Egocentric Capture Hardware: Choosing a Camera for Data Collection

Quick facts

Baseline rig

Head-mounted RGB, 1080p/30, ≥120° horizontal FOV

Highest throughput

Phone-on-head (collector-owned) or GoPro head-mount

Synchronized sensors

Project Aria — RGB + gaze + IMU 1000/800 Hz + SLAM

3D hand ground truth

Apple Vision Pro — device-native hand/finger tracking

Optional streams

Head/wrist IMU, gaze, metric depth, 21-keypoint hand pose

Sensor sync tolerance

Wrist IMU aligned within ~20 ms of the video clock

Delivery

H.265 + per-clip metadata; consent artifact per clip

Comparison

What counts as an egocentric capture camera

An egocentric capture camera is a head-mounted device that records the world from the wearer's own point of view while they do a real task — hands entering frame from below, gaze leading the motion, the same viewpoint a robot policy will see at inference. That last part is the whole reason the category exists. Egocentric data is first-person video or sensor data captured from the perspective of a person performing real tasks, and the camera is what fixes the viewpoint, the sensor stack, and — quietly — the consent story.

Three things get miscalled egocentric and shouldn't. A chest-mount sees your hands from the wrong angle and misses whatever you look up at. A handheld phone isn't head-anchored, so the frame stops tracking gaze the moment you set it down. A dashcam or in-cabin camera is vehicle-mounted, not person-mounted — a different modality entirely. If the camera doesn't move with the head, it isn't egocentric, and a manipulation policy trained on it learns the wrong geometry.

The rest of this guide is a buyer's trade-off matrix: five rigs, scored on the five axes that actually decide a data program — sensors, field of view, cost, throughput, and consent visibility. There is no best egocentric camera, only the right rig for a given bottleneck.

The five rigs, head to head

Start with the bottleneck. If your constraint is scale and budget, a GoPro or a phone-on-head rig wins. If it's synchronized IMU and gaze, Project Aria. If it's mocap-grade 3D hand tracking, Apple Vision Pro. The table above scores each on the axes buyers weigh; the sections below explain the numbers behind them.

One axis rarely tabulated but decisive at procurement time is consent visibility — how obvious it is to a bystander that they're being recorded, which sets your release and blur obligations under egocentric data licensing before a single frame is sellable. An obvious head-strapped GoPro makes consent trivial to obtain and document; near-invisible smart glasses shift that burden onto your capture protocol. Hardware and legal risk are the same decision, not two.

GoPro head-mount: cheap RGB at scale

The GoPro is the workhorse of egocentric capture for one reason: resolution per dollar. A HERO13 Black records 5.3K60 (plus 4K120 and 2.7K240) with HyperSmooth 6.0 stabilization on a 1/1.9-inch sensor, and a 156-degree native field of view that the HB-Series Ultra Wide Lens Mod widens to 177 degrees. That wide FOV is a feature, not a bug — you capture more than the deployment view and crop down to your robot's camera later, instead of discovering at training time that the useful pixels were off-frame.

What you give up is every non-RGB signal: no IMU you can trust for head pose, no gaze, no depth. Runtime is the other catch — continuous high-resolution capture drains batteries fast, so a serious program budgets Enduro cells and hot-swaps to keep collectors recording through a shift. Build AI's Egocentric-10K — 10,000 hours, 1.08 billion frames across 192,900 clips at 1080p/30, 16.4 TB, Apache-2.0 — is the existence proof that head-mounted RGB scales to five-figure hours. It's also the reminder that raw RGB alone ships no labels and no documented consent chain, so the cheap rig buys you volume and hands you the annotation and diligence bill.

Project Aria: synchronized sensors for research-grade capture

When you need the sensors aligned rather than just the pixels sharp, Project Aria is the reference rig. Gen 1 glasses carry an RGB camera at 2880x2880 (downsampled to 1408x1408), two mono SLAM cameras at 640x480, and two IMUs running at 1000 Hz and 800 Hz — the combination that makes head pose, spatial tracking, and eye gaze recoverable instead of estimated after the fact. That's why the research corpora lean on it: EgoMimic captures human video on Aria and co-trains it with robot teleoperation, reporting that one hour of extra hand data is worth more than one hour of extra robot data. The Gen 2 wave is what has pulled smart-glasses capture from a lab curiosity toward something a collector network can run.

The catch is scale and rights. Aria units cost far more than a GoPro and are seeded rather than sold in bulk, so throughput is lower. And nearly every polished Aria corpus ships under a non-commercial or signed research license — great for benchmarking, unusable as the base of a product you plan to ship. For teleoperation-adjacent programs where head pose and gaze feed the policy, the sensor fidelity earns its cost; for raw volume it usually doesn't.

Apple Vision Pro: the dexterity rig

Apple Vision Pro is the choice when the label you care about is the hand itself. Apple built EgoDex — 829 hours of egocentric video over 194 tabletop tasks with paired 3D hand and finger tracking ^[1] — by recording on Vision Pro and reading pose straight off its on-device SLAM at capture time. No post-hoc hand-pose model matches tracking that comes from the device's own inside-out cameras; this is mocap-grade ground truth without a mocap stage, and it's why Vision Pro data is the sharpest input for dexterous-manipulation and VLA fine-tuning.

You pay for it in throughput and cost. Vision Pro is a headset, not a discreet pair of glasses, so it suits bounded, seated, tabletop capture far better than all-day in-the-wild recording, and the per-seat price caps how many you can put in the field. Treat it as your gold-standard rig for hand-precise data, not your volume rig — the two jobs want different cameras.

Smart glasses and phone-on-head: the throughput plays

The two rigs that scale fastest are the cheapest. Consumer smart glasses add an IMU and sometimes audio or gaze to a near-invisible form factor — and here the consent-visibility problem inverts, because bystanders often can't tell they're on camera, which raises your release burden even as it lowers your hardware bill. Phone-on-head is the blunt-instrument version: a modern phone strapped to a head mount gives you RGB plus a usable IMU, and on Pro models a LiDAR depth stream, on hardware collectors already own. That last fact is why it wins on throughput — you don't ship devices, you ship a mount and a capture spec, and a collector network is recording the next day.

For manipulation specifically, a handheld capture rig like the UMI gripper is the low-cost alternative that puts the camera at the end-effector instead of the head, trading the natural first-person view for a geometry that matches a wrist-mounted robot camera. Figure's Project Go-Big is the scale argument for the cheap-rig approach: it trains the Helix VLA on 100% egocentric human video collected passively as people go about tasks in real homes. You do not get that volume out of a headset.

How to choose: match geometry first, then scale

Pick the camera by working backward from your robot's deployment view, then from your volume target — in that order, because geometry mismatch is unrecoverable and volume is just money.

Geometry first. A 15-degree field-of-view mismatch changes the apparent size and position of grasped objects enough to break precise manipulation, so the rig's FOV and mounting angle have to approximate the camera your policy runs at inference. Head-mounted footage sees hands entering from below; a wrist-mounted end-effector camera looks down from 15 to 30 cm — if those don't match, no amount of data fixes it. It's also why the highest-fidelity public corpora, Ego4D's 3,670 hours across nine countries included, still can't stand in for capture matched to your specific rig.

Then sensors. RGB-only (GoPro, phone) is enough for perception and video-generation training; add IMU and gaze (Aria) for head-pose-conditioned policies; add device-native 3D hand tracking (Vision Pro) or metric depth — an Intel RealSense D455 gives stereo depth at 1280x720 up to 90 FPS, a 0.6-to-6 m range, and under 2 percent error at 4 m — when the model actually consumes it. Sync tolerance is the trap: a wrist IMU wants alignment within about 20 ms of the video clock or the extra stream is noise, not signal.

Then scale. Below a few hundred hours, buy the fidelity — Aria or Vision Pro. Above it, the economics force RGB rigs on a collector network, with sensors added only where the model earns them.

How Truelabel captures egocentric data at scale

Truelabel runs the phone-on-head and GoPro path deliberately, because the bottleneck for most buyers isn't sensor exotica — it's accepted hours of consented, task-matched footage. Collectors record on a device floor (iPhone 11+, Galaxy S21+, Pixel 6+) against your capture spec; per-clip machine QA gates stability, field of view, and hands-in-frame before a human ever reviews it; and every accepted clip ships with a consent artifact and, where the setting needs it, a location release. Sensor upgrades — head or wrist IMU, gaze, depth, 21-keypoint hand pose — are priced as add-ons on top of the RGB baseline, so you buy fidelity only where your policy consumes it.

That is the difference between a camera and a data program. The hardware sets the ceiling on signal; the consent chain, QA, and taxonomy decide whether the data is legally and technically usable. Start from your deployment view and volume target, then route the request through the physical AI data marketplace — industrial, humanoid, and custom sourcing programs all run on the same capture stack, so switching rigs never means switching vendors.

Related pages

Use these to move from category-level context into specific task, dataset, format, and comparison detail.

External references and source context

1. EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video

   EgoDex pairs egocentric video with device-native 3D hand and finger tracking captured on Apple Vision Pro using on-device SLAM.

   arXiv ↩
2. Apple Vision Pro - Technical Specifications - Apple

   Apple Vision Pro technical specifications — sensor and display reference for the tabletop dexterity rig.

   apple.com
3. Project Aria Hardware Specifications

   Project Aria Gen 1 glasses carry an RGB camera at 2880x2880 (downsampled to 1408x1408), two mono scene (SLAM) cameras at 640x480, and two IMUs at 1000 Hz and 800 Hz.

   Meta / Project Aria
4. Project Aria Gen 2 research glasses

   Aria Gen 2 announcement — the device wave pulling smart-glasses egocentric capture from a lab curiosity toward collector-scale programs.

   Meta
5. GoPro HERO13 Black

   The GoPro HERO13 Black records 5.3K60 (also 4K120 and 2.7K240) with HyperSmooth 6.0 stabilization on a 1/1.9-inch sensor, a 156-degree native field of view widened to 177 degrees by the HB-Series Ultra Wide Lens Mod.

   gopro.com
6. GoPro Enduro Battery

   GoPro Enduro battery — extended-runtime cell used to keep high-resolution head-mounted capture recording through a collector's shift.

   gopro.com
7. Project site

   The UMI gripper is a low-cost handheld capture interface that puts the camera at the end-effector rather than the head, matching a wrist-mounted robot camera's geometry.

   umi-gripper.github.io
8. Egocentric-10K

   Egocentric-10K (Build AI) is 10,000 hours of factory-worker first-person video totaling 1.08 billion frames across 192,900 clips at 1080p/30fps (16.4 TB), released under Apache-2.0.

   Hugging Face
9. Egocentric Data Collection for Robot Training

   A 15-degree field-of-view mismatch changes the apparent size and position of grasped objects enough to matter for precise manipulation, and a wrist IMU should be synced to within about 20 ms of the video stream; head-mounted footage sees hands entering from below while a wrist-mounted camera looks down from 15 to 30 cm.

   unidata.pro
10. Intel RealSense Depth Camera D455

    The Intel RealSense D455 captures stereo depth at 1280x720 up to 90 FPS with an 86-by-57-degree depth field of view, an ideal range of 0.6 to 6 m, and less than 2 percent depth error at 4 m.

    Intel RealSense
11. Egocentric video remains useful but incomplete for robot data buyers

    Ego4D provides over 3,670 hours of daily-life first-person video from 931 participants across 74 worldwide locations in 9 countries, captured with multiple synchronized wearable cameras.

    ego4d-data.org

FAQ