Merging several local frames into a shared frame is the premise of multi-user persistence

What it is

Two people face the same part on a table. Each headset’s “here” is a local frame grown from its own boot origin. Until those frames are merged into one shared frame (coordinate frame alignment / map merging), persistent content pins to a different physical point for each of them: the screw you indicate sits half a desktop away for the other person.

People in the room have to colocate first; only then can an overnight cloud anchor land on the same already-aligned table. The merge is a premise, not an option.

Why it happens

Each device’s tracking starts at its session origin, whose place in the room depends on which way it faced and where it stood at boot. An unknown rigid transform sits between the two frames. Merging estimates that transform: from environment features both see, from a marker both scan, or from one device handing its map to the other for an ICP-style registration. A wrong estimate puts the same bias on every later shared object; people patch with language (“not that one, further left”) and collaboration drops from spatial to spoken correction.

The merge has to finish before content is written into shared persistence. Pinning first and aligning later is trying to reconcile two already-frozen coordinates; the conflict will not autoscale. Once the shared frame exists, late joiners localize into it. They do not open a new origin.

Studying it

Colocation accuracy: two people enter the same room in sequence or together, merge, then both point at a real fiducial. Measure each virtual cursor against the mark, and the two cursors against each other.

Operations: merge method (cloud map, QR, host share), entry order, feature richness. Scores: relative translation/rotation, rate of agreeing on the same object, time to merge, whether the session still proceeds on a “looks together” after failure.

Single-user revisit cannot see merge error. You need at least two independent origins.

Where it stops holding

Solo, single-device persistence does not need a merge; the onboard origin is enough. Remote collaboration that does not share a physical room merges a virtual-scene origin, not this room’s walls. Feature-starved or symmetric rooms (two identical walls) can yield a mirrored solution that looks aligned and has left/right flipped. Someone who loses tracking mid-session and relocalizes without rejoining the shared frame comes back in a drifted private frame. Permission models govern who may edit; they do not say whether the coordinates are the same copy. Do not collapse the two into one setting.

Applying it

• Make merge the first step of a multi-user session. Do not let either side place persistent content before it succeeds. Use an explicit “not aligned yet” state rather than letting both pin.
• Late joiners may only localize into the existing shared frame. Forbid a second map that “also looks like the world origin” from being written into the same venue.
• On merge failure, give an actionable recovery (look at this marker together, walk another loop). Do not open on the last guessed transform.
• How to check: after a merge, both people put a cursor on one screw on the table. Both cursors should land on that screw. A stable few-centimetre split means residual is still on the task path; one person hits the screw and the other hits empty table, and the shared frame was never built.

Related

• Same group：N5.09.1 Cloud anchors let different devices at different times see the same content at the same place · N5.09.2 Persistent content must be periodically re-matched to the environment to stop long-term drift · N5.09.3 Anchor lifetime needs an explicit policy, not a permanent-environment assumption · N5.09.5 After a structural change to the environment, old anchors should be invalidated, not silently kept
• Nearby：N5.04 Spatial Anchors · N5.05 Shared Space · N5.10 Shared Space and Multi-user Collaboration
• Search terms：coordinate frame alignment · map merging · colocation