Ego4D is the largest egocentric (first-person) video dataset, created by Meta AI Research. It contains 3,670 hours of daily-life video captured by 931 participants wearing cameras across 74 locations in 9 countries. It includes annotations for episodic memory, hand-object interaction, audio-visual diarization, social interaction, and forecasting.

Can I use Ego4D commercially?

No. Ego4D is released under CC-BY-NC-4.0 (non-commercial). You must register at ego4d-data.org to access the data, and commercial use is prohibited.

Ego4D is approximately 7 TB in total, making it one of the largest video datasets in any domain.

Ego4D: The Largest Egocentric Video Dataset

3,670 hours of first-person video from 931 participants across 9 countries. From Meta AI Research.

CC-BY-NC -- Non-Commercial MP4 + JSON ~7 TB

Key Stats

Metric	Value
Duration	3,670 hours of first-person video
Participants	931 across 74 locations in 9 countries
Size	~7 TB
Modalities	RGB video, audio, gaze tracking, hand tracking, 3D environment scans
Benchmarks	Episodic memory, hand-object interaction, AV diarization, social interaction, forecasting
License	CC-BY-NC-4.0 (requires registration)
Origin	Meta AI Research

What is Ego4D?

Ego4D is the largest egocentric video dataset ever created, produced by Meta AI Research in collaboration with 13 universities and labs worldwide. It captures 3,670 hours of daily activities -- cooking, shopping, crafting, socializing -- filmed from the first-person perspective of 931 participants across 9 countries.

What sets Ego4D apart from other video datasets is the depth of its annotations. Beyond standard action labels, it includes narrations, object detections, hand-object contact annotations, gaze tracking, and 3D environment scans. This enables five distinct benchmark tracks: episodic memory (finding past events), hand-object interaction, audio-visual diarization, social interaction understanding, and future activity forecasting.

Why roboticists care

Ego4D provides the most extensive source of egocentric hand-object interaction data available. For robotics, this translates to: better understanding of human manipulation strategies that robots can imitate, training data for action anticipation modules, and visual priors for grasp detection and contact prediction. Several recent VLA models use Ego4D-derived features for pre-training their visual encoders.

Access

Note: Ego4D requires registration and is CC-BY-NC-4.0 licensed (non-commercial use only).

Related datasets

EPIC-KITCHENS -- focused egocentric kitchen activities
Assembly101 -- egocentric + multi-view assembly tasks
HumanML3D -- human motion with language descriptions

Need egocentric data for your robot?

We can capture custom first-person manipulation data paired with robot teleoperation recordings.

Request Custom Data Contact Us