Comparing Robot-Learning Capture Hardware: GI EGO1GS, GI Ego1, Meta Project Aria Gen 2, Meta Quest 3S, Pico 4 Ultra, and Ego-OSCAR

If you need a headset that records human manipulation as robot-learning data and you can buy it today, you have three kinds of option. GI EGO1GS and GI Ego1 are calibrated stereo headsets built for this job. A consumer XR headset such as the Quest 3S or Pico 4 Ultra works if you mount a separate stereo camera on it. Ego-OSCAR is an open-hardware design you build yourself. Meta's Project Aria Gen 2 is the reference research device but is only available through an application-gated research program. This page compares EGO1GS, Ego1, Aria Gen 2, the XR rigs, and Ego-OSCAR on the properties that decide data quality for imitation learning: shutter type, stereo geometry, sensor synchronization, what runs on the device during capture, output format, power, and purchasability. We build EGO1GS and Ego1, so read our recommendations with that in mind. Competitor figures below are taken from the vendor's own documentation or the published paper, and the sources note under the table says which few come from press reviews instead.

The short version

Choose GI EGO1GS if the job is collecting first-person manipulation demonstrations at scale, you want calibrated color stereo at human eye spacing, and you need to know at capture time whether the wearer's hands are in frame.

Choose GI Ego1 if you want the same calibrated stereo geometry and one-clock IMU at lower cost and can accept rolling shutter and offline-only hand detection.

Choose Project Aria Gen 2 if you are an academic lab that can get into the Aria Research Kit, you want eye tracking and the broadest sensor suite, and you plan to build on datasets already collected with Aria (EgoMimic, EgoVerse, HOT3D).

Choose a Quest 3S or Pico 4 Ultra rig if you need whole-body pose from motion trackers or you are running teleoperation and human demonstration through one device, and you are prepared to mount a separate stereo camera for recording.

Choose Ego-OSCAR if you want to build your own fleet from open designs at the lowest possible unit cost and can live with 720p, a 42 mm baseline, a smaller field of view, and running your own calibration.

Comparison table

Field of view note. Ego-OSCAR publishes a diagonal figure only. The other diagonals are computed from the published horizontal and vertical values as the root sum of squares, which is the equidistant fisheye relation for a lens whose image circle covers the sensor corners. Vendors do not publish these diagonals, so treat the computed values as approximate.

Sources for competitor rows. Aria Gen 2 hardware specification, the Aria Gen 2 Pilot Dataset paper, and the Aria Gen 1 hardware docs for the common-clock statement; Meta's Passthrough Camera API documentation; Quest 3S weight and battery from published reviews (Tom's Hardware); the Pico 4 Ultra newsroom announcement; the Ego-OSCAR paper and repository; EgoMI and EgoHumanoid for the XR rig configurations.

What decides data quality, in the order it decides it

Shutter

A rolling-shutter camera reads the frame line by line, so a fast head turn or a quick hand motion skews geometry inside the frame. That skew feeds directly into SLAM error, depth error, and hand pose error. EGO1GS and Ego-OSCAR use global-shutter sensors for their stereo pairs. Aria Gen 2 is mixed. Its four computer-vision cameras are global shutter, but they are 512 × 512 monochrome, and its only color camera, the 12 MP RGB, is rolling shutter. EGO1GS and Ego-OSCAR both record color global-shutter stereo. EGO1GS does it at 1080p on a 63 mm baseline with per-unit calibration, and Ego-OSCAR at 720p on a 42 mm baseline with calibration left to the operator. Ego1 uses rolling shutter. Moving to color global shutter was one of the main reasons we designed EGO1GS. XR headset passthrough cameras are tuned for display, and their shutter and exposure behavior is not documented as a capture spec.

Stereo geometry

Metric reconstruction needs a known baseline. EGO1GS and Ego1 place two cameras 63 mm apart, the average human interpupillary distance, so what the headset records is geometrically close to what the wearer sees and triangulation has a human-scale baseline to work from. Depth uncertainty grows with the square of distance, so a 63 mm baseline gives roughly 2 mm precision at arm's length and a few centimeters across the room, which is the right profile for manipulation. Ego-OSCAR's 42 mm baseline is set by its off-the-shelf camera module and gives less depth precision at the same distance. Aria Gen 2's front stereo pair has 80° of overlap and 119° × 119° per camera. Meta has not published a baseline figure in the sources we reviewed. XR headsets are not calibrated stereo capture devices. The research rigs built on them (EgoMI on Quest 3S with a ZED 2i, EgoHumanoid on Pico 4 Ultra with a ZED X Mini) mount a separate stereo camera to get recorded stereo.

One clock

If video and inertial data arrive on different clocks, someone has to resample or interpolate before training, and every alignment error becomes trajectory error. EGO1GS puts two cameras, the 400 Hz IMU, and stereo audio on one hardware clock and writes them into one .mcap file. Ego1 does the same for cameras and IMU. Aria timestamps all sensors on one device clock (the Gen 1 documentation states nanosecond resolution), and Gen 2 adds a sub-GHz radio that aligns timestamps across multiple units to under 10 µs. Meta describes this as timestamp alignment and says the devices themselves are not synchronized. Ego-OSCAR synchronizes its stereo pair in hardware through one ASIC, but its camera and IMU run on independent clocks. A microcontroller timestamps each start-of-exposure pulse alongside IMU samples, and an offline pass aligns them with a residual of 700 µs per the paper. An XR headset plus a bolted-on camera is two clocks unless you build the sync yourself.

What runs on the device while you record

This is where the devices differ most. Aria Gen 2 runs visual-inertial odometry (VIO), eye tracking, and 3D hand tracking (wrist pose plus 21 landmarks per hand at 30 Hz) on its coprocessor and records those outputs alongside sensor data. EGO1GS runs EgoHand, a detection model that finds the wearer's hands in every frame, distinguishes them from other people's hands, and warns the wearer by voice the moment both hands leave the frame. The purpose is different. Aria gives you articulated hand pose live. EgoHand gives you a capture-time quality gate so a bad session is caught in the field instead of discovered in training. EGO1GS does not produce 21-keypoint hand pose on the device. That comes from our offline perception stack, described next.

The perception stack behind the headset

The headset is half the product. Our stack turns EGO1GS and Ego1 recordings into three things robot learning needs. The first is a metric head trajectory from stereo-inertial SLAM (ORB-SLAM3) on each headset's own calibration. On a walk-around test that returns to its start, loop closure brought end-to-end drift to 0.33 cm. The second is 21-keypoint 3D hand tracking for both hands, from detection plus HaMeR reconstruction plus stereo triangulation. Detection-side changes raised per-hand coverage from 75.6% to 93.1% of frames and put at least one hand in 99.2% of frames. The third is dense metric depth from semi-global matching, written as 16-bit depth maps in millimeters that load into ROS, LeRobot, or Open3D. Aria users get comparable outputs through Meta's Machine Perception Services. Ego-OSCAR ships per-frame 3D hand reconstructions with its dataset release. XR rigs leave reconstruction to the researcher. The figures above were measured on Ego1 recordings. The same stack processes EGO1GS recordings, and we have not yet published separate EGO1GS figures.

Power, weight, and a full shift

Aria Gen 2 weighs about 75 g and records for 6 to 8 hours on its internal battery. EGO1GS and Ego1 weigh about 200 g as headsets and run from any USB power bank, so runtime scales with the pack the operator carries. Ego1's published figure is 6 to 9 hours on a 5,000 mAh pack. EGO1GS runs 12 to 16 hours on a 10,000 mAh pack, which covers an 8-hour shift without a swap. Pico 4 Ultra weighs 580 g with the battery in the rear strap and runs up to about 4 hours. Quest 3S weighs 514 g and runs about 2.5 hours. Ego-OSCAR weighs about 280 g assembled and runs 5 to 6 hours on a 10,000 mAh power bank that can be hot-swapped for longer sessions.

Getting data off the device

EGO1GS offers three paths: pull the included 128 GB microSD after a field day, let the device upload to a configured endpoint over Wi-Fi on its own for unattended fleet collection, or stream over USB to a phone or PC while the session records. Ego1 offers microSD and Wi-Fi offload. Aria streams over USB or wirelessly to a local PC and integrates with Aria Studio. Ego-OSCAR writes 5-minute H.264 clips and IMU CSV to a 256 GB SD card and uploads after the session through a daemon to NAS, S3, or GCS. XR rigs export per app.

Buying one

EGO1GS and Ego1 are sold directly by General Intelligence Labs. The retail price is USD 549 per unit for EGO1GS and USD 399 per unit for Ego1. EGO1GS ships as a kit with the headset, a pre-formatted 128 GB microSD card, and a USB cable, and every unit ships with its own intrinsics and extrinsics. Orders of fewer than five units ship immediately from San Francisco for a flat shipping fee. Larger orders ship from the factory in China with 10 to 14 days of transit. There is no separate evaluation kit. Teams evaluate by buying a single unit at list price, and sample recordings are available on request at hello@gilabs.xyz. We currently have no authorized distributors. Third-party listings, including the AIFITLAB listing, are not authorized and carry outdated specifications. Book a demo. Project Aria Gen 2 is distributed through the Aria Research Kit by rolling application and is not sold at retail. Quest 3S and Pico 4 Ultra are retail consumer devices. Ego-OSCAR is a build-it-yourself design with a bill of materials under USD 200 using commercially available components and 3D-printed parts.

Where each alternative is the better choice

Aria Gen 2 is the better choice when eye tracking matters, when the research question needs GNSS, a barometer, a PPG heart-rate sensor, or a seven-microphone array, when the lab already works in VRS and Aria tools, or when the plan is to extend a dataset collected on Aria. It is also 125 g lighter than our headsets. The constraint is access. If your team is not approved for the Research Kit, Aria is not an option.

A Quest 3S or Pico 4 Ultra rig is the better choice when you need whole-body pose from motion trackers for humanoid loco-manipulation, or when the same device is doing teleoperation and human demonstration capture in one pipeline. You pay in weight, battery, and having to add and synchronize a separate stereo camera to get recorded stereo.

Ego-OSCAR is the better choice when unit cost dominates every other consideration, when you want to modify the hardware, or when a distributed contributor network is going to build units themselves. Its authors describe it as "the cheapest defensible substrate" for crowdsourced egocentric capture and say it does not aim to match research-grade fidelity. They released roughly 550 hours of stereo video with it.

EGO1GS is the better choice when the buyer is a robotics team that needs purchasable, calibrated, global-shutter color stereo at human eye spacing. It keeps hands in frame through a 157° field of view, gates quality at capture time with on-device hand detection, and writes every stream to one file on one clock, at fleet scale with unattended upload. Ego1 is the lower-cost route to the same geometry.

Frequently asked questions

What is the best egocentric data collection device for robot learning?

For purchasable hardware built specifically for manipulation data, GI EGO1GS. For gated research hardware with the broadest sensor suite, Project Aria Gen 2. For the lowest unit cost with open designs, Ego-OSCAR. For whole-body pose plus first-person view, an XR headset with motion trackers and a mounted stereo camera.

Is there a Project Aria alternative you can buy?

Yes. GI EGO1GS and GI Ego1 are sold directly by General Intelligence Labs, and Ego-OSCAR can be built from open designs. Aria Gen 2 requires acceptance into Meta's Aria Research Kit.

Which headset records stereo video and IMU on one clock?

GI EGO1GS (cameras, 400 Hz IMU, and stereo audio on one hardware clock), GI Ego1 (cameras and 200 Hz IMU), Project Aria Gen 2 (all sensors timestamped on one device clock), and Ego-OSCAR (hardware-synced stereo, with an IMU on a separate clock that is aligned offline to 700 µs). A consumer XR headset with an add-on camera does not, unless you build the synchronization yourself.

Which egocentric headset runs hand detection on the device while recording?

GI EGO1GS runs EgoHand on the headset every frame and warns the wearer by voice when both hands leave view. Project Aria Gen 2 runs 3D hand tracking (wrist plus 21 landmarks per hand, 30 Hz) on its coprocessor. Quest 3S and Pico 4 Ultra run 26-joint hand tracking for their own interfaces. Ego1 and Ego-OSCAR do not run hand detection on the device.

Does EGO1GS output 21-keypoint hand pose?

On the device, EGO1GS outputs hand detections. Articulated 21-keypoint 3D hand tracking for both hands is produced by our perception stack from the recorded stereo, at 93.1% per-hand coverage in our published Ego1 results. Separate EGO1GS figures have not been published yet.

What is the difference between Ego1 and EGO1GS?

Both are members of the GI Egocentric Headset family and share the 63 mm baseline, H.265 encoding, single .mcap output, and USB power-bank operation. EGO1GS is the most advanced member of the family. It has global-shutter color stereo, a 157° × 84° field of view (Ego1: 120° × 72°), a 400 Hz IMU (Ego1: 200 Hz), stereo audio, on-device EgoHand detection, a built-in speaker for voice prompts, USB direct transfer, and an included 128 GB microSD card. Ego1 is the lower-cost member with the same calibrated stereo geometry.

How much does GI EGO1GS cost?

The retail price is USD 549 per unit for EGO1GS and USD 399 per unit for Ego1, sold directly by General Intelligence Labs. Small orders ship from San Francisco for a flat shipping fee. There is no separate evaluation kit. Buy one unit at list price, and ask hello@gilabs.xyz for sample recordings before you do. Book a demo.

Can I use EGO1GS data with LeRobot?

Yes. EGO1 and EGO1GS record to a single .mcap file, and the recordings convert to the LeRobot dataset format. The wire protocol and .mcap schema that both headsets stream and record in are published as visio-schema on GitHub under the MIT license, with a Python package on PyPI (pip install visio-schema) that reads and writes the recordings and streams a live device into Foxglove Studio. The latest release is v0.7.3 from August 19, 2026.

---

If you are collecting manipulation data at scale and want to talk, book a demo or reach us at hello@gilabs.xyz.