cd ~/
Smart GlassesRay-Ban MetaVPSWearablesMeta

VPS for Ray-Ban Meta AI Glasses

Bringing visual positioning to consumer smart glasses - extreme power constraints and new use cases.

Evyatar Bluzer
3 min read

Ray-Ban Meta AI glasses are launching with Meta AI built in, and the first generation ships without VPS. The integration work we are doing now, the glasses exploration on this year's list, targets the generation after, because localization on a 40g wearable is a different engineering problem from anything we shipped on Quest.

Glasses vs Headset Constraints

ConstraintQuest 3Ray-Ban Meta
Weight500g40g
Battery3-4 hrs4+ hrs always-on
ComputeHigh-end mobileUltra-low power
CamerasMulti-camera rigSmall forward camera
Use caseSessionsAlways-on

Add the rows up and the problem gets orders of magnitude harder.

Glasses VPS Architecture

Quest's VPS architecture cannot run on glasses, so this is a complete rethink. Queries go out sparsely, about once per minute rather than continuously. Processing moves cloud-heavy, with minimal work on device. The focus turns semantic: the question becomes "Where am I?" rather than "Exactly where am I?". And localization goes multi-modal, combining GPS, WiFi, and visual. Different product, different architecture.

Privacy Amplified

Glasses worn in public continuously raise the privacy stakes. Bystanders cannot tell when the cameras are active, the social norms around glasses differ from those around headsets, and always-on operation means more potential capture. The requirements get more stringent to match: extremely limited data retention, clear indicators during any capture, and location computed only when the user explicitly requests it.

What is VPS on glasses actually for?

The use cases look different from Quest's. Contextual AI is the big one: you ask "What building is this?", VPS identifies the location, and the AI provides the information. Navigation becomes step-by-step walking directions delivered as audio. Memory features like "Remember I parked here" get visual confirmation. And sharing your location comes with context: "I'm at this café". Less about AR overlay, more about AI context.

Technical Approach

The plan rests on ultra-efficient features - a new network architecture targeting 5mW inference - plus compressed queries that send minimal data up for cloud matching, predictive localization that guesses the location before the query completes, and graceful degradation, because on glasses a GPS fallback is acceptable.

One glasses localizationSequence diagram with three lifelines, wearer, glasses and cloud, four arrows and two notes; the arrow carrying the compressed query from glasses to cloud is highlighted. WearerRay-Ban Meta glasses (40 g)Cloud: VPS and Meta AI asks what building this is compressed query;about once per minute, only on request location plus AI context answer as audio;GPS fallback if the query failsCapture indicator on;ultra-efficient features,5 mW inference target;query compressed to minimal dataMatch against GPS, WiFi and visual;predictive guess before the query completes;extremely limited retention
One localization on the glasses: the wearer asks, the glasses compute features at a 5 mW target and send a compressed query about once per minute, the cloud matches against GPS, WiFi and visual, and the answer comes back as audio.

Timeline

Ray-Ban Meta launches in 2023 without VPS; the next generation, 2024 or later, is the target for integration. Two years ago I wrote this same paragraph about Ray-Ban Stories: V1 without VPS, the following generation with it. This is the following generation, and it ships without VPS too. The once-a-minute, sub-10mW architecture from that post survived intact; the timeline did not. Meanwhile this generation is our field lab. It gives us real sensor characteristics, consented training data, refinement cycles for the ultra-low-power algorithms, and glasses-specific maps. When the next pair ships, the goal is for spatial awareness to be ready and waiting for it.

Comments