The Gold Standard of Virtual Reality | Keeping an Anchored Object Fixed When a Reference Space Changes

Put a small model on a table and describe its position relative to a chosen corner. Now choose the opposite corner as the origin. The position numbers change even though nobody touched the model. Coordinates describe a relationship; they are not labels permanently attached to matter.

A virtual scene has a similar distinction, with additional software choices. An object may be attached to a room-like reference, to a controller or to the viewer. Before discussing a jump, state which relationship the application intends to preserve. A label intended to follow a hand should behave differently from a sculpture intended to remain in the gallery.

For this guide, the object is meant to remain fixed in one represented gallery. The viewer is also fixed during the instantaneous bookkeeping comparison. We change the coordinates used to describe both. Later chapters add genuine movement and explain why that is a different operation.

A useful opening record has four entries: the represented object, its intended attachment, the frame in which its pose is stored and the frame used for the view calculation. If any entry is missing, a plausible subtraction may answer the wrong spatial question.

Contents · Next chapter

Call the earlier coordinate frame O and the replacement frame N. Let x_O and x_N describe the same point. Define the supplied transformation by x_O = R x_N + t. In words, R and t convert coordinates from the new frame into the old frame.

Here R is a rotation matrix, while t is the position of the new origin expressed in old coordinates. Every position in the calculation uses metres. This direction statement matters more than the letter chosen for the matrix. Another source may define its transformation in the reverse direction.

A rotation preserves lengths and angles. Its transpose Rᵀ is its inverse. Therefore the reverse conversion is x_N = Rᵀ(x_O − t). The subtraction happens before the inverse rotation because t was supplied in old coordinates.

Read that equation aloud before calculating: subtract the new origin’s old position, then express the remaining vector along the new axes. Writing Rᵀx_O − t would generally be wrong. It subtracts an old-frame translation after the coordinates have already been rotated. The units still look correct, so dimensional checking alone would miss the error.

Contents · Previous chapter · Next chapter

Begin on a line. The new origin lies one metre to the positive side of the old origin, and the axes have the same direction. The conversion from new to old is x_O = x_N + 1. An object at old coordinate 3 therefore has new coordinate 2.

Place the viewer at old coordinate 0.5. Its new coordinate is −0.5. The object’s relative displacement is 3 − 0.5 = 2.5 metres before the change and 2 − (−0.5) = 2.5 metres afterwards. Changing both coordinates consistently preserves the relationship.

Now deliberately keep the object’s old number 3 while interpreting it in the new frame. It represents old coordinate 4. The object has acquired an unintended one-metre shift. That is an application bookkeeping error in this example, not evidence that the real object or the viewer moved.

If the learner subtracts one from the object but adds one to the viewer, return to the direction statement. Both positions describe points in the same new frame and use the same conversion. A change of origin does not select a different rule for whichever quantity happens to be called the camera.

Contents · Previous chapter · Next chapter

Use a planar rotation embedded in three dimensions. Let R turn a new-frame vector through positive ninety degrees about the old positive z-axis. Thus R maps (a,b,c) to (−b,a,c). Set t = (1,2,0) metres.

The new origin is therefore at old position (1,2,0). The new positive x-axis points along old positive y, and new positive y points along old negative x. These statements define the picture independently of any software’s naming convention.

Take an object point at x_O = (3,4,1). First subtract t to obtain (2,2,1). Applying Rᵀ gives x_N = (2,−2,1). Check by converting forward: R(2,−2,1) + (1,2,0) = (2,2,1) + (1,2,0) = (3,4,1).

The recovered old point is an independent check of the conversion. It also reveals why subtracting origins is not enough once axes rotate. The pair (2,2) is an old-axis displacement from the new origin; it is not yet the pair of coordinates along the new axes.

Contents · Previous chapter · Next chapter

Place the viewer at h_O = (2,3,1) metres. The viewer’s displacement from the new origin is (1,1,1), so h_N = Rᵀ(1,1,1) = (1,−1,1).

Before changing frames, the displacement from viewer to object is x_O − h_O = (1,1,0). Afterwards, the new-frame displacement is x_N − h_N = (1,−1,0). These two coordinate triples differ because their axes differ. Their lengths are both √2 metres.

It would be incorrect to announce that the relative direction changed merely because the triples changed. A vector is a geometric quantity with components relative to a basis. To compare components directly, express both in the same basis.

At this stage we have preserved displacement geometry, but have not yet calculated what the viewer sees. The viewer’s own orientation supplies the final coordinate conversion into the view.

Contents · Previous chapter · Next chapter

Let Q_O map vectors from the viewer’s axes into old-frame axes. Choose Q_O = R in the worked example. The viewer is oriented through the same ninety-degree rotation used above.

The object point in viewer coordinates is q = Q_Oᵀ(x_O − h_O). Substituting gives Rᵀ(1,1,0) = (1,−1,0). This is a view-space displacement, not a location in the gallery frame.

After the reference change, the viewer orientation is Q_N = RᵀQ_O. Because Q_O = R, Q_N is the identity rotation. The viewer’s axes now align with the new frame. The new view calculation is Q_Nᵀ(x_N − h_N) = (1,−1,0), exactly as before.

A position-only correction would leave Q_O attached to a position already expressed in N. Applying that unchanged rotation to the new displacement produces (−1,−1,0), which is the wrong view-space result. The apparent error comes from mixing frame meanings, even though every array contains three tidy numbers and every matrix still describes a valid rotation in isolation.

Contents · Previous chapter · Next chapter

For the general change, x_N = Rᵀ(x_O − t), h_N = Rᵀ(h_O − t) and Q_N = RᵀQ_O. Subtracting the first two positions cancels the common translation: x_N − h_N = Rᵀ(x_O − h_O).

Now Q_Nᵀ = Q_OᵀR. The new view displacement is therefore Q_OᵀR Rᵀ(x_O − h_O). Since R Rᵀ is the identity, this reduces to Q_Oᵀ(x_O − h_O), the earlier view displacement.

This proof explains what must remain consistent. The same rigid transformation applies to object position, viewer position and viewer orientation. The cancellation is not a fortunate feature of the chosen ninety-degree example.

It also identifies the boundary of the claim. We have shown an algebraic invariance within a supplied coordinate model. We have not shown that a tracker estimated R or t correctly, that timestamps match, or that a display presented the expected image. Those questions need their own observations. A correct identity is powerful because its assumptions are explicit, rather than because it replaces every test in the pipeline.

Contents · Previous chapter · Next chapter

A sculpture has orientation as well as a representative point. Let S_O map the object’s local axes into the old gallery axes. A vertex with object-local coordinates u has old position S_Ou + x_O.

After the reference change, use S_N = RᵀS_O and x_N = Rᵀ(x_O − t). The same vertex then has new position S_Nu + x_N. Converting this forward with R and t recovers S_Ou + x_O exactly.

For an object initially aligned with O, S_O is identity. In the ninety-degree example, its new orientation is Rᵀ. Leaving its orientation as identity while changing only its centre keeps one point fixed but rotates the rest of the sculpture relative to the represented gallery.

This is why checking a single anchor point is insufficient for a full pose update. Inspect at least enough independent geometry to test orientation as well. A model can look stable at its origin while a long feature points in the wrong direction. The check should match what the stored quantity claims to represent: a point, a direction or a complete rigid pose.

Contents · Previous chapter · Next chapter

A homogeneous transform packages a rotation and translation into one matrix. Write T for the transform from N to O, C_O for the object’s pose in O and H_O for the viewer’s pose in O. Each pose maps its local coordinates into the named gallery frame.

The consistent new poses are C_N = T⁻¹C_O and H_N = T⁻¹H_O. The viewer-relative object pose becomes H_N⁻¹C_N = (T⁻¹H_O)⁻¹T⁻¹C_O. Reversing the product when taking the inverse gives H_O⁻¹T T⁻¹C_O = H_O⁻¹C_O.

The order is essential. Matrix multiplication generally does not commute. Replacing T⁻¹C_O with C_OT⁻¹ usually applies a different operation, associated with a different frame relationship.

The matrix notation earns its place when it reduces separate position and orientation bookkeeping. It should not hide the meaning of the arrows. Keep a small annotation beside each transform: what frame it accepts and what frame it returns. A multiplication is meaningful when the output frame of one factor matches the input frame of the next, just as units must match in an ordinary calculation.

Contents · Previous chapter · Next chapter

Suppose the application deliberately moves the sculpture one metre within the represented gallery. That changes its relationship to the viewer unless the viewer moves in a corresponding way. The view should generally change. Preserving the old view would defeat the intended object action.

A reference change instead alters the description used for the same intended placement. The consistent coordinate update is designed to preserve that placement. The two operations may use similar mathematical objects, but their purposes differ.

For a classroom exercise, give two event cards. One says that a sculpture was repositioned within the gallery. The other says that the gallery’s coordinate origin changed while the sculpture was meant to stay put. Ask for the expected viewer-relative result before calculating anything.

This protects against a common implementation habit: adding a compensating transform whenever a jump appears. Compensation is justified only after identifying the relationship meant to remain fixed. An intentionally viewer-attached menu, a controller-held object and a room-anchored sculpture need different scene policies. No universal instruction to keep everything visually still can replace those definitions.

Contents · Previous chapter · Next chapter

An application may deliberately choose an offset from a runtime’s natural reference origin. That offset is another transform with its own direction and purpose. It should not be silently merged into a reported reference-origin change without checking the API’s definition.

The OpenXR reference-space-change event documentation describes the new natural origin relative to the previous natural frame. It also distinguishes that change from origin offsets supplied by an application. The distinction matters because compensating twice can create a new displacement rather than remove one.

Our paper model does not prescribe the transform order inside a particular engine. It provides a way to check it. Name the natural frame, the application’s chosen scene frame and the object’s local frame separately. Then compose only transforms whose frame meanings match.

A useful diagnostic is to set the application offset to identity in a controlled mathematical test, verify the natural-frame change, and then restore one explicitly defined offset. This is a calculation sequence, not a request to change a live deployment. The point is to expose which transformation describes the reference change and which describes an intentional placement policy.

Contents · Previous chapter · Next chapter

A reference change has a boundary in time. Combining an object pose from before that boundary with a viewer pose already expressed after it mixes coordinate systems even if both records have valid timestamps.

In an abstract ledger, attach a frame version to each pose: O before the declared transition and N afterwards. A calculation can either convert the earlier pose into N or convert the later pose back into O when the supplied relationships allow it. Merely sorting the timestamps does not perform the conversion.

OpenXR’s event includes a change time, and the documentation defines when location calls respect the replacement reference. An actual application must follow its runtime and engine semantics. Our algebra does not authorise inventing a transition instant from an observed visual jump.

The separate guide on Pose Sample Age and Rendering Time develops information age and presentation timing. Here the question is which spatial frame a record belongs to. A very fresh pose in the wrong frame is still wrong, while an older correctly transformed pose has a separate timing limitation.

Contents · Previous chapter · Next chapter

The algebra requires a known R and t. If the system cannot determine how the new natural origin relates to the previous one, there is no justified numerical compensation to calculate from the event alone.

The OpenXR documentation includes a validity field for the reported relationship. When the relationship is unavailable, its position and orientation are not defined by that event. Treating missing information as an identity transform would add an unsupported claim that the origins and axes stayed aligned.

For a paper exercise, replace the transformation card with the words unknown relationship. Ask which earlier conclusions survive. Object and viewer coordinates remain numbers within their own stated frames, but cross-frame comparison cannot be completed without additional information or an explicit policy that changes the scene’s meaning.

This is a useful stopping rule. A method is not incomplete because it refuses to manufacture an input. The next step may be to obtain a supported relationship, recreate a represented placement under a declared new convention or report that continuity cannot be established. The appropriate product behaviour depends on the application; mathematical certainty should not be used to conceal that design decision.

Contents · Previous chapter · Next chapter

Let T₁ map N into O using the ninety-degree rotation and translation (1,2,0). Now let T₂ map a third frame M into N with a negative ninety-degree rotation and translation (2,0,0).

To convert an M-coordinate into O, apply T₂ first and T₁ second. The combined rotation is identity. The combined translation is (1,2,0) plus R(2,0,0), which equals (1,4,0). Thus x_O = x_M + (1,4,0).

Reversing the composition produces a different translation: (2,0,0) plus the negative ninety-degree rotation of (1,2,0), giving (4,−1,0). Both products are mathematically valid rigid transforms, but they connect different imagined sequences.

For an old point (6,5,0), the correct M-coordinate is therefore (5,1,0). Convert it forward through both stages to check the result. This small example makes order visible without needing a complicated matrix library. A reliable transform chain is an ordered explanation of frame relationships, not an unordered collection of offsets waiting to be added.

Contents · Previous chapter · Next chapter

A supplied matrix should satisfy RᵀR = I and determinant +1 to represent a proper three-dimensional rotation in this model. Orthogonality checks length preservation; the determinant condition excludes a reflection.

These conditions address geometry, not provenance. A numerically perfect rotation can still be the wrong rotation for the actual reference change. Conversely, small numerical departures may arise from finite precision and require an explicitly chosen tolerance rather than an exact floating-point equality.

Do not silently normalise a suspect matrix and then claim the original event was correct. First identify whether the problem is numeric representation, a convention mismatch, a transcription error or a genuinely different transformation. Preserve the original record and state any correction.

A scale change is another distinct operation. Converting metres to millimetres changes component magnitudes according to units; stretching a scene by a factor of two changes represented distances. Both may involve multiplication by one thousand or two in code, but neither is an ordinary rigid reorientation. A frame check should name the length unit and reject an unexplained scale factor before interpreting a changed distance as physical movement.

Contents · Previous chapter · Next chapter

Pairwise distances between fixed scene points remain the same under a common rigid reference change. Angles between attached directions also remain the same. The object’s pose relative to the viewer remains the same when both poses are transformed consistently for the instantaneous comparison.

These invariants support a layered check. If pairwise object distances change, inspect scaling or inconsistent position conversion. If object distances stay fixed but the view changes, inspect the viewer pose, orientation, frame version and composition order. The pattern narrows the next investigation without proving a unique cause.

One preserved distance is a weak check. Two different configurations can share a distance. Similarly, a single object at a special symmetric location may conceal an incorrect rotation. Use deliberately asymmetric test points and orientations so that the error cannot hide behind the geometry.

Avoid claiming that passing these algebraic tests certifies a headset or tracker. A common erroneous global placement can preserve internal distances while the whole virtual gallery is misplaced relative to the physical room. Consistency of a model and accuracy of its registration are different properties, requiring different evidence.

Contents · Previous chapter · Next chapter

In the worked model, the relationship between frames is supplied exactly. Real tracking estimates can change as new evidence becomes available. A correction to the estimated scene can therefore combine coordinate bookkeeping with a changed belief about physical placement.

The algebra can describe how poses transform under a declared relationship. It cannot tell whether that relationship is an accurate account of the room. To assess registration accuracy, use an appropriate independent reference and a defined measurement question. A stable rendered object alone does not establish the answer.

For learners, separate three statements. The coordinate conversion is internally consistent. The application preserved its intended represented relationship. The represented relationship matches the physical world accurately. Each statement adds a requirement beyond the preceding one.

This distinction is useful when reviewing a demonstration. A developer may correctly show that an object does not jump under a simulated origin change. That is valuable verification of a transformation pathway. It is not a measurement of long-term physical tracking error, nor evidence that a particular environment or participant arrangement is suitable. Report the achievement precisely so that the remaining checks stay visible.

Contents · Previous chapter · Next chapter

Instructions such as move left or turn clockwise depend on a viewpoint and a reference. A coordinate rebase should not quietly change the meaning of those words for the user. Language, object attachment and interaction feedback need to remain coherent.

A viewer-relative arrow can keep pointing to the same side of the viewer while a world-relative arrow changes its displayed direction as the viewer turns. Both behaviours can be correct. The intended relationship should be documented before translations or visual adjustments are evaluated.

Localising AR, VR and Spatial Interfaces develops those language and access questions. Our coordinate derivation supplies a narrower mathematical check, rather than a replacement for testing understandable interaction.

A useful paper task gives the learner a transformed scene and an instruction containing an ambiguous direction. Ask them to rewrite it with an explicit reference: relative to the participant, relative to the object or along a named scene axis. This connects geometry with communication. Correct coordinates are more useful when the explanation also tells a person which relationship they are being asked to follow.

Contents · Previous chapter · Next chapter

Use a new planar example. The supplied map from N into O rotates through 180 degrees and translates by (−2,1) metres. An object is at old position (4,−3), and the viewer is at (1,−1). The old viewer orientation is identity.

Predict before computing. A half-turn reverses both horizontal coordinate directions, and the translation must be removed before that reversal. The old viewer-to-object displacement is (3,−2).

The object in N is the half-turn of (4,−3) − (−2,1), giving (−6,4). The viewer in N is the half-turn of (1,−1) − (−2,1), giving (−3,2). The new viewer orientation is the half-turn, not identity.

Their new-frame displacement is (−3,2). Applying the new viewer orientation’s inverse gives (3,−2), matching the earlier view displacement. Verify the positions by the forward map as a second check. The case changes the angle, translation, positions and viewer orientation; retaining only a memorised subtract-one rule cannot solve it. The stable principle is consistent transformation of every quantity used in the same relationship.

Contents · Previous chapter · Next chapter

Suppose two records describe the same fixed object on opposite sides of the reference change. Its old coordinates are (3,4,1), while its new coordinates are (2,−2,1). Averaging those triples component by component gives (2.5,1,1). That mixture does not describe a justified intermediate position because its inputs belong to different frames.

Convert both records into one common frame first. In O, both are (3,4,1), so any ordinary weighted interpolation between them remains that point. In N, both are (2,−2,1), giving the matching result there. A stationary object should not acquire a fictional excursion because the coordinate labels changed between records.

For positions expressed in one common frame, a fixed affine map commutes with linear interpolation when the weights sum to one. If x(s) = (1−s)x₀ + sx₁, then Rx(s)+t equals (1−s)(Rx₀+t)+s(Rx₁+t). The translation is included once because the weights add to one.

That identity checks coordinate consistency, not the represented object’s actual motion between samples. Interpolation still supplies an assumed path. Orientation also needs an appropriate rotation treatment; arbitrary component averaging is not justified by the position equation. The practical order is to identify the frames, reconcile them and only then apply a stated interpolation model to comparable quantities.

Contents · Previous chapter · Next chapter

A position point and a free direction vector do not respond identically to a changed origin. The point uses x_O = Rx_N+t. A direction uses v_O = Rv_N, with no translation. Moving an origin does not rotate or stretch an arrow representing a displacement.

In the worked example, the viewer-to-object vector is (1,1,0) in O and (1,−1,0) in N. Applying R to the latter recovers the former. Adding t would produce (2,3,0), incorrectly treating a displacement as a position.

This distinction also explains why subtracting two transformed positions cancels their common translation. It is the relationship between the points that becomes a vector. In homogeneous coordinates, the extra coordinate is one for a point and zero for a direction; that convention encodes whether the translation participates.

For a useful check, translate the origin while keeping axes aligned and ask whether a surface’s stated direction should change. It should not in this rigid model. Then rotate the axes and express that same direction with the inverse rotation. Attach the frame label to the result rather than announcing that the surface physically turned. This small exercise catches a category error that can survive otherwise correct matrix arithmetic.

Contents · Previous chapter · Next chapter

The first explanation says the object did not move because its stored numbers did not change. That confuses coordinates with represented placement. Under a changed origin or basis, unchanged numbers can refer to a different point.

The second says the object stayed fixed because its centre was corrected. That may be true for one point while the object’s orientation or the viewer’s orientation remains inconsistent. Check a vertex away from the centre and compute the viewer-relative pose.

The third says a successful invariant check proves physical tracking accuracy. It establishes only the tested consistency property. A common registration error can preserve every internal distance. Ask what independent physical information supports the accuracy claim.

For each error, repair the first missing definition rather than adding another unexplained matrix. Name the fixed relationship, state the direction of the map, identify the frame of each input and convert a deliberately asymmetric case forward and backward. This sequence makes debugging an explanation exercise. It also helps a learner recognise when the evidence is insufficient, instead of interpreting every unresolved scene change as a failure of their arithmetic.

Contents · Previous chapter · Next chapter

For repair, stay on one line. Use two origins, one object and one viewer. The learner should explain why both coordinates change while their difference stays constant. Ask for a sketch and a sentence before asking for another numerical answer.

For stabilisation, introduce a ninety-degree planar rotation. Check positions first, then viewer orientation and a non-central object vertex. Require a forward conversion back to the original data. This separates a genuinely understood inverse from a formula copied in the right order once.

For extension, compose two reference changes, attach frame versions to a short event ledger and compare a world-anchored object with a viewer-attached panel. The learner should identify which invariant belongs to each policy. An intentional object action should change the appropriate relationship rather than be cancelled automatically.

The final product can be a small transformation notebook: input definitions, a labelled sketch, the stated mapping direction, intermediate coordinates, independent checks and one limit. No headset is required. What matters is the ability to transfer the reasoning to a new frame convention while preserving the meaning of the quantities.

A final check can use two object vertices and the viewer’s forward direction. Convert all three into the common frame, then reconstruct the same viewer-relative geometry. Deliberately reverse one transform in a copy of the calculation. Explain which check exposes the error and why a single centre-point check could have missed it.

Contents · Previous chapter · Next chapter

Do unchanged coordinates prove an object stayed fixed? Only within an unchanged frame and attachment convention. A new frame can give the same numbers a different meaning.

Should every reference change be visually cancelled? The application must first define which represented relationships it intends to preserve. A deliberate scene action and a reference redefinition are different events.

Why transform orientation? Position locates an origin; orientation locates its axes. An object or viewer pose needs both for a consistent view calculation.

Can the reported transform always be used? Its meaning, direction, validity and relevant time must match the calculation. Unknown information cannot be replaced by identity without an explicit new assumption.

Does this prove the tracker is accurate? No. It checks a declared rigid-coordinate model and a particular consistency property. Physical registration accuracy needs additional evidence.

What is the best independent arithmetic check? Convert the new coordinates forward into the old frame, then verify the unchanged viewer-relative result using transformed orientation as well as position.

Contents · Previous chapter · Next chapter

Check the interface definition

Khronos: OpenXR Reference Space Change Pending defines the event’s natural-origin relationship, validity and change-time semantics. The page was checked on 23 October 2026. Our coordinates and invariance proofs are original teaching examples, not a device implementation or a claim that every runtime exposes identical higher-level behaviour.

Follow the relevant spatial question

Return to Virtual Reality for representation, interaction and learning. Vectors supports component reasoning, while Pose Sample Age and Rendering Time keeps the timing question distinct.

Preserve the participant's task

Human–Computer Interaction examines understandable feedback. Universal Design considers participation beyond one input method. Continue through How X Works and Learning That Transfers into Life. Explain the unchanged relationship in the changed example without referring to its original numbers; that is the useful transfer check.

Contents · Previous chapter · Return to Virtual Reality