randomization of lighting intensity and rotation:

yaml
env_map:
  envmap_lib: envmap_lib              # Path to HDR environment maps
  apply_randomization: True           # Enable random lighting
  intensity_range: [4000, 7000]       # Light intensity range
  rotation_range: [0, 180]            # Environment rotation range (degrees)

Robots ​

Robot configuration specifies the robot name, its configuration file path, initial orientation, and collision filter substrings:

yaml
robots:
  -
    name: "split_aloha"
    robot_config_file: workflows/simbox/core/configs/robots/split_aloha.yaml
    euler: [0.0, 0.0, 90.0]
    ignore_substring: ["material", "table", "gso_box"]

Robots ​

  • name (str): Robot identifier (must match skill definitions).
  • robot_config_file (str): Path to robot config file, relative to asset_root.
  • euler (list): Initial rotation [roll, pitch, yaw] in degrees, in world frame.
  • ignore_substring (list): Substrings for collision filtering; objects with matching name prefixes are excluded from CuRobo collision checking.

For detailed robot configuration, see Robots.

Objects ​

Objects define items in the scene that can be manipulated, along with their metadata and placement properties:

yaml
objects:
  -
    name: pick_object_left
    path: task/sort_the_rubbish/non_recyclable_garbage/obj_0/Aligned_obj.usd
    target_class: RigidObject
    dataset: oo3d
    category: bottle
    prim_path_child: Aligned
    translation: [0.0, 0.0, 0.0]
    euler: [0.0, 0.0, 0.0]
    scale: [1, 1, 1]
    apply_randomization: True
    orientation_mode: random

Objects ​

  • name (str): Unique object identifier (must match skill definitions).
  • path (str): USD file path relative to asset_root.
  • target_class (str): Object type: RigidObject, GeometryObject, ArticulatedObject, XFormObject, ConveyorObject.
  • dataset (str): Source dataset (e.g., oo3d, gso).
  • category (str): Object category for grasp detection.
  • prim_path_child (str): USD sub-prim name (default: Aligned). This prim contains the target mesh for collision and grasping.
  • translation (list): Initial position [x, y, z] in world frame.
  • euler (list): Initial rotation [roll, pitch, yaw] in degrees, in world frame.
  • scale (list): Scale factors [sx, sy, sz].
  • apply_randomization (bool): Enable pose randomization within the category.
  • orientation_mode (str): Orientation constraint: random (pure random), suggested (follows category defaults), or keep (use euler values).

For detailed object configuration, see Objects.

Regions ​

Regions define spatial constraints for object placement, specifying which object to place, the target surface, and the allowed position/orientation ranges:

yaml
regions:
  -
    object: pick_object_left
    target: table
    random_type: A_on_B_region_sampler
    random_config:
      pos_range: [
        [-0.3, -0.20, 0.0],
        [-0.025, 0.10, 0.0]
      ]
      yaw_rotation: [-45.0, 15.0]

Regions ​

  • object (str): Name of the object to place.
  • target (str): Target surface name from arena fixtures (default: table).
  • random_type (str): Sampler type for placement.
  • pos_range (list): Position range [[x_min, y_min, z_min], [x_max, y_max, z_max]] in world frame.
  • yaw_rotation (list): Yaw angle range [min, max] in degrees in world frame.

Cameras ​

Camera configuration defines viewpoint, intrinsic parameters, and extrinsic randomization:

yaml
cameras:
  -
    name: split_aloha_hand_left
    translation: [0.0, 0.08, 0.05]
    orientation: [0.0, 0.0, 0.965, 0.259]
    camera_axes: usd
    camera_file: workflows/simbox/core/configs/cameras/astra.yaml
    parent: "split_aloha/split_aloha_mid_360_with_piper/split_aloha_mid_360_with_piper/fl/link6"
    apply_randomization: False

Cameras ​

  • name (str): Camera identifier.
  • translation (list): Position offset [x, y, z] from parent link.
  • orientation (list): Rotation offset as quaternion [qx, qy, qz, qw] from parent link.
  • camera_axes (str): Coordinate convention: usd or ros.
  • camera_file (str): Path to camera parameter file.
  • parent (str): USD prim path to attach the camera to.
  • apply_randomization (bool): Enable extrinsics randomization.

For detailed camera configuration, see Cameras.

Data ​

The data section stores metadata for dataset generation:

yaml
data:
  task_dir: "sort_the_rubbish_part0"
  language_instruction: "Sort the garbage on the desktop into recyclable and non-recyclable."
  detailed_language_instruction: "Pick the bottles and place them into the recyclable trashbin with right arm, and pick the other garbage and place it into the non-recyclable trashbin with left arm."
  collect_info: ""
  version: "v2.0, head camera 1280x720, wrist 640x480"
  update: True
  max_episode_length: 4000

Data ​

  • task_dir (str): Output directory name.
  • language_instruction (str): Short task description.
  • detailed_language_instruction (str): Detailed instruction for training.
  • collect_info (str): Additional collection metadata.
  • version (str): Dataset version string.
  • update (bool): Whether to overwrite existing data.
  • max_episode_length (int): Maximum steps per episode.

Skills ​

Skills define the action sequence for the robot:

yaml
skills:
  -
    split_aloha:
      -
        right:
          -
            name: pick
            objects: [pick_object_right]
            filter_y_dir: ["forward", 60]
            filter_z_dir: ["downward", 150]
            pre_grasp_offset: 0.05
            gripper_change_steps: 10
            t_eps: 0.025
            o_eps: 1
            process_valid: True
            lift_th: 0.02
            post_grasp_offset_min: 0.10
            post_grasp_offset_max: 0.15
          -
            name: place
            objects: [pick_object_right, gso_box_right]
            place_direction: vertical
            x_ratio_range: [0.3, 0.4]
            y_ratio_range: [0.3, 0.4]
            pre_place_z_offset: 0.3
            place_z_offset: 0.3
            success_mode: height
          -
            name: heuristic__skill
            mode: home
            gripper_state: 1.0

For detailed skill configuration, see Skills.

Complete Example ​

Here is a minimal task configuration:

yaml
tasks:
  -
    name: banana_base_task
    asset_root: workflows/simbox/assets
    task: BananaBaseTask
    task_id: 0
    offset: null
    render: True

    arena_file: workflows/simbox/core/configs/arenas/example.yaml

    env_map:
      envmap_lib: envmap_lib
      apply_randomization: False

    robots:
      -
        name: "fr3"
        robot_config_file: workflows/simbox/core/configs/robots/fr3.yaml
        euler: [0.0, 0.0, 0.0]

    objects:
      -
        name: bottle
        path: objects/bottle.usd
        target_class: RigidObject
        dataset: oo3d
        category: bottle
        prim_path_child: Aligned
        translation: [0.3, 0.0, 0.0]
        euler: [0.0, 0.0, 0.0]
        scale: [1, 1, 1]
        apply_randomization: False

    regions:
      -
        object: bottle
        target: table
        random_type: A_on_B_region_sampler
        random_config:
          pos_range: [[0.2, -0.1, 0.0], [0.4, 0.1, 0.0]]
          yaw_rotation: [-30.0, 30.0]

    cameras:
      -
        name: head_camera
        translation: [0.0, 0.0, 0.5]
        orientation: [0.707, 0.0, 0.0, 0.707]
        camera_file: workflows/simbox/core/configs/cameras/realsense_d455.yaml
        parent: ""

    data:
      task_dir: "pick_bottle"
      language_instruction: "Pick the bottle."
      detailed_language_instruction: "Pick the bottle."
      collect_info: ""
      version: "v1.0"
      update: True
      max_episode_length: 500

    skills:
      -
        fr3:
          -
            name: pick
            objects: [bottle]
            pre_grasp_offset: 0.05

Released under the MIT License.

Copyright © 2024 InternDataEngine Team