Principle:Facebookresearch Habitat lab Task Dataset Selection
| Knowledge Sources | |
|---|---|
| Domains | Configuration_Management, Embodied_AI |
| Last Updated | 2026-02-15 02:00 GMT |
Overview
Selection of task type and episode dataset through YAML configuration files that define the sensors, measures, actions, and evaluation criteria for benchmarking.
Description
Task and Dataset Selection determines what the agent is evaluated on. Task YAML files define the observation sensors, evaluation metrics, available actions, and success criteria. Dataset YAML files specify the episode data path and scene dataset. Together they fully parameterize the evaluation environment.
Habitat supports multiple tasks: PointNav, ObjectNav, ImageNav, VLN, and Rearrangement, each with their own sensors and metrics.
Usage
Select the appropriate task and dataset YAML files before configuring the Benchmark. The choice determines which metrics are reported and what the agent must accomplish.
Theoretical Basis
Task specification follows a declarative pattern:
# Abstract task definition
task:
type: "Nav-v0"
sensors: [pointgoal_with_gps_compass, rgb, depth]
measurements: [distance_to_goal, success, spl]
actions: [stop, move_forward, turn_left, turn_right]
success_distance: 0.2
dataset:
type: "PointNav-v1"
data_path: "data/datasets/pointnav/gibson/v1/{split}.json.gz"