Principle:Facebookresearch Habitat lab Dataset and Scene Preparation
| Knowledge Sources | |
|---|---|
| Domains | Embodied_AI, Task_Planning |
| Last Updated | 2026-02-15 02:00 GMT |
Overview
Provisioning of articulated 3D scenes and PDDL-based task definitions required for multi-step object rearrangement training.
Description
Dataset and Scene Preparation for rearrangement tasks extends basic environment setup with articulated object assets (objects with joints, drawers, fridges) and PDDL (Planning Domain Definition Language) task specifications. ReplicaCAD and HSSD scene datasets contain interactive objects that agents must manipulate. The PDDL domain files define entity types, predicates, and actions that describe valid task goals (e.g., "set the table" requires placing specific objects at target locations).
This preparation ensures the simulation can instantiate scenes with graspable, articulatable objects and that the task planner has access to valid goal specifications.
Usage
Perform this step before training any rearrangement or manipulation policy. Required for both hierarchical RL training and human-in-the-loop evaluation with interactive objects.
Theoretical Basis
The PDDL formalization enables structured task specification:
- Domain definition: Entity types (robot, object, receptacle), predicates (on, in, holding), and action schemas (pick, place, navigate)
- Problem specification: Initial state + goal conditions expressed in predicate logic
- Task graph: The planner (or learned policy) must find a sequence of actions transforming the initial state to satisfy all goal predicates
Abstract task structure:
# PDDL-like task specification
initial_state = {"plate": "on_counter", "cup": "in_cabinet"}
goal_state = {"plate": "on_table", "cup": "on_table"}
# Agent must plan: navigate->open_cabinet->pick_cup->navigate->place_cup->...