Project Aria

Project Aria is an egocentric data-collection research program run by Meta's Reality Labs Research. It was announced on September 16, 2020 ^[1][2]. The program centers on a pair of sensor-equipped research glasses, also called Project Aria, that Meta describes explicitly as a research tool rather than a consumer product. The glasses have no display and show nothing to the wearer; their purpose is to record first-person (egocentric vision) multimodal sensor data that can be used to develop the machine perception and AI needed for future augmented reality glasses and contextual AI assistants ^[2][5].

Aria is worn by trained Meta researchers and contractors, and the device and accompanying software are provided to academic and commercial research labs through the program. Meta's site describes the glasses as used "by over 200 academic and corporate partners" ^[11]; the Aria Gen 2 device whitepaper, published in October 2025, gives a larger and more specific figure of "over 290 academic and industrial partners operating more than 1,000 devices across 27 countries" ^[9]. A second-generation device, Aria Gen 2, was announced on February 27, 2025 and opened to applications on October 19, 2025, with a broad rollout targeted for the second quarter of 2026 ^[5][10].

Project Aria is distinct from Meta's consumer AI glasses and from the Orion AR prototype: it is a research instrument, never sold, and neither generation carries a display of any kind ^[14][15].

Background and purpose

When Meta (then Facebook) announced Project Aria, it framed the glasses as a way to figure out how all-day wearable AR could work in practice, and to develop the safeguards, policies, and social norms that such devices would require ^[1][2]. The first-person data the glasses capture was intended to feed research on machine perception, scene understanding, and the construction of a persistent spatial map of the world that Meta called LiveMaps ^[2].

The announcement was explicit that the glasses are not a consumer device and carry no display. As the launch post put it, the glasses "won't display any information on the inside of the lens, and research participants cannot view or listen to the raw data captured by the device" ^[2]. The breakthroughs the program targets, including enhanced audio and visual input, contextualized AI, and a lightweight frame, are presented as prerequisites for AR glasses rather than features of Aria itself ^[1][2].

The first-generation device

The first-generation Aria glasses weigh 75 grams and carry a multi-sensor suite for capturing synchronized egocentric data; the team documented the design in a 2023 device paper ^[3][4]. All cameras and other sensors are calibrated and timestamped on a shared clock with nanosecond resolution, which lets researchers fuse the streams precisely ^[3]. The glasses do not run AR applications and cannot be programmed by partners; they record and securely upload data for offline research ^[3][13].

The table below lists the Gen 1 sensor suite as documented by Meta ^[3].

Scroll sideways for more →

Nominal capture rates are lower than the maxima above: in the default profile the mono scene and eye-tracking cameras run at 10 fps and the RGB camera at 1 fps, which is what makes the 1.5-hour recording figure achievable ^[3]. Meta's own comparison page quotes Gen 1 battery life more loosely as "1-2 hrs" ^[13]. A phone companion app and desktop tools handle device control, secure upload, and access to the recorded data ^[12].

Aria Gen 2

Meta announced Aria Gen 2 on February 27, 2025, followed by a detailed technical write-up on June 4, 2025 and a device whitepaper in October 2025 ^[5][6][9]. Like the original, Gen 2 has no display and is not sold to consumers; it is aimed at researchers working on machine perception, egocentric and contextual AI, and robotics ^[5][15].

Form factor

Gen 2 weighs 74 to 76 grams depending on size and adds folding arms for storage and transport ^[6]. Meta produced eight size variations, built from two options each for frame width, temple arm length, and nose pad shape, and says wearers undergo a fitment process to pick the size that maximizes both comfort and sensing quality ^[6][9]. Gen 1 came in two frame widths. Meta says Gen 2 supports six to eight hours of continuous recording on a custom lithium-ion battery, against roughly 1.5 hours for Gen 1 ^[3][5][7]. That battery figure is vendor-reported; because the device is never sold, no independent measurement of it exists.

Sensors

The headline change is on the computer-vision side. Gen 2 carries four computer-vision cameras, double the two mono scene cameras in Gen 1, which widens the combined field of view and supports 3D hand and object tracking ^[6]. Each CV camera has a 119 by 119 degree field of view and a custom stacked digital-pixel sensor with a global shutter and 4.6 micrometer pixels in a 512x512 array; the recorded stream runs at 30 Hz ^[7][8]. An overlapped triple quantization scheme produces single-frame, single-exposure high dynamic range. Meta's blog gives that dynamic range as 120 dB against about 70 dB in Gen 1, while both the developer documentation and the whitepaper state the more conservative "exceeding 110 dB" ^[6][7][9]. Stereo overlap between the front-facing CV cameras rises to 80 degrees from Gen 1's 35 degrees, improving off-device depth reconstruction ^[6].

Three sensors are genuinely new in Gen 2 and were absent from Gen 1 ^{[5][6][7][13]}:

• A photoplethysmography (PPG) sensor in the nosepad, sampling at typically 128 Hz and up to 512 Hz, which allows heart-rate estimation. Meta positions it for physiological and affective computing studies ^[7][9].
• A contact microphone, also in the nosepad: a Knowles V2S200D sampling at up to 48 kHz, which isolates the wearer's voice from bystanders and ambient noise. Meta demonstrated it picking up a whisper in a wind tunnel where the acoustic microphones could not ^[6][7].
• A calibrated ambient light sensor with an ultraviolet channel, used for exposure control and to distinguish indoor from outdoor lighting ^[6][7].

Gen 2 also adds a proximity sensor that detects when the glasses are worn, dual-frequency GNSS (L1 plus L5 for GPS, E1 plus E5 for Galileo), Wi-Fi 6E and Bluetooth 5.3, and two force-cancelling open-ear speakers, the first audio output on an Aria device ^[5][7]. It retains an RGB camera, two eye-tracking cameras, seven spatial microphones, two IMUs, a barometer, and a magnetometer ^[5][7].

Scroll sideways for more →

On-device compute

Where Gen 1 recorded raw streams for offline processing, Gen 2 runs machine perception on the device itself, on what Meta calls a custom, energy-efficient coprocessor. Meta has not published a part number, clock speed, or process node for it. The coprocessor handles visual inertial odometry for 6DoF tracking, eye tracking (gaze per eye, vergence point, blink detection, pupil center, pupil diameter, corneal center), articulated 3D hand tracking at 30 Hz per hand, and speech recognition, plus hardware H.265 and Opus encoding ^[5][6][7][9]. This is what turns SLAM from an offline post-process into a live signal a prototype application can consume, and why Gen 2 can be programmed while Gen 1 could not ^[9][13].

Gen 2 also adds multi-device time alignment over a sub-GHz radio: one device broadcasts its timestamps and others compute their offset, without the round-trip handshake earlier software methods required. Meta reports a measured time-offset error below 10 microseconds over ranges of 30 meters indoors and 100 meters outdoors ^[6][7].

Availability and access

Aria has never been sold. Access runs through the Aria Research Kit, which Meta describes as including "Aria Gen 2 glasses, tooling, and support" and as "available to select partners through a rolling application process" ^[12]. The kit bundles a companion app, Project Aria Tools, a client SDK, cloud Machine Perception Services (eye gaze, hand tracking, 6DoF trajectory) and Aria Studio ^[12]. No price is published on any Meta page, and Meta has not said whether partners pay anything.

Meta's dated blog posts are the clearest signal on status. Applications for Gen 2 opened on October 19, 2025, applications for Gen 1 closed the same day, and Meta said it was "targeting a broad rollout of Aria Gen 2 devices to qualified applicants in Q2 2026" ^[10]. Alongside that it published the Gen 2 device whitepaper and the Aria Gen 2 Pilot Dataset ^[9][10].

The website is less consistent, and the contradiction sits on a single page. Retrieved on August 2, 2026, projectaria.com/glasses/ carries the banner "Aria Gen 2 applications are open. Our next generation hardware unlocks new possibilities" at the top, while its own comparison table further down the same page lists Gen 2 availability as "Coming soon" and Gen 1 as "Available" ^[13]. Gen 1 has in fact been closed to new applicants since October 2025 ^[10]. The research-kit page and the homepage carry the same banner and describe a kit that ships Gen 2 glasses ^[11][12]. Meta has published no post confirming that the Q2 2026 rollout was completed, and no such announcement surfaced in searches of Meta's blog or the XR trade press through August 2, 2026.

Privacy safeguards

Because the glasses record video, audio, and location in public and private spaces, Meta built in several safeguards described at launch ^[2]. Each device shows a prominent white recording light so that people nearby can tell when data is being collected, and it has a physical mute button that stops data capture when pressed ^[2]. Wearers are trained on where use is appropriate and where it is not: recording is barred in restrooms, prayer rooms, locker rooms, and private meetings, is only permitted in a wearer's home with the consent of everyone in the household, and requires written consent from venues such as shops and restaurants ^[2]. In public, participants must wear a lanyard and clothing identifying them as part of the study and pointing to a website about it ^[2].

Data is moved using encryption and a secure ingestion pipeline. Uploaded recordings are quarantined for three days before researchers can see them, during which a participant can still delete segments by reviewing low-resolution thumbnails ^[2]. Before researchers get access, the data is automatically processed to blur faces and vehicle license plates ^[2]. Meta later published the model behind that step, EgoBlur, as an open-source face and license-plate anonymizer, with an analysis comparing it against other systems ^[21]. Meta states that the glasses do not use facial recognition technology and that Aria data is not used to target advertising ^[2].

The initial deployment was limited in scale: roughly 100 Facebook employees and contractors, primarily in the San Francisco Bay Area and Seattle, made up the first group of wearers ^[2]. In a September 13, 2021 update to the launch post, Meta said employees and contractors would begin capturing data at a handful of public landmarks in Singapore, the first use of the device in public outside the United States ^[2].

Both hardware safeguards survive into Gen 2. The whitepaper describes "a privacy switch on the glasses that prevents the device from capturing audio, video, or other sensor data when engaged" and an "outward-facing privacy LED" on the front of the frame that signals to bystanders when the device is capturing ^[9]. Meta's Gen 2 posts do not revisit the wider policy framework, and no change to the training rules or the anonymization pipeline has been published ^[5][6][10]. The rules also travel downstream: datasets built with the glasses cite the Project Aria Research Community Guidelines as the standard their collection adhered to, and access requires signing a license ^[23]. Broader questions about bystander consent and recording indicators on head-worn cameras are covered at smart glasses and privacy.

Criticism

The Electronic Frontier Foundation published a critique on October 16, 2020, a month after launch. It acknowledged the precautions, noting that Meta "asks the owners' concerns before recording in privately owned venues such as a bar or restaurant", "avoids sensitive areas, like restrooms and protests" and "blurs peoples' faces and license plates", but argued the anonymization was insufficient: "there are still many other ways to identify individuals, from tattoos to peoples' gait, and these should be obfuscated, too" ^[25]. The EFF's broader objection was that egocentric capture is a more intimate class of data than the street-level imagery collected by mapping cars, because it records what a person looks at and how they interact rather than what a vehicle drives past ^[25].

The program has attracted less scrutiny than Meta's consumer glasses, in part because it is small, opt-in for wearers, and governed by documented rules. It has also drawn less independent testing. Because the device is never sold there are no teardowns and no third-party measurements of its sensors, so every performance figure in circulation, including battery life, dynamic range, and timing accuracy, traces back to Meta.

Academic partners and the research-tool model

From the start, Project Aria was tied to outside research. Meta began a pilot with Carnegie Mellon University's Cognitive Assistance Laboratory to build 3D maps of indoor spaces such as museums and airports, work aimed at AR-based navigation aids for people with visual impairments ^[2]. Over time the program expanded into a broader model in which Meta supplies the glasses, software development kits, and machine perception services to qualifying academic and commercial labs, which then run their own egocentric studies and publish results ^[12][16].

By November 2024, Meta had named partners working on goal-driven human interaction (University of Bristol), sound localization for hearing-aid research (University of Iowa), driver intent prediction (IIIT Hyderabad), and audio-based indoor navigation for blind and low-vision users (Carnegie Mellon) ^[16]. An early industrial collaboration with BMW explored how AR content could be kept stable inside a moving vehicle ^[16]. In its Gen 2 announcement Meta cited Georgia Tech work using the Research Kit to teach humanoid robots to assist people at home, and a collaboration with Envision, an accessibility company, which combined Gen 2's on-device SLAM with spatial audio to guide blind and low-vision users around indoor spaces ^[5][15]. Meta's whitepaper notes that Aria's real-time streaming has also been used to feed sensor data directly into robot control policies ^[9].

Open datasets

A major output of Project Aria has been a set of open datasets recorded with the glasses, distributed through projectaria.com and the open-source Project Aria Tools. They give the wider computer vision and robotics community egocentric data with high-quality ground truth that is otherwise hard to obtain ^[17][18][19].

Scroll sideways for more →

What makes them useful is that they pair raw sensor streams with the kind of ground truth (precise poses, geometry, gaze) that is expensive to obtain any other way ^[18][19]. The A2PD release also ships example outputs from Meta's own hand-object interaction, speaker diarization, and 3D bounding-box models, and from Nvidia's FoundationStereo depth estimator, which Meta ran over every recording in the dataset ^[10][14].

Relationship to Ego4D and Ego-Exo4D

Project Aria connects to Meta's wider egocentric research effort. The glasses are the capture device for the egocentric portion of Ego-Exo4D, a multimodal dataset released on December 13, 2023 that records the same activity from first-person and surrounding third-person cameras at once ^[22][23]. It was built by Meta FAIR, the Project Aria team, and roughly 15 university partners, and its first-person video comes with the Aria glasses' seven-channel audio, IMU data, and grayscale SLAM camera streams alongside the RGB feed ^[22].

The published sizes differ depending on which Meta surface you read. The dataset site gives 1,286.3 hours of combined ego and exo video from 740 participants across 13 cities and 123 scene contexts, matching the peer-reviewed paper; Meta's launch blog and the Project Aria documentation instead quote more than 800 participants and roughly 1,400 hours ^[17][22][23]. The lower numbers describe the dataset as actually released.

Ego-Exo4D followed the earlier Ego4D effort, a consortium dataset released in 2021 by FAIR and university partners, offering 3,670 hours of daily-life video from 931 camera wearers in 74 locations across nine countries ^[24]. Ego4D was not captured on Aria hardware; its wearers used a mix of head-mounted cameras. Both have become widely used resources in egocentric computer vision and embodied AI research ^[22][23].

Relationship to Meta's other eyewear

Aria is easy to confuse with Meta's consumer and prototype eyewear, and Meta itself keeps the categories separate: it calls Aria "research glasses", not AI glasses or smart glasses ^[14]. The Ray-Ban Meta line is a shipping consumer product with a camera, speakers, and an assistant, but no research sensor suite. Meta Ray-Ban Display adds a monocular waveguide display and pairs with the Meta Neural Band surface electromyography wristband. Orion is an internal AR prototype with binocular waveguides that Meta has said it does not intend to sell.

Aria sits outside all three. It has the densest sensor payload of any of them and the least user-facing function: no display, no assistant, and, until Gen 2's speakers, no output to the wearer at all. Writing for Road to VR, Scott Hayden described the trade Meta is making as putting out "substantively the skeleton of a future device" so that outside teams do the exploratory work more cheaply than Meta could by hiring them ^[14].

References

1. ^"Announcing Project Aria: A Research Project on the Future of Wearable AR." Meta Newsroom, 2020-09-16. about.fb.com/...oject-on-the-future-of-wearable-ar
2. ^"Announcing Project Aria: a research project on the future of wearable AR." Tech at Meta / Facebook Reality Labs, 2020-09-16 (updated 2021-09-13; retrieved via the Internet Archive, as the original host no longer serves the page). web.archive.org/...ct-on-the-future-of-wearable-ar
3. ^"Hardware Specifications." Project Aria Gen 1 Documentation, Meta. facebookresearch.github.io/...hardware_spec
4. ^Engel, J. et al. "Project Aria: A New Tool for Egocentric Multi-Modal AI Research." arXiv:2308.13561, 2023-08-24. arxiv.org/...2308.13561
5. ^"Introducing Aria Gen 2: Unlocking New Research in Machine Perception, Contextual AI, Robotics, and More." Meta Blog, 2025-02-27. meta.com/...earch-glasses-reality-labs-ai-robotics
6. ^"Inside Aria Gen 2: Explore the Cutting-Edge Tech Behind the Device." Meta AI Blog, 2025-06-04. ai.meta.com/...glasses-under-the-hood-reality-labs
7. ^"Hardware Specification." Project Aria Gen 2 Documentation, Meta. facebookresearch.github.io/...hardware
8. ^"Technical Specifications." Project Aria Gen 2 Documentation, Meta. facebookresearch.github.io/...technical-specs
9. ^The Project Aria Team at Meta. "Aria Gen 2: An Advanced Research Device for Egocentric AI Research" (device whitepaper, PDF). Meta, October 2025. projectaria.com/ariagen2devicepaper
10. ^"Aria Gen 2 Applications Are Now Open." Meta Blog, 2025-10-19. meta.com/...aria-gen-2-updates
11. ^"Project Aria." Meta (homepage, retrieved 2026-08-02). projectaria.com
12. ^"Project Aria Research Kit." Meta (retrieved 2026-08-02). projectaria.com/research-kit
13. ^"Aria Gen 2 Glasses." Project Aria, Meta (retrieved 2026-08-02). projectaria.com/glasses
14. ^Hayden, Scott. "Meta to Ship Project Aria Gen 2 to Researchers in 2026, Paving the Way for Future AR Glasses." Road to VR, 2025-10-29. roadtovr.com/...ving-the-way-for-future-ar-glasses
15. ^Hayden, Scott. "Meta Reveals Next Generation Aria Glasses for Research and Experimentation." Road to VR, 2025-02-27. roadtovr.com/...en-aria-smart-glasses-for-research
16. ^Hayden, Scott. "Meta Opens Project Aria to Researchers Tackling All-day AR Challenges." Road to VR, 2024-11-18. roadtovr.com/meta-project-aria-research-ar-challeges
17. ^"Open Datasets." Project Aria Gen 1 Documentation, Meta. facebookresearch.github.io/...open_datasets
18. ^Lv, Z. et al. "Aria Everyday Activities Dataset." arXiv:2402.13349, 2024-02-20. arxiv.org/...2402.13349
19. ^Pan, X. et al. "Aria Digital Twin: A New Benchmark Dataset for Egocentric 3D Machine Perception." arXiv:2306.06362, 2023-06-10. arxiv.org/...2306.06362
20. ^"Aria Gen 2 Pilot Dataset." arXiv:2510.16134, 2025-10-17. arxiv.org/...2510.16134
21. ^Raina, N. et al. "EgoBlur: Responsible Innovation in Aria." arXiv:2308.13093, 2023-08-24. arxiv.org/...2308.13093
22. ^"Introducing Ego-Exo4D: A foundational dataset for research on video learning and multimodal perception." Meta AI Blog, 2023-11-30. ai.meta.com/...ego-exo4d-video-learning-perception
23. ^"Ego-Exo4D." Ego-Exo4D dataset site (retrieved 2026-08-02). ego-exo4d-data.org
24. ^Grauman, K. et al. "Ego4D: Around the World in 3,000 Hours of Egocentric Video." arXiv:2110.07058, 2021-10-13. arxiv.org/...2110.07058
25. ^"Augmented Reality Must Have Augmented Privacy." Electronic Frontier Foundation, 2020-10-16. eff.org/...ted-reality-must-have-augmented-privacy
26. ^"Reading Recognition in the Wild." Project Aria dataset page, Meta (retrieved 2026-08-02). projectaria.com/...reading-in-the-wild
27. ^"Nymeria Dataset." Project Aria dataset page, Meta (retrieved 2026-08-02). projectaria.com/...nymeria
28. ^"Digital Twin Catalog" and "Aria Everyday Objects (AEO)." Project Aria dataset pages, Meta (retrieved 2026-08-02). projectaria.com/...dtc and projectaria.com/...aeo