CNA-GAP-042: CUBICSPLINE bone and node channels are replayed linearly between samples taken at least every 1/60 s, an approximation of the authored curve

CNA snapshot 3b43bd86  ·  Known Issues › Functional gaps  ·  source links pinned to 3b43bd86

BuildTrack samples each CUBICSPLINE interval at least every 1/60 s and AnimationPlayer lerps/slerps between the samples, so the curve between samples is approximated.

Identifier

CNA-GAP-042

Category

Functional gap

Subsystem

Models & glTF

Status

Narrowed (partially fixed; describes only what survives)

Verified against

CNA 3b43bd86 (3b43bd8687ed46523d873f9bec4b32bd8481dcb5)

Evidence basis

Source-verified: read at TARGET, not executed

Tests touching this area

Yes: see Current tests

Affected contract

AnimationPlayer and SkinnedModelEXT playback of glTF clips with CUBICSPLINE interpolation

Expected behaviour

glTF CUBICSPLINE channels define a Hermite curve between keys, which a player is expected to evaluate.

Actual behaviour at TARGET

BuildTrack in GltfImportCore.cpp resamples a bone's channels onto the union of their key times and, since a9d378c06, also samples every CUBICSPLINE interval at least every 1/60 s (four samples at the least) and adds a hold key 10 µs before each STEP key, each evaluated with the channel's own mode. SampleTrack in AnimationPlayer.cpp still interpolates linearly (Slerp for rotation) between those samples, so a CUBICSPLINE curve is piecewise-linear between samples rather than evaluated. gltf-limitations.md records the result: anim-cubicspline plays within 0.0021 units of its spline (it was 0.962 off), “a bound for that fixture, not a universal one, because authored tangents may bend a spline more sharply”. Morph-weight tracks keep their tangents and are evaluated as Hermite curves.

The dense sampling is CNA's fix for CNA-DEMOS-ISSUE-031 (STEP and CUBICSPLINE playing as LINEAR), which CNA closed on a complete checkpoint (6c8b042c4, 2026-10-24); the approximation between samples is what remains.

Source locations

• modules/content/src/GltfImport/GltfImportCore.cpp — BuildTrack: union-of-key-times resampling
• modules/graphics/src/Xna/AnimationPlayer.cpp — SampleTrack: Lerp/Slerp between keys
• docs/gltf-limitations.md — resampling row with the measured error

Evidence

Checked by reading at 3b43bd86; not executed. The error figures are CNA's own measurements on its anim-cubicspline fixture.

Re-verified at c6dec91f: a9d378c06 (CNA-DEMOS-ISSUE-031) added kStepHold and kCubicSpacing (1/60 s) sampling to BuildTrack and two linear-playback sampling tests; AnimationPlayer playback is unchanged. CNA_DEMOS_REPAIR_MATRIX.json lists ISSUE-031 as PENDING_VERIFICATION. Not executed.

Re-verified at 3b43bd86: GltfImportCore.cpp, AnimationPlayer.cpp and gltf-limitations.md are unchanged; CNA_DEMOS_REPAIR_MATRIX.json lists ISSUE-031 as FIXED_VALIDATED (accepted at a9d378c06). Not executed.

Focused reproduction

No focused reproduction is known. Nothing has been invented here; the evidence above is what exists.

Current tests

GltfAnimationSamplingTests.cpp now plays STEP and CUBICSPLINE tracks linearly and checks the held step and the spline within its tolerance; GltfAnimationRobustnessTests.cpp covers resampling. No test bounds the between-sample error of an arbitrary, sharply bent spline.

Regression test

If Hermite playback is added, a mid-sample assertion on a sharply bent spline (large authored tangents), where the 1/60 s sampling still deviates.

Blast radius

Clips authored with CUBICSPLINE and large tangents (sharp overshoot or easing within one 1/60 s interval); typical clips play within CNA's measured bound, and LINEAR clips are unaffected.

Workaround

Bake animations densely at export time.