From ab728c7e9aaf9a5e91b0a17b2820289cb4c0fa19 Mon Sep 17 00:00:00 2001
From: crosstyan <crosstyan@outlook.com>
Date: Wed, 11 Mar 2026 11:54:41 +0800
Subject: [PATCH] fix: use shortest-path rotation interpolation in playback

Switch viewer playback from Magnum Math::slerp() to Math::slerpShortestPath() when interpolating adjacent OpenSim frame orientations.

Why:

- Adjacent OpenSim quaternions can cross sign while representing nearly identical orientations.

- Non-shortest-path interpolation can create artificial long-arc spins between valid sampled poses.

- That makes playback exaggerate or invent visible bone flips that are not present in the sampled frame states.

What changed:

- Updated playback interpolation in src/ViewerApp.cpp to use shortest-path quaternion slerp.

- Added docs/motion-troubleshooting.md documenting the distinction between viewer interpolation artifacts and upstream IK discontinuities.

- Added a README pointer to the troubleshooting note.

Investigation log:

- Verified the viewer loads .mot through OpenSim state storage and renders PhysicalFrame transforms directly.

- Reproduced the target Sports2D/Pose2Sim/OpenSim clip and confirmed the .mot already contains large coordinate discontinuities and limit clamping, indicating upstream IK failure.

- Confirmed the viewer also had a separate interpolation issue due to non-shortest-path quaternion slerp.

Validation:

- Rebuilt with: cmake --build build -j

- Relaunched the viewer successfully against the problematic .osim/.mot pair after the fix.
---
 README.md                      |  4 +++
 docs/motion-troubleshooting.md | 45 ++++++++++++++++++++++++++++++++++
 src/ViewerApp.cpp              |  5 +++-
 3 files changed, 53 insertions(+), 1 deletion(-)
 create mode 100644 docs/motion-troubleshooting.md

diff --git a/README.md b/README.md
index 35ba0b7..76de816 100644
--- a/README.md
+++ b/README.md
@@ -31,3 +31,7 @@ Sample run:
   /home/crosstyan/Code/opensim-core/OpenSim/Moco/Test/walk_gait1018_subject01.osim \
   /home/crosstyan/Code/opensim-core/OpenSim/Moco/Test/walk_gait1018_state_reference.mot
 ```
+
+## Notes
+
+- Motion troubleshooting notes: [`docs/motion-troubleshooting.md`](docs/motion-troubleshooting.md)
diff --git a/docs/motion-troubleshooting.md b/docs/motion-troubleshooting.md
new file mode 100644
index 0000000..0f1543a
--- /dev/null
+++ b/docs/motion-troubleshooting.md
@@ -0,0 +1,45 @@
+# Motion Troubleshooting
+
+## Bone Flips During Playback
+
+The viewer loads the `.mot` file through OpenSim, converts it to model states,
+and renders each body using `PhysicalFrame::getTransformInGround()`. It does
+not solve IK, rebuild bones from markers, or apply any custom joint logic.
+
+Because of that, a visible flip can come from two different places:
+
+1. The upstream motion data already contains a discontinuity.
+2. Playback interpolation between two valid sampled poses introduces an
+   artificial long-arc rotation.
+
+## Viewer Behavior
+
+Playback now uses shortest-path quaternion interpolation between sampled OpenSim
+poses. This avoids false in-between spins when adjacent quaternions have
+opposite signs but represent nearly the same orientation.
+
+If a flip is still visible after this fix, the sampled OpenSim motion itself is
+already discontinuous.
+
+## Upstream IK Failures
+
+For Sports2D / Pose2Sim / OpenSim pipelines, broken IK usually shows up as one
+or more of the following:
+
+- sudden large frame-to-frame jumps in joint coordinates
+- joint values snapping to model limits
+- marker RMS staying low while anatomically implausible body orientations appear
+- left/right ambiguity or bad depth reconstruction in the source TRC
+
+Sports2D ultimately delegates IK generation to Pose2Sim, which then calls
+`opensim.InverseKinematicsTool(...)`. If the generated `_ik.mot` already
+contains large discontinuities, the viewer is only displaying that failure.
+
+## Practical Check
+
+To separate viewer issues from upstream IK issues:
+
+1. Load the `.osim` and `.mot` in the viewer.
+2. Scrub near the suspicious frame while paused.
+3. If the sampled pose itself is wrong, the IK output is broken upstream.
+4. If only the in-between motion looks wrong, interpolation is the likely cause.
diff --git a/src/ViewerApp.cpp b/src/ViewerApp.cpp
index 3d2d0b8..f4da96f 100644
--- a/src/ViewerApp.cpp
+++ b/src/ViewerApp.cpp
@@ -343,7 +343,10 @@ void ViewerApp::updateSceneAtCurrentTime() {
 
     for(std::size_t i = 0; i != _frames.size(); ++i) {
         const Vector3 translation = Math::lerp(_sceneData.frames[i].translations[sample], _sceneData.frames[i].translations[next], alpha);
-        const Quaternion rotation = Math::slerp(_sceneData.frames[i].rotations[sample], _sceneData.frames[i].rotations[next], alpha).normalized();
+        const Quaternion rotation = Math::slerpShortestPath(
+            _sceneData.frames[i].rotations[sample],
+            _sceneData.frames[i].rotations[next],
+            alpha).normalized();
         _frames[i].object->setTransformation(makeFrameMatrix(translation, rotation));
     }