Implementation:Haosulab ManiSkill ReplicaCADSceneBuilder

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

Overview

Concrete tool for building scenes from the ReplicaCAD dataset in ManiSkill, handling background stages, static and dynamic objects, articulated objects, lighting, and collision configuration for the Fetch robot.

Description

ReplicaCADSceneBuilder is the base scene builder for the ReplicaCAD dataset (https://aihabitat.org/datasets/replica_cad/). It reads JSON scene configuration files to construct complete indoor environments. The builder loads the scene background (stage) as a static actor, then iterates over object instances to create either dynamic (movable) or static actors. Articulated objects (such as kitchen cabinets with doors) are loaded from URDF files.

Key implementation details:

All ReplicaCAD assets are rotated 90 degrees via a quaternion transform because ReplicaCAD uses a different coordinate convention than SAPIEN/ManiSkill.
Collision groups are configured to prevent the Fetch robot's wheels and base from colliding with the scene background and certain ground-level objects (mats, rugs, carpets).
The builder supports ray-traced lighting with an HDR environment map or standard ambient/point lighting.
Navigable positions are cached from precomputed mesh files for path planning.
The class is registered as a scene builder via the @register_scene_builder("ReplicaCAD") decorator.

Usage

Use this builder for any task that requires a ReplicaCAD indoor environment. It serves as the base class for more specific builders like ReplicaCADRearrangeSceneBuilder.

Code Reference

Source Location

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

Signature

@register_scene_builder("ReplicaCAD")
class ReplicaCADSceneBuilder(SceneBuilder):
    robot_initial_pose = sapien.Pose(p=[-1, 0, 0.02])
    builds_lighting = True
    build_configs: list[str] = None

    def __init__(self, env, robot_init_qpos_noise=0.02, include_staging_scenes=False): ...
    def build(self, build_config_idxs: Union[int, list[int]]): ...
    def initialize(self, env_idx: torch.Tensor): ...
    def disable_fetch_move_collisions(self, actor: Actor, disable_base_collisions=False): ...
    @property
    def navigable_positions(self) -> list[trimesh.Trimesh]: ...
    @cached_property
    def ray_traced_lighting(self) -> bool: ...

Import

from mani_skill.utils.scene_builder.replicacad import ReplicaCADSceneBuilder

I/O Contract

Method	Input	Output	Description
`__init__`	`env`, `robot_init_qpos_noise=0.02`, `include_staging_scenes=False`	Populated `build_configs` list	Loads scene config JSON and caches navigable positions
`build`	`build_config_idxs: Union[int, list[int]]`	Populates `scene_objects`, `movable_objects`, `articulations`, `bg`	Builds stage, objects, articulations, lighting
`initialize`	`env_idx: torch.Tensor`	None	Resets object poses, articulation qpos, and robot position
`disable_fetch_move_collisions`	`actor: Actor`, `disable_base_collisions=False`	None	Sets collision group bits for Fetch robot avoidance

Usage Examples

# Build a ReplicaCAD scene
scene_builder = ReplicaCADSceneBuilder(env)
scene_builder.build(build_config_idxs=[0] * env.num_envs)
scene_builder.initialize(env_idx=torch.arange(env.num_envs))

# Access navigable positions for path planning
nav_positions = scene_builder.navigable_positions

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment