Near-Eye Eye Tracking

Scope. Eye tracking in head-mounted devices — smart glasses, AR/VR headsets and research eyewear — where cameras, photodiodes and infrared illumination sit within centimetres of the eye. This page covers the sensor stack, optics, performance limits, companies and patents. Device-category context is on smart-glasses; display optics on augmented-vision; clinical use of the resulting gaze data on eye-tracking-biomarker-neurodegeneration.

Why it matters

Eye tracking is the enabling sensor for several functions that otherwise cost power, hardware or usability:

  • Foveated rendering — render only the fixated region at full resolution; the review reports foveated rendering tolerates roughly 50–70 ms of latency, while common head-mounted trackers measured 45–81 ms end-to-end. 1
  • Gaze interaction and accessibility — selection, scrolling and hands-free control for users with motor impairments. 1
  • Optics and comfort — dynamic eye-box steering and pupil steering reduce the trade-off between field of view and display power; vergence and IPD estimates improve stereo rendering and fit. 2
  • Presence and avatars — eye contact, blinks and eyelid aperture drive social presence in VR. 1
  • Analytics and health — attention, cognitive load, pupillometry and clinical biomarkers. 1

The counterweight is privacy: gaze reveals attention, intent and physiological state, and eye images can carry biometric identity. 1

How near-eye tracking works

Camera-based video-oculography (VOG)

The dominant architecture is PCCR (pupil-centre corneal reflection): infrared LEDs illuminate the eye; one or more cameras image the pupil and the corneal glints. The pupil centre plus glint positions give a gaze vector, usually via a model that estimates five degrees of freedom — eye position in X, Y, Z and rotation in azimuth/elevation. 1 3 Dark-pupil and bright-pupil illumination, glint detection and machine-learning gaze regression are the main variants; modern systems increasingly replace per-user calibration with learned, slippage-invariant models. 4 5

In a headset the components typically sit in a ring between the user and the display: cameras plus IR light sources around the lens, with the tracker output fused with head-pose data. 6

Camera-free, scanning and split-cadence designs

Several 2025–2026 designs remove the eye camera or reduce its duty cycle:

  • Photodiode oculography (PSOG) — multiple IR emitters and photodiodes around the frame infer eye rotation from differential reflectance of sclera, iris and pupil; no imaging sensor. 7
  • Scanning-source, camera-less tracking — a MEMS scanning light source plus photodiodes behind the lens detects reflections; Apple’s granted US12687925B2 describes this architecture. (See patents below.)
  • Split-cadence tracking — run a slow five-DoF pipeline for eye position and a fast two-DoF pipeline for rotation from glints, interpolating gaze between reference frames (Apple WO2026030005A1). 5
  • Self-mixing interferometry (SMI) — a laser senses its own returned light; the same beam-steering element can serve display and eye sensing (Apple US20260339096A1).

Illumination and optics

Illumination is the hard part in prescription-style glasses. IR light must reach the eye and return to a detector without contaminating the display or leaking visibly. Diffractive waveguides couple light in and out but typically stay below 5% efficiency end-to-end (about 0.25% through two passes); geometric/reflective waveguides reach roughly 10%; holographic mirror combiners can reflect up to 85% of NIR while transmitting visible light. 2 The 2026 PBOE review argues that polarization-selective Pancharatnam–Berry elements can approach 100% diffraction efficiency and suppress display crosstalk. 2

Metasurface cameras

Flat meta-optics replace the refractive lens stack with nanostructured surfaces. NILT’s metaEye is the first commercial camera built entirely from meta-optical elements, using two metalenses plus an integrated bandpass filter and apertures in a solid cube mounted on the sensor cover glass. 8 3

Performance metrics

MetricTypical rangeNotes
Gaze accuracy0.5–2°Device-dependent; 1.8° uncalibrated / 1.3° offset-corrected for Neon 5
Precision0.05–0.2°Reported for desktop-grade trackers; less often specified for HMDs 1
Sampling rate60–200+ Hz200 Hz common for research; 240–480 fps sensors in metaEye 5 8
Latency<10 ms best; 20–80 ms typicalPhoton-to-output; foveated rendering tolerates ~50–70 ms 1
Tracking field~90–100° × 80°Neon: 100° × 80°, covering the human FOV 5
Power15 mW sensor; headset-level higherNeon sensor system <15 mW; metaEye low-power mode ~15 mW 5 9
Weight7–13 g module; 12.9 mg cameraNeon module 7 g; metaEye 12.9 mg 5 8

These are vendor or review figures, not a single benchmark; the same device can differ by participant, glasses, makeup, contact lenses and lighting. 1 5

Companies and systems

  • Tobii — the dominant XR integrator. Its eye tracking is in PlayStation VR2, HTC Vive Pro Eye, Pimax Crystal, Pimax Dream Air and Play for Dream MR. 6
  • Pupil Labs — research-grade wearable eye tracking (Neon, below) plus NeonNet, its calibration-free ML pipeline; the company reports 10,000+ users and 1,100+ publications. 4 5
  • NILT — Danish meta-optics manufacturer (founded 2006; owned by Radiant Opto-Electronics; manufacturing in Malaysia); supplies the metaEye camera. 3
  • Apple — no announced eye-tracking product beyond Vision Pro, but a dense 2026 patent cluster (below).
  • Meta, Magic Leap, Samsung, DigiLens — named in the PBOE review as proposing waveguide- or combiner-based IR illumination for in-field eye tracking. 2
  • Varjo, Pico, Fove — HMDs with integrated trackers in the review’s comparison table (Varjo VR-1/VR-3, Pico Neo 3 Pro Eye, Fove-0). 1
  • AdHawk Microsystems — Canadian MEMS eye-tracking startup; Bloomberg reported in March 2025 that Google was in final talks to acquire it for about USD 115 million. The deal was not confirmed by either company, so this remains a report, not a completed acquisition.

Pupil Labs Neon

Neon is a wearable eye tracker built as a frame plus a clip-in module, aimed at research and product development rather than consumer displays.

ParameterNeon
PriceEUR 6,250 (academic EUR 5,515) 4
Module size / weight35 × 40 × 10 mm; 7 g 5
Eye cameras2 × IR, 192 × 192 px at 200 Hz 5
Scene camera1600 × 1200 at 30 Hz; 103° × 77° (128° diagonal) 5
Gaze output2D gaze in scene coordinates at 200 Hz; eye position, orientation, pupil diameter 5
Accuracy1.8° uncalibrated; 1.3° with offset correction 5
Tracking field100° × 80° 5
Latency<10 ms photon-to-output 5
Power<15 mW sensor system 5
Recording4 h at 200 Hz; 25+ h storage; Samsung S25 compute 5
CalibrationCalibration-free; slippage-invariant 4 5

The module can be moved between frames (including child and sports frames), and the software stack provides Pupil Cloud and a recording API. 4 Its accuracy and robustness claims come from the vendor and a linked Zenodo test report; independent replication is limited. 5

NILT metaEye

NILT launched metaEye on 28 October 2026: an ultra-compact NIR camera specified for eye tracking in smart AI glasses and AR/VR headsets. 3 10

  • Architecture — two meta-optical elements (MOEs) plus a bandpass filter and multiple apertures integrated into one solid lens cube on the sensor cover glass; no refractive lenses, no air gaps and no plastic parts. 8 3
  • Size and weight — 1.68 × 1.68 × 2.11 mm; 12.9 mg. 8
  • Sensor — 400 × 400 px, 2.2 µm pixels (OG0TC1B); 850 nm design wavelength ±20 nm; integrated 850 nm bandpass. 8
  • Frame rate — 240 fps at full resolution; 480 fps at 200 × 200. 8
  • Optics — 90° diagonal field (H 65°, V 65°); F/2.4; object distance 15–40 mm; TTL 1.7 mm; −8.3% SMIA distortion. 8
  • Interface — 1-lane MIPI D-PHY or 1-trio MIPI C-PHY. 8
  • Applications — eye tracking, iris recognition, blink analytics, VR perimetry, attention monitoring. 8
  • Claims — NILT calls it the first all-meta-optics camera and the first product with two MOEs, and says it was recognised with a SPIE Prism Award in 2025; these are company claims. 3 10

On 19 January 2026 Pupil Labs and NILT announced a pilot partnership to integrate metaEye into Pupil Labs’ next-generation eye-tracking product and NeonNet pipeline. 9 The partnership release quotes a rounded 1.6 × 1.6 × 2.0 mm and a low-power mode of about 15 mW, and says the sensor/illumination can run at up to 480 Hz. 9

Eyewear-integrated photodiode tracking: Pezzoli et al. (2026)

A Politecnico di Milano / EssilorLuxottica Smart Eyewear Lab study designed a fully frame-integrated PSOG system: four NIR emitters and four photodiodes in one half-frame, multi-frequency modulation to separate sources and reject ambient IR, and lightweight regression (lookup tables, shallow neural networks, Gaussian process regression) for azimuth/elevation. 7

On a controllable artificial eye, the best configuration (GPR on a square-modulated signal) reached an RMSE of 0.025° and a median error of 0.018°, with lookup-table baselines around 0.9–1.0° median; other signal/method combinations were worse (1–3°). 7 The authors present this as a feasibility study: human-subject validation, anatomical variability and on-device calibration remain future work, so the sub-degree result should not be read as wearable performance. 7

Apple’s 2026 eye-tracking patents

Apple published a cluster of filings covering complementary architectures. These are applications or grants, not announced products.

  • US12687925B2 — Camera-less eye tracking system (granted 21 Jul 2026; filed 12 Nov 2024; priority 5 Dec 2023). A MEMS scanning light source projects spots through the lens; a set of photodiodes behind or within the refractive/diffractive medium detects the returning reflections, so no eye camera is needed. Inventors include Gelareh Babaie. Patents
  • US20260339096A1 — Eye tracking (published 29 Oct 2026; filed 28 Jan 2026; continuation of 18/105,709). Self-mixing interferometry sensors share the display’s beam-steering element; the return signal yields angular velocity and a gaze vector. Inventors: Tong Chen, Mark T. Winkler, Edward Vail, Xibin Zhou, Ahmet Fatih Cihan, Walter Nistico. USPTO PDF
  • US20260329078A1 — Eye-tracking module in the nose bridge (published 22 Oct 2026; filed 12 Sep 2025; priority 22 Apr 2025). Camera, illuminator and motion sensor stacked on a PCB and flexible circuit inside the nosepiece, under 3 mm thick. Inventors: Nivay Anandarajah, Darshan R. Kasar, Guy Lederman, Katya Gotlib. Patents
  • US20260147215A1 — Head-mountable device with eye tracking (published 28 May 2026; filed 27 Nov 2024). Cameras and IR projectors move into the arms; inner and outer IR-reflective coatings on the lens route eye illumination and returning light, leaving the front frame passive. Inventors: Adam S. Meyer, Peter C. Tsoi. Patents
  • US20260126855A1 — Eye tracking data filtering (published 7 May 2026; filed 5 Jan 2026). Replaces pupil positions outside the lens bounding box, and gaze pixels outside the visible region, before the gaze estimate is used. Inventors: Jacob Wilson, Tobias Eble, Sabine Webel, Hariprasad Puthukkootil Rajagopal, Andreas Gapel, Ritesh Gangadhar Sholapur. Patents
  • WO2026030005A1 — Split-cadence eye tracking (published 5 Feb 2026; filed 17 Jul 2025). A low-cadence pipeline estimates five DoF; a high-cadence pipeline estimates only rotation from glints and interpolates gaze. Inventors: Lionel E. Edwin, Tom Sengelaub. Patents

Integration with smart glasses and displays

Eye tracking is coupled to the display in both directions. It enables foveated rendering and dynamic eye-box/pupil steering, which relax the resolution and uniformity requirements on waveguides. 2 In return, the display’s optics constrain the tracker: cameras and illuminators compete for the frame, the lens must pass NIR, and waveguide illumination loses most of the light. 2 The 2026 design responses split into (i) moving components to the arms or nose bridge (Apple filings), (ii) removing the camera entirely (PSOG, scanning photodiodes), and (iii) shrinking the camera with meta-optics (NILT). 8 7

For glasses specifically, the social constraints are as important as the optical ones: the front frame is the most style-sensitive part, which is why arm- and nose-bridge-mounted designs recur. 8 3

Research and clinical crossover

Neon is widely used in psychology, neuroscience and human factors, and its 200 Hz gaze/eye-state streams support fixation, blink and pupillometry analyses. 4 5 Clinical and cognitive applications — including VR perimetry (a listed metaEye use case) and eye-movement biomarkers — are covered on eye-tracking-biomarker-neurodegeneration. 8

Open questions

  • Real-world accuracy — most headline numbers come from bench or controlled tests; glasses, contact lenses, makeup, sunlight and anatomical variation degrade them. 1 5 7
  • Calibration-free generalization — learned models (NeonNet) promise no calibration, but independent, population-wide validation is thin. 4
  • Latency vs power — high-rate tracking and always-on sensing compete with battery life; low-power modes trade field or rate. 8 9
  • Privacy and biometrics — gaze and eye images are sensitive; on-device processing and data minimisation are not yet standardised. 1
  • Patent thicket — Apple, Tobii, Meta and component suppliers are all filing; freedom to operate is unclear. (See the patent entries above.)
  • Clinical translation — PSOG and camera-less designs still need human validation before they can support perimetry or biomarker claims. 7

References

Footnotes

  1. raw/papers/adhanom-2023-eye-tracking-vr-review.md ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12

  2. raw/papers/benoit-2026-pboe-eye-tracking-ar-glasses.md ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  3. raw/articles/nilt-metaeye-press-release-2026.md ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  4. raw/articles/pupil-labs-neon-overview-2026.md ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  5. raw/articles/pupil-labs-neon-specs-2026.md ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14 ↩15 ↩16 ↩17 ↩18 ↩19 ↩20 ↩21 ↩22

  6. raw/articles/tobii-xr-headsets-2026.md ↩ ↩2

  7. raw/papers/pezzoli-2026-eyewear-infrared-eye-tracking.md ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7

  8. raw/articles/nilt-metaeye-product-page-2026.md ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8 ↩9 ↩10 ↩11 ↩12 ↩13 ↩14

  9. raw/articles/pupil-labs-nilt-partnership-2026.md ↩ ↩2 ↩3 ↩4

  10. raw/articles/opticsorg-nilt-metaeye-2026.md ↩ ↩2