Implementation:Haosulab ManiSkill AI2THORSceneBuilder

Knowledge Sources	Haosulab_ManiSkill
Domains	Robotics, Simulation, Scene_Building
Last Updated	2026-02-15 08:00 GMT

Overview

Concrete tool for building AI2-THOR scenes in ManiSkill, using configurations and assets from the ai2thor-hab HuggingFace dataset to create indoor environments with static and dynamic objects.

Description

AI2THORBaseSceneBuilder extends SceneBuilder to load and construct scenes from AI2-THOR datasets (including ProcTHOR procedurally generated scenes). It reads scene instance JSON files that describe background stages, object placements, rotations, and motion types.

Key implementation details:

Loads scene configurations from a metadata JSON file (e.g., ProcTHOR.json) containing paths to individual scene instance files.
Applies a 90-degree rotation around the x-axis to convert from AI2-THOR's coordinate system to SAPIEN/ManiSkill conventions. ProcTHOR scenes additionally rotate around the y-axis.
Distinguishes between static objects (built as static actors with non-convex collision) and movable objects. Only objects whose names match WORKING_OBJS (apple, potato, tomato, lettuce, soap, sponge, plate, book) are built as dynamic actors.
Uses _should_be_static() to check object semantic IDs against a movable object mapping loaded from metadata.
Configures collision groups to prevent the Fetch robot's wheels from colliding with the scene background.
Provides per-scene robot start positions via FETCH_BUILD_CONFIG_IDX_TO_START_POS mapping.
Caches navigable positions from precomputed numpy files for path planning.

The builder creates a merged background actor from individual background objects across parallel environments.

Usage

Use this builder for tasks that require AI2-THOR indoor environments in ManiSkill. Subclass it and set the scene_dataset class attribute to specify which AI2-THOR dataset to use (e.g., "ProcTHOR").

Code Reference

Source Location

Repository: Haosulab_ManiSkill
File: mani_skill/utils/scene_builder/ai2thor/scene_builder.py

Signature

WORKING_OBJS = ["apple", "potato", "tomato", "lettuce", "soap", "sponge", "plate", "book"]

FETCH_BUILD_CONFIG_IDX_TO_START_POS = {
    0: (-3, 0), 1: (-2, -2), 2: (0, 0), 3: (-3.5, 0),
    4: (0, -2), 5: (-1, 1.5), 6: (1, -0.5), 7: (3.25, 1),
    8: (1, 2), 9: (1, 1),
}

class AI2THORBaseSceneBuilder(SceneBuilder):
    scene_dataset: str

    def __init__(self, env, robot_init_qpos_noise=0.02): ...
    def _should_be_static(self, template_name: str) -> bool: ...
    def build(self, build_config_idxs: Union[int, list[int]], convex_decomposition="none"): ...
    def initialize(self, env_idx): ...
    def disable_fetch_move_collisions(self, actor: Actor, disable_base_collisions=False): ...
    @property
    def navigable_positions(self) -> list[np.ndarray]: ...

Import

from mani_skill.utils.scene_builder.ai2thor.scene_builder import AI2THORBaseSceneBuilder

I/O Contract

Method	Input	Output	Description
`build`	`build_config_idxs`, `convex_decomposition`	Populated `scene_objects`, `movable_objects`, `bg`	Builds stage background and all objects from scene JSON
`initialize`	`env_idx`	None	Resets Fetch robot position and all object poses
`_should_be_static`	`template_name: str`	`bool`	Checks if object should be static based on semantic ID

Usage Examples

# Subclass for ProcTHOR scenes
class ProcTHORSceneBuilder(AI2THORBaseSceneBuilder):
    scene_dataset = "ProcTHOR"

# Usage
scene_builder = ProcTHORSceneBuilder(env)
scene_builder.build(build_config_idxs=[0] * env.num_envs)
scene_builder.initialize(env_idx=torch.arange(env.num_envs))

# Access navigable positions
nav = scene_builder.navigable_positions

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment