
Smart glasses training data is first-person video and sensor data recorded from the wearer's point of view: hands, objects, gaze and audio as the user sees and hears them. Most public egocentric datasets, including Meta's Aria sets, are licensed for non-commercial research only, so teams shipping a product usually commission consented capture they own the rights to.
You can buy AI glasses in a shopping mall, but you probably cannot legally train your glasses model on the datasets everyone cites. Global AI glasses shipments rose 263% year over year in the first half of 2026, with display-less models accounting for 96% of the total, according to Counterpoint Research — and the largest public first-person datasets behind that wave are still licensed for non-commercial research only.

That gap is the whole problem with smart glasses training data. The research community solved the hard part years ago: there is a lot of high-quality egocentric video in the world, carefully annotated, captured with consent. What there is not much of is first-person data a company can put into a shipping product without breaching a license. This guide covers what wearable AI models need to learn, why the public sets will not get you there, and what commercially usable capture actually looks like.
Egocentric data is video and sensor data recorded from the point of view of the person doing the task, using a camera mounted on the head, chest or glasses. A smart glasses assistant sees exactly that view and nothing else, so a model trained on third-person footage is learning from a camera position its product will never occupy.
The difference is not cosmetic. From the wearer's viewpoint, hands enter the frame from the bottom, objects are held close and partially occluded, the camera moves whenever the head moves, and the most important thing in the scene is often what the wearer is looking at rather than what is centred. Those are the exact conditions an "what am I holding" or "what do I do next" feature has to handle. The same argument applies to multimodal models generally, which is why real-world video data matters more than more parameters.
Wearable camera video data also carries signals a fixed camera cannot produce:
Because their licenses say you cannot. The best-known smart glasses datasets were built by and for research consortia, and their terms allow non-commercial research only — which is the whole reason this page exists.
The Aria Everyday Activities Dataset License Agreement is explicit about it:
Ego4D sits in the same place. It is a 3,670-hour, 923-participant collection from 74 locations in 9 countries, released under a license agreement you must sign and have approved before you receive the data. We covered that agreement in detail in why most egocentric datasets aren't commercially licensed.
None of this makes the research sets bad. It makes them the wrong input for a commercial roadmap. A model fine-tuned on a research-only dataset carries a licensing question into every funding round, acquisition and enterprise security review it ever faces, and "we'll swap the data later" rarely survives contact with a retraining budget.
It has to answer questions about the wearer's current situation, which is a harder target than labelling a clip. In practice that breaks into a handful of capabilities, each with its own data demands:

There is a second, quieter problem: the device. Ego4D deliberately spread capture across seven different head-mounted cameras — GoPro, Vuzix Blade, Pupil Labs, ZShades, ORDRO EP6, iVue Rincon 1080 and Weeview — so that models would not overfit to one rig. That is good science, and it still means the footage does not look like what your glasses produce.
Research hardware has also moved. Meta's Aria Gen 2 glasses carry four SLAM cameras, one RGB camera and two eye-tracking cameras, plus PPG sensors and a contact microphone, with 6-8 hours of battery against the 1-2 hours of Gen 1. If your production device has one RGB camera and an IMU, a dataset recorded on a richer rig trains a model that expects inputs you will not have at inference time. Matching the capture rig to the shipping device matters as much as matching the viewpoint.
You commission it, with the rights written down before capture starts. A workable brief covers five things:

That is the service Firsthand runs: paid contributors capturing first-person video to a written spec across phone, head-rig and multi-camera tiers, anonymized before anything ships, with the same QA bar at each tier.
Being straight about the limits saves everyone a discovery call:
The smart glasses market grew 263% year over year in H1 2026, and the data situation did not move with it. The public egocentric sets are excellent research assets and unusable product inputs: Aria's license confines use to non-commercial research and forbids redistributing derived works, and Ego4D's footage was captured on seven assorted head cameras rather than on anything you will ship. If you are building wearable AI for a product, the practical path is commissioned, consented, commercially licensed capture on hardware that resembles your own — briefed once, owned outright, and defensible in the diligence that follows.
It depends on the device, but research glasses collect far more than video. Meta's Aria Gen 2 carries four SLAM cameras, an RGB camera, two eye-tracking cameras, IMU, barometer, magnetometer, microphones, GNSS, PPG sensors and a contact microphone. Consumer AI glasses are usually narrower: one camera, microphones and motion sensors, with processing split between the glasses and a phone.
Egocentric data is video and sensor data recorded from the viewpoint of the person performing a task, using a head, chest or glasses-mounted camera. It captures hands, tools and objects as the wearer sees them, along with head motion and often gaze and audio. That viewpoint is what wearable assistants and embodied AI models have to operate from at inference time.
It is first-person recordings used to teach robots manipulation and navigation, either from a human wearing a camera or from a camera on the robot itself. Human-worn capture is far cheaper to scale, since a person can record real tasks anywhere without a robot present, and the viewpoint transfers reasonably well to robots built around a head-mounted or chest-mounted camera.
Recording in public is broadly lawful in most jurisdictions, and glasses makers add capture indicators and privacy controls to stay within those rules. Dataset builders go further: Ego4D de-identified most footage before release, and Aria's public sets were captured with consented actors and had faces and license plates blurred by human annotation first.
Paid first-person recording is a real and growing line of work, since commercial AI teams cannot use research-licensed datasets and must commission their own. A legitimate program says in writing what is recorded, how contributors are paid, how faces and screens are anonymized, and what rights the buyer receives. Programs that are vague about any of those are worth avoiding.

Compact, ready to go anywhere
Interchangeable lens that’s upgradeable
Dual 1-inch sensors for improved clarity and low light performance
Dynamic range and 6K 360° capture
360° photo resolution at 21MP

8K 360° video recording for ultra-detailed visuals.
4K single-lens mode for traditional wide-angle shots.
Invisible selfie stick effect for drone-like perspectives.
2.5-inch touchscreen with Gorilla Glass protection.
Waterproof up to 33ft for underwater shooting.

360° photo resolution in 23MP
Slim design at 24 mm thick
Built-in image stabilization for smooth video capture.
Internal 19GB storage for photo and video storage.
Wireless connectivity for remote control and sharing.

60MP 360° still images for high-resolution photography.
5.7K 360° video recording at 30fps.
2.25-inch touchscreen for intuitive control.
USB Type-C port for fast charging and data transfer.
MicroSD card slot for expandable storage.
.png)
.png)

Try it free. No credit card required. Instant set-up.


