Deployed VIO

The highest-volume SLAM systems in the world are not research codebases — they are the VIO stacks inside consumer XR and mobile devices. Meta Quest headsets, Apple’s ARKit (iPhone/iPad, Vision Pro), and Google’s ARCore each run visual-inertial tracking on hundreds of millions of devices. They are worth studying as case studies because their constraints differ sharply from the benchmarks that academic systems optimize for.

What a commercial XR stack looks like

The common recipe is inside-out tracking:

Public technical detail is limited (these are proprietary systems), but the broad architecture — multi-camera + IMU tightly-coupled estimation with a relocalization/mapping layer — matches the open literature this level covers.

Benchmark priorities vs product priorities

Academic benchmarkShipped device
Primary metrictrajectory RMSEperceived tracking failures per hour
Environmentcurated datasetsevery bedroom, office, train, and mirror
Computedesktop CPU/GPUshared mobile SoC, hard power budget
Calibrationdone offline, assumed correctdrifts with temperature and drops; refined online
Initializationexpert-recorded excitationwhatever motion an untrained user provides
Failure handlingsequence excluded / reportedmust degrade gracefully, recover invisibly

What deployment changes about the problem

Why study them without source code?

Three reasons. First, the constraint set explains why the open literature looks the way it does: efficiency-obsessed filters, online calibration, fast relocalization, and consistency machinery are exactly what shipping products need. Second, the public artifacts — developer APIs (anchors, plane detection, tracking-state notifications), device teardowns, and published talks/papers from these teams — are informative about architecture even when implementations stay closed. Third, many open systems in this roadmap were written by people who came from or went to these teams; the design DNA is shared.

What to take away

Treat deployed VIO as the requirements document for this level: any open system you study can be graded against the deployment checklist — What happens in the dark? How fast is relocalization? Does it estimate its own calibration? Does it degrade gracefully? How many milliwatts? — and the gaps you find are the research problems industry actually pays for.

Why it matters for SLAM

Deployed VIO is proof that the material in this level is not academic: MSCKF-style filtering, preintegration, online calibration, and observability-aware design are exactly what ships in consumer devices. Studying the deployment constraints (robustness, latency, power, calibration drift) tells you which research problems industry actually pays for — and explains design choices in open systems written by authors who came from or went to these teams.