Principle:Facebookresearch Habitat lab Data Recording

Knowledge Sources	Habitat 3.0 Habitat-Lab
Domains	Human_in_the_Loop, Data_Collection
Last Updated	2026-02-15 02:00 GMT

Overview

Structured capture and persistence of interactive session data including per-frame actions, episode metadata, and user connection records for later analysis.

Description

Data Recording captures all relevant data from interactive HITL sessions: per-frame agent actions and observations, episode-level metrics, user connection/disconnection events, and session metadata. The data is serialized to compressed JSON or pickle format and optionally uploaded to cloud storage (S3) for large-scale data collection studies.

Usage

Enable during HITL interactive sessions when collecting data for analysis, user studies, or training demonstration datasets.

Theoretical Basis

Session recording follows a hierarchical data structure:

1. Session level: User connections, total duration, session metadata
2. Episode level: Start/end events, task configuration, metrics
3. Frame level: Agent positions, actions, observations, interaction events

# Abstract recording structure
session = SessionRecord(
    episodes=[
        EpisodeRecord(
            frames=[FrameData(actions, observations, metrics), ...],
            metadata={task, scene, duration}
        ),
        ...
    ],
    connections=[ConnectionRecord(user_id, connect_time, disconnect_time)]
)
serialize_and_save(session, format="json_gzip")

Related Pages

Implemented By

• Implementation:Facebookresearch_Habitat_lab_SessionRecorder_init