Bone inventory while walking

Which bones the library moves today, what moves them and how much, compared with what those bones do in a person. It is the starting point for making the gait move all the bones that deform the body.

node scripts/run-tests.mjs zzz && node .test-build/tests/tools/boneInventory.js
# another model: PH_MODEL=modelos/human.glb PH_MAPPING=rigify (in PowerShell, with $env:PH_MODEL=…)

How it is measured

• The character walks in a straight line at 1.2 m/s for 8 s. For each bone, it measures how much it oscillates around its average pose, not its rest pose (each skeleton has its own: T-pose, A-pose…).
• The rotation is expressed in the character's axes: flexion (lateral axis), turn (vertical axis) and tilt (forward axis). Range = 95th percentile − 5th percentile, in degrees. The pelvis is measured relative to the heading.
• People, measured the same way:
  • CMU: the four normal walking takes of the naturalness test (subjects 35 and 16, ~1.2 m/s). It is the reference: its license allows using it for tuning.
  • SFU: subject 0018 in their straight stretches (1.1–1.4 m/s). Local only (naturalness): it confirms or breaks ties, but no value comes from SFU alone. It is a single person, with a much more mobile pelvis and head than in CMU.
• Limits of the mocap:
  • The SFU knee rotates about a single axis: its "tilt" is the knee axis rotated with the thigh.
  • The CMU clavicles don't move in these takes (constant channels).
  • In both, the toe joint is poorly measured with markers.
  • The CMU skeleton has its legs 20° apart at rest, so its foot can't be compared in absolute values, only in ranges.

What moves each part

Left side (the right one is symmetric). Flexion / turn / tilt, in degrees. Library: Mixamo model; with Rigify it is the same except where noted.

Segment	Mixamo	Rigify	Moved by	Library	CMU	SFU
Pelvis	Hips	spine	PelvisController	3.3 / 9.8 / 7.8 (previously 0.2 of flexion)	4.2 / 10.3 / 10.6	6.7 / 9.0 / 16.6
Lower spine	Spine	spine.001	TorsoController	4.2 / 5.9 / 6.2 (previously 3.1 / 5.9 / 8.7)	3.0 / 4.7 / 11.4	3.0 / 4.3 / 3.7
Middle spine	Spine1	spine.002	nothing	still	1.2 / 6.6 / 2.6	(included in the chest)
Chest	Spine2	spine.003	TorsoController	0.0 / 6.0 / 2.7 (previously 0 of tilt)	1.7 / 3.9 / 6.0	4.6 / 3.6 / 13.9
Neck	Neck	spine.004	TorsoController (only resets it)	still	3.1 / 1.6 / 2.0	7.4 / 6.4 / 3.6
Upper neck	—	spine.005	nothing	still	4.9 / 2.2 / 2.8	—
Head	Head	spine.006	TorsoController	3.3 / 3.8 / 1.4	1.7 / 1.0 / 1.0	15.0 / 4.7 / 6.6
Clavicle	LeftShoulder	shoulder.L	ArmController	0.4 / 5.7 / 0.0	(doesn't move)	5.9 / 7.1 / 4.1
Upper arm	LeftArm	upper_arm.L	ArmController	21.9 / 0.7 / 3.5	15.3 / 19.6 / 7.5	42.0 / 25.6 / 12.4
Forearm	LeftForeArm	forearm.L	ArmController	14.7 / 1.4 / 2.6	14.8 / 2.3 / 3.2	23.7 / 10.2 / 8.2
Forearm twist	—	forearm_twist.L (optional)	TwistController	doesn't oscillate	1.5 / 10.0 / 3.5	—
Hand	LeftHand	hand.L	ArmController	8.5 / 1.0 / 0.0	13.9 / 4.6 / 5.1	8.4 / 3.5 / 6.7
Fingers, 1st phalanx	LeftHandIndex1…	f_index.01.L…	HandFingers	10 / 1 / 0 (+ relaxed pose)	—	—
Fingers, 2nd and 3rd	LeftHandIndex2-3…	f_index.02-03.L…	HandFingers	fixed pose, they don't oscillate	—	—
Thigh	LeftUpLeg	thigh.L	LegIK	44.9 / 13.8 / 15.0	44.5 / 10.2 / 16.2	45.1 / 19.5 / 26.6
Leg	LeftLeg	shin.L	LegIK	48.8 / 2.2 / 3.4 (Rigify: 40.4)	52.5 / 2.5 / 4.3	69.7 / 4.9 / 11.9
Foot	LeftFoot	foot.L	LegIK + FootRoll	31.6 / 7.9 / 19.3 (Rigify: 36.7 / 2.2 / 2.4)	30.8 / 22.0 / 9.8	34.1 / 17.3 / 25.9
Toes	LeftToeBase	toe.L	FootRoll	42.8 / 0.0 / 12.4	23.6 / 3.3 / 3.2	19.0 / 3.5 / 5.3
Twist (optional)	—	upper_arm_twist.L…	TwistController	upper arm 9.6 of turn, thigh 6.1	—	—

Bones nothing moves

Model	Bones	What they are
Mixamo	Spine1	middle spine: without a mapping, the curve of the back bends only at Spine and Spine2
Mixamo	*_End, *Index4, *Ring4, *Thumb4, HeadTop_End	tips: they don't deform the mesh
Rigify	spine.002, spine.005	middle spine and upper neck, like Spine1
Rigify	pelvis.L/R, breast.L/R	pelvis and chest of the mesh: they follow their parent; they could bounce a little
Rigify	palm.01–04.L/R	palm: opening or cupping it with the relaxed hand
Rigify	forearm.L.001	second segment of this model's forearm
Rigify	heel.02.L/R	metarig helper for the heel: doesn't deform
Rigify	face (jaw, eye.L/R, lid.*, brow.*, lip.*…, ~90)	outside the gait: gaze, blinking, jaw

The Mixamo foot doesn't twist

In the inventory, the Mixamo foot seems to tilt 19° sideways and the Rigify one 2°. It isn't a fault of the foot roll:

• The Mixamo model has the toes pointing 16° outwards at rest (Rigify: 2°), and the library keeps that angle when walking (16.2° in stance).
• The foot bends about its own lateral axis, which is rotated by those 16°. Seen in the character's axes, part of that flexion shows up as sideways tilt (in world space, the character's lateral axis tilts from −5° to 13°).
• Measured about the foot's own axis, the sideways tilt is exactly 0° in both models, in stance and in flight.

Two things that do come out of this:

• The toe-out angle depends on the model: 16° with Mixamo, 2° with Rigify, 0° with base_pbr_test_03. The clinical literature gives about 5–7° outwards when walking. A toe-out angle parameter, independent of the model, would make everyone walk the same.
• The foot never twists sideways on flat ground (inversion and eversion). In people it twists a few degrees: inwards at toe-off, outwards when taking the weight.

What the gait is missing

Sorted by how much the usable data (CMU and the literature) supports it and by how noticeable it is.

1. The upper arm barely rotates about its own axis. Library: 0.7° of turn; CMU: 19.6°; SFU: 25.6°. In a hanging arm, the turn about the vertical is mostly the rotation of the humerus about its axis (inwards and outwards), not a swing out of the plane: measured with points, the CMU arm opens outwards when going forward (±4.7°; the library, ±2.3°, in the same direction), and the wrist trajectory already matches CMU's (tests/tools/mocapArm.ts). What is missing is the rotation of the humerus, which shows in the twist of the arm mesh and in the plane of the elbow. The size of the arm swing is left alone: CMU swings less than the library and SFU more.
2. The pelvis doesn't tilt forward and back. Done (pelvisTiltSwing): ±1.8° twice per stride; error against CMU from 1.3° to 0.2°; range of the bone from 0.2° to 3.3° (CMU: 4.2°). Its lateral drop is still somewhat short (7.8° vs 10.6° in CMU; see naturalness).
3. The chest only turns. Done (chestRollShare): the chest does ~1/3 of the compensation of the pelvic drop (sideways tilt from 0° to 2.7°; CMU: 6°) and the lower spine the rest (from 8.7° to 6.2°). The ratio between the two is CMU's, but in CMU both move more (11.4° and 6°) and with phase offsets: that needs the middle spine (point 4). The shoulders stay just as level (error against CMU: 1.0° → 1.1°).
4. The middle spine doesn't move. In CMU it takes part of the counter-rotation (6.6° of turn). The turn and tilt need to be distributed among all the segments of the spine.
5. The neck doesn't move. In CMU its two segments bend 3–5° at each step and the head barely at all (1–2°). The library's head moves somewhat more (3–4°): part of that movement should go to the neck. SFU moves head and neck much more (15°), but that subject is the exception.
6. The wrist: no data to decide. CMU's lhand bone is ~1 cm long (almost the whole hand is lwrist, which only rotates about the forearm), so CMU doesn't measure the wrist: measured with points, its hand barely departs from the forearm (±1°). SFU does measure it: somewhat more towards the body than the library (range 6° vs 3.4°) and less forward and back (±1.4° vs ±5°). With a single source (and local only), it is left alone.
7. The forearm doesn't rotate (pronation and supination): CMU, 10°. It would only show in rigs with a forearm twist bone.
8. The fingers are a fixed pose: only the first phalanx oscillates (10°). As a chain of pendulums, each phalanx would follow the previous one with some lag. There is no finger mocap: this is tuned by eye.
9. The foot is rigid sideways and its toe-out depends on the model (see above).

Unchanged:

• Clavicle: only SFU moves it more (CMU doesn't measure it). By the SFU rule, it isn't touched without support from the literature.
• Toes: the library bends them 43° and both mocaps ~20°, but markers measure that joint poorly and the clinical literature gives quite a lot more than the mocap.
• Fine as it is: the turn of the pelvis, the lower spine, the thigh, the forearm and the flexion of the foot match CMU.
• To look at later: with Rigify the knee bends less when walking (40° vs 49° with Mixamo and 52° in CMU), probably because of the proportions of the model.

Mapping proposal

Without breaking the current mappings:

• Intermediate spine and neck bones, automatic: the bones in the parent chain between spine and chest (Spine1, spine.002) and between neck and head (spine.005) are detected automatically when creating the character, and the turn is distributed among all the segments. The inventory tool already finds them this way. The mapping doesn't need to change. An optional field (for example spineChain: string[]) would allow setting them by hand.
• Neck: TorsoController distributes the movement of the head between neck, the upper neck (if it exists) and head.
• Palm (palm.*): optional palm field in each hand's fingers, to cup it with the relaxed hand. Low priority.
• Rigify mesh bones (pelvis.L/R, breast.L/R): not mapped. They follow their parent, which already moves. The bounce is a secondary effect that can come later.
• Face: outside this stage.

This changes the public API (new optional fields in HumanRigMapping): it has to be approved first. Points 1 and 9 of the list above don't need new bones: they can be done with today's mapping (2 and 3 were already done that way).