Repository navigation
3.49.0
Creature animation and camera performance capture.
Creature motion — template clips now retarget correctly (#1085)
Animal animations played with the legs "moving in weird ways" on rigs whose own native clips looked fine. The cause was a chain of retarget bugs; a self-retarget (a clip applied back onto its own rig, which should be near-perfect) measured 39.9° of world-orientation error. It is now 3.7°.
| action | before | after |
|---|---|---|
| walk | 39.9° | 3.7° |
| idle | ~40° | 1.0° |
| jump | ~40° | 4.8° |
| run | ~40° | 10.7° |
| death | ~40° | 12.5° |
| dragon flying (winged biped) | — | 4.3° |
Library-wide: median 5.8°, nothing above 20°.
What was wrong
- Keys were solved against the bind parent, and against the previous frame. These rigs carry more limb segments than the canonical skeleton, so one bone per limb is unmapped and frozen; a key solved against the bind parent never reaches its intended world orientation. The retarget now runs a per-frame pass that poses the skeleton from the keys solved so far and solves against the live parent.
- The delta was conjugated in the parent's frame. Ogre post-multiplies a key onto the bind pose, so the correct frame is the bone's own — the parent frame left a near-constant ~30° bias under every clip.
- Root translation was discarded entirely. Rotation alone cannot express a jump (the body rises), a death (it topples) or a lunge: the body pivoted about a pinned root and folded through its own legs. Translation is now carried, scale-normalised by hip height so a horse's leap does not fling a pug, and applied to the skeleton's top-level bone.
- Half of some clips were dead.
Horse|Walkanimates for 50.6% of its 3.33 s and holds the rest — the "walk stops halfway and the legs stick". Clips are trimmed against the final pose; every walk clip is now ≥97% live. - A bare
FrontLeg.Lwas mapped to the lower leg, where it lost its role to the explicitly-named bone and received no animation at all, so the limb swung from the wrong joint.
Auto-rigging
The generate → auto-rig → animate path previously could not play any clip: an auto-rigged creature was refused outright because AutoRig names its spine SpineFront/Mid/Back while the gate required a root and a chest. Also fixed:
- the quadruped template has three-segment legs, matching the canonical skeleton;
- facing is detected from the mesh rather than assumed — a generated mesh facing −Z was rigged back-to-front and played every clip in reverse;
- left and right are preserved when the facing is reversed (a 180° turn mirrors both in-plane axes).
Known limitation
An auto-rigged creature still breaks on fast gaits. The retarget drives canonical roles, so a shorter chain absorbs the same rotation in fewer joints — measured on Cow|Run, an authored 4-bone foreleg spreads 128° with only 12° at the upper leg, while a 3-bone chain concentrates 146° into one shoulder, folding the foreleg under the chest. Walk and idle are clean; run and jump are not. Authored rigs are unaffected. A real fix needs a fourth canonical leg role, which would invalidate the existing clip library.
Tooling
scripts/creature-retarget-error.py measures world-orientation error against a rig's own native clip — a self-retarget must score ~0°. It exists because three cheaper checks each passed a visibly broken clip: per-joint angle range matched the source to 0.1° while the limb was placed completely wrong; frame-to-frame liveness reported a 50%-dead clip as 98.9% live; and sprite sheets at 5–10 frames alias badly on a walk cycle. docs/CREATURE_MOTION.md records all of it, including the metric's own blind spot (it measures orientation only, so a clip with missing root travel scores well while a jump stays grounded).
Performance capture — full-body retarget and root motion (#1084)
Camera body capture follows anatomical poses more closely, including seated positions and knee bends.
- Retargets from filtered landmark geometry using anatomical frames, rig bind axes and transported knee-bend planes; occluded joints hold their last local pose.
- Camera heading is calibrated without erasing initial body lean, preserving full turns; hip translation is estimated from image position and torso scale.
- Body-pose processing is shared across live preview, video/CLI and MCP capture, so the three cannot drift.
- Estimated body movement is now included by default — pass
--in-placeto keep translation locked. - Combined face/body recordings bake compensated head orientation into the body clip.
- Sample timestamps and neutral calibration are preserved in recorded clips; skeleton culling bounds account for root motion.
Validated by replaying two videos on a humanoid rig: 1,715 and 676 reliable leg-segment checks passed with a maximum direction error of 0.04°. See docs/MOCAP_RETARGETING_VALIDATION.md, which also documents the remaining monocular depth and foot-contact limits.
Also
- macOS 27: the export file dialog no longer crashes the app (an uncaught AppKit exception during view layout).