Dataset guide · Updated November 2026

Egocentric Video Datasets: Academic vs Commercial

Every egocentric dataset decision starts with the same fork: public research datasets or commercially licensed data. The academic sets (Ego4D, EPIC-KITCHENS, EgoDex) are large and well-benchmarked but carry research-shaped licenses; the commercial sources (Luel's catalog, commissioned collection) exist precisely because the open licenses stop at the production door. This page compares them row by row — scale, sensors, annotation, and the license terms that actually govern.

All facts below are sourced first-party (fetched November 2026) — ledger in docs/seo/claims/egocentric-video-datasets.md.

The comparison

DatasetKindScaleSensors / modalityAnnotationsLicenseSource
Ego4DAcademic3,670 h · 923 participants · 74 locations · 9 countries7 wearable camera types; portions add gaze, stereo, audio, 3D scansTextual narrations; episodic memory / HOI / forecasting benchmarksExecuted license agreement defines permitted usesego4d-data.org
EPIC-KITCHENS-100Academic100 h · 45 kitchens · 4 citiesFull HD head-mounted camera + audio90K action segments · 20K unique narrations · 97 verb / 300 noun classesCC BY-NC 4.0 (non-commercial research); commercial license via University of Bristolepic-kitchens.github.io
EgoDexAcademic829 h · 194 tabletop tasksApple Vision Pro / ARKit egocentric capture3D head, upper-body and finger pose at recording time + language annotationCC BY-NC-ND (dataset)apple/ml-egodex
Luel — Diverse Egocentric POVCommercial1,000+ h · 50+ environments · 400+ contributorsPineye egocentric camera systemRich metadata; action-narration, 3D hand tracing on requestCommercially licensed (commercial-use rights confirmed up-front)luel.ai
Luel — General Egocentric VideoCommercial20,000+ unique tasksEgocentric video + synchronized IMU (accelerometer + gyroscope); ~95% hand visibilityAction-narration, 3D hand tracing; more on requestCommercially licensed (commercial-use rights confirmed up-front)luel.ai
Claru — custom collectionCommercialHousehold humanoid program: 344K+ h · 220+ tasks · 32 countries (2025 collection: 386K+ clips, ~500 contributors that program)Wearable cameras, smartphones, activity-specific rigsDepth, pose, segmentation, optical flow, captions, action labelsCommercially licensed + consent documentation includedclaru.ai

License column reflects the license text itself: "executed agreement" means the license you sign defines permitted uses — never infer commercial eligibility from download access.

If your deployment set cannot come from a research license, commissioned collection is the other column of this table.

Scope licensed egocentric data

How to select between them

  1. 01

    License first, scale second

    Decide what you are legally allowed to train and deploy before admiring hour counts. Ego4D's permitted uses are defined by its executed agreement; EPIC-KITCHENS is CC BY-NC with a commercial path through the University of Bristol; EgoDex is CC BY-NC-ND. If production use is the goal, the commercial column is the shortlist.

  2. 02

    Distribution match

    Datasets are frozen distributions: 45 kitchens in 4 cities (EPIC-KITCHENS-100), tabletop tasks on one device (EgoDex), 9 countries (Ego4D). If your robot operates elsewhere, catalog data pretrains but does not finish the job — custom collection controls tasks, environments and camera geometry.

  3. 03

    Annotation and enrichment

    Check what arrives computed: narrations and benchmarks (Ego4D), dense action segments (EPIC-KITCHENS-100), 3D hand/finger pose (EgoDex), IMU pairing and hand tracing (Luel), or full enrichment stacks — depth, pose, segmentation, flow, captions, action labels (Claru).

  4. 04

    Sensor pairing

    Modern VLA stacks consume video plus IMU, depth or proprioception. EgoDex's pose-at-recording and Luel's per-clip IMU synchronization are examples of the sensor pairing to demand; raw video alone is increasingly not enough.

Ego4D — the benchmark consortium dataset

Ego4D is "3,670 hours of video collected by 923 unique participants from 74 worldwide locations in 9 different countries," led by "an international consortium of 13 universities in partnership with Facebook AI." Seven wearable camera types were deployed (GoPro, Vuzix Blade, Pupil Labs, ZShades, ORDRO EP6, iVue Rincon 1080, Weeview), and portions add "3D scans, audio, gaze, stereo, multiple synchronized wearable cameras, and textual narrations."

Access is gated by the license agreement: review and execute it through the sign-up flow, and "you will be emailed a set of AWS access credentials when your license agreement is approved, which will take ~48hrs" — credentials expire after 14 days, with local download expected rather than cloud consumption. Permitted uses are defined by that agreement; treat Ego4D as a research and benchmarking source unless your counsel says otherwise.

EPIC-KITCHENS-100 and HD-EPIC — dense action annotation

EPIC-KITCHENS-100 is "the large-scale dataset in first-person (egocentric) vision": "100 hours of recording - Full HD" across "45 kitchens - 4 cities," with 20M frames, "90K action segments," "20K unique narrations" and 97 verb / 300 noun classes. In March 2025 the team released HD-EPIC, a CVPR 2025 "large-scale highly-detailed validation dataset" (see hd-epic.github.io).

The license is CC BY-NC 4.0: non-commercial research by default, with commercial use requiring a separate agreement with the University of Bristol team. For food-preparation and kitchen-workflow distributions it remains the densest public annotation set available.

EgoDex — dexterous manipulation at recording time

Apple's EgoDex has "829 hours of egocentric video with paired 3D hand and finger tracking data collected at the time of recording" across "194 different tabletop tasks," captured with Apple Vision Pro / ARKit. Because pose is captured by the recording device rather than estimated later, the hand-object geometry is unusually clean.

The dataset is available under "CC-by-NC-ND terms" (per the ml-egodex repository) — research use, no derivatives, no commercial training. It is the reference set for dexterous manipulation research, and off-limits for production policies.

Luel — commercially licensed egocentric catalog data

Luel is the marketplace route: Diverse Egocentric POV ("1,000+ hours of footage" across "50+ real-world environments" from "400+ unique contributors" on the Pineye camera system) and General Egocentric Video ("20,000+ distinct tasks," every clip "synchronized with IMU data," "~95% hand visibility on average"). Annotations include "action-narration, 3D hand tracing, and more upon request."

The licensing posture is the differentiator: "All datasets include commercial-use rights confirmed up-front," with licensing arranged with their team per use case.

Claru — commissioned collection at program scale

When no catalog matches your distribution, Claru collects it. The household humanoid program holds 344,000+ accepted hours across 220+ distinct household tasks in 32 countries (126 tasks past 100 hours, 62 past 1,000 hours); the 2025 egocentric collection delivered 386K+ clips from approximately 500 contributors that program, across wearable-camera (219K), smartphone (155K) and activity-specific (12K) pipelines; the workplace program adds 9,800+ hours across 17 categories in 42 countries.

Scope requested annotation and derived enrichment, delivery artifacts, commercial usage rights and consent documentation for your collection program. Evidence: 2025 collection case study and the collection service page.

When to commission capture instead

Commission collection when your target distribution is specific — tasks, environments, camera geometry, sensor pairing — or when enrichment must align with your pipeline. Open datasets are frozen; your robot's world is not.

The practical pattern most teams adopt: pretrain and benchmark on open data where the license allows (Ego4D for breadth, EPIC-KITCHENS for dense action, EgoDex for manipulation), buy catalog data (Luel) for immediate licensed coverage, and commission collection (Claru) for the deployment distribution itself. The best egocentric data providers guide covers the supplier side of that decision.

Frequently Asked Questions

What is the largest egocentric video dataset?

Ego4D is a major public benchmark with 3,670 hours from 923 participants across 74 locations in 9 countries, not the largest by hours. RekaDaily-10k reports 10,865 untrimmed hours and 10,200 processed egocentric hours under Apache 2.0. Compare released hours, tasks and terms rather than assuming one dataset leads every measure. Among commercial programs, Claru's household humanoid program holds 344,000+ accepted hours across 220+ distinct household tasks in 32 countries — an order of magnitude more, but collection is commissioned rather than downloaded.

Can I use Ego4D for commercial training?

Only as the executed license agreement defines — the license text governs, not the ability to download. Access itself is administrative: review and accept the license through the sign-up flow, and access credentials arrive when your agreement is approved (~48 hours per the dataset docs, expiring after 14 days). Whether commercial training is a permitted use under your agreement is a legal review question. Do not assume commercial eligibility from download access alone.

Which egocentric datasets allow commercial use?

The open academic sets generally do not: EPIC-KITCHENS ships under CC BY-NC 4.0 (commercial use requires contacting the University of Bristol for a separate license) and Apple's EgoDex dataset is available under CC BY-NC-ND terms. For commercial training, the realistic sources are commercial datasets with rights confirmed up-front (Luel's egocentric catalog) or commissioned collection (Claru), where licensing and consent documentation are part of the delivery.

How do egocentric datasets differ in sensors and annotation?

Ego4D adds gaze, stereo, audio, 3D scans and textual narrations across seven wearable camera types; EPIC-KITCHENS-100 provides dense action segmentation (90K action segments, 20K unique narrations); EgoDex pairs 829 hours of egocentric video with 3D hand and finger tracking captured at recording time on Apple Vision Pro; Luel's General Egocentric Video offering states synchronized IMU data and ~95% hand visibility; its separate Diverse POV page lists 1,000+ hours, 50+ environments and 400+ contributors. Confirm requested annotation and derived enrichment with each supplier, including Claru.

When should I commission custom egocentric collection instead of using an existing dataset?

Commission collection when (1) your target distribution is specific — particular tasks, environments, camera geometry or demographics a catalog cannot match; (2) you need enrichment layers (depth, pose, segmentation, action labels) aligned to your pipeline; or (3) you need airtight commercial licensing with consent documentation. Use existing datasets for pretraining, benchmarking and method development where their licenses allow.

Commercially licensed egocentric data

Skip the license archaeology — collect what you need

344K+ household humanoid hours across 220+ tasks, 9.8K+ workplace hours across 17 categories — enriched, QA-verified, commercially licensed. Tell us your distribution.

Talk to the Claru team