Scenario Maya Rigging

Maya rigging

Expert rigging means every value is explained and every behavior measured: joints at real pivots with computed orientation, controls at zero with their placement in offsetParentMatrix, IK/FK and space switches that move nothing, and a rig that survives scale, distance, several characters per shot and an animator keying everything. The rig is the output of a build script and data. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-maya-expert (execution channel, review loop, 2027 version traps).

Module scripts/mx_rig.py (sys.path.insert(0, "<skills>/scenario-maya-rigging/scripts"); import mx_rig as R): not yet run in Maya; its pure math passed offline (test_rig_offline.py). With Maya installed, run tests/code/maya-rigging/run_all.sh and fix what test_rig_probe contradicts.

Stance (the expert delta)

• Clean values, one place for each. Rotations zero at rest, orientation in jointOrient (or nowhere: Fragapane), control placement in offsetParentMatrix (antCGi; Maya Learning Channel). OPM composes, world = local * OPM * dagParentWorld (2027 transform node): feed .matrix for a relative drive, or multiply a worldMatrix by the DAG parent's worldInverseMatrix, then zero translate, rotate and jointOrient or they apply twice (Maya Learning Channel; antCGi's leg rotating backwards).
• Compute, never eyeball. -orientJoint silently skips joints with rotations or no child (2027 cmds.joint); a secondary world axis along the bone is ill-defined (antCGi's toes). R.joint_chain computes jointOrient from positions; a pole on the chain plane adds zero drift (R.pole_vector_position) [added].
• The animator defines done. Zero pop switching IK/FK at neutral, only the active mode visible, everything else locked because animators "select hierarchy and key all", limits at breaking points, behavior mirroring on limbs, same orientation on eyes and face (antCGi).
• Matrices before constraint stacks, and know the blend. OPM, blendMatrix and pickMatrix keep channels free (Maya Learning Channel cut 20 joints from one hand). blendMatrix is an ordered stack, each target overriding the ones before (equal thirds: 1, 1/2, 1/3); parentMatrix (2025) normalizes and stores per-target offsets (2027 Help). Spaces blend offset targets, never the space itself, or the IK flips (antCGi).
• The built rig is disposable. Guides, build script, serialized weights and settings; fixes go back into the data (Miquel Campos, mGear: "we serialize everything and we build clean every single time"; Bungie). Metadata and joint labels, not names, identify chains (Bungie). A skinCluster reads only influence world matrices and bindPreMatrix: reset the bind pose, never rebuild joints for it (Fragapane).
• Structure decides speed. The Evaluation Graph leaves static curves out and rebuilds at the first differing key; without curves a rig "is going to be paralyzed terribly" (Fragapane; Using Parallel Maya 2027). Split components so cheap global work finishes early, even by duplicating cheap math (Fragapane); no Python nodes, scene-querying expressions or driven frozen; hosts out of the limbs (Miquel Campos); DG, Serial and Parallel agree before optimizing.
• Test at the extremes. Scale 1.24 and 1.6 (antCGi), a million units away and upside down (Fragapane), three referenced characters, a quick bind while building (antCGi).
• Deformation skeleton apart from the control rig. Engines see joints, skin and blendshapes only: one root at the origin, engine names, a joint budget (antCGi, Unreal and Unity docs, Bungie).

Establish first

Choose the path

Workflow

1. Gate the model. mx_validate.validate(profile="model") clean; cm; height to spec (about 180); feet at y = 0; centered on X; frozen; eyes dead ahead; loops at knuckles, elbows, knees; mouth bag for a face (antCGi). Send problems back as a list; never edit a client model unasked. GATE: no fail; wire sheet looked at.
2. Place joints at pivots on edge loops: spine joint under the rib cage (antCGi), wrist decoupled from the palm (Fragapane), elbows pre-bent back, knees forward; the first upper-arm and thigh twist joint stays at the top with the shoulder or hip, mid twist joints are placed by position so they keep the parent's orientation (antCGi). GATE: R.chain_plane does not raise on any limb; joints over wireframe looked at.
3. Build and orient. R.joint_chain(spec, parent, up="plane") for limbs, up=(0, 0, 1), up_axis="y" for spine and neck, world=[...] for root, cog, head, eyes and feet (antCGi: animators, engine foot placement), end_aim= for clavicles; existing skeletons: R.orient_with_cmds ('zup' for legs and spine). R.mirror_chain: behavior for limbs, orientation for eyes and face; rotate orders on controls (R.rotate_order(twist, main); Fragapane); R.set_preferred_angles(root). GATE: R.rig_check(joints=bind) clean; mirror within 1e-3 cm; all digits curl on one axis and sign, thumb fixed, so one attribute makes a fist (antCGi).
4. Hand the bind skeleton to scenario-maya-deformation (it binds while you build controls): joints only, one hierarchy, labels, rig_check JSON.
5. Control rig. R.ikfk_limb(bind[:3], name, rig_parent, ctrl_parent, own_solver=char + "_ikRPsolver"): handles of a type share one solver node, so several characters in a scene conflict (Fragapane: "can make a mess"; 2027 Help). Legs: extra=[ball, toe], orient_end=False, then R.reverse_foot with R.foot_pivots_from_mesh; spine P9; R.space_switch(ctrl, [None, cog, chest], names=[...]); R.ikfk_match(meta, to="fk"|"ik", key=True) uses stored offsets (antCGi instead adds a pole locator under the FK elbow and an ankle locator oriented like the IK foot). GATE: R.neutral_switch_test and pole drift below 1e-3; R.space_test as expected; FK and IK posed apart, R.blend_sweep_test finds no flip; roll sweep plants heel, ball, toe in turn; no shared-solver error; controls at zero.
6. Scale and stretch. Root scale only on groups that do not inherit it; stretch base lengths times root scaleY; IK spine joints under a world-pivot scale group, as segmentScaleCompensate blocks scale (antCGi). GATE: R.scale_test at 1.24 and 1.6, neutral and stretched, below 1e-3 cm.
7. Lock down. R.lockdown; FK hinges keep their bend axis; limits; controller tags; selection sets. GATE: R.lockdown_audit and R.opm_audit(rig) empty; the rig still works.
8. Performance. Key first (R.rom_keys, two different values per used attribute): a keyless rig profiles fast and lies. R.eval_census(rig), R.eval_ab(bind, meshes, frames); R.prepare_eval_graph(root=rig) tags every control (with the animator's Preferences > Settings > Animation > Include controllers in evaluation graph, first keys stop rebuilding the graph) and turns on the curve manager for the session. R.gpu_override_census(meshes), then the GUI Evaluation HUD must show a non-zero k count, else deformerEvaluator -meshes says why. Profile DG, Serial, Parallel and GPU Override with 1 and 3 referenced characters; keep dated profiles (Miquel Campos). GATE: no census error; modes agree within 1e-5; each dense mesh on the GPU or its reason written.
9. Review. R.rom_keys, then mx_review.playblast (GUI) or R.pose_review (headless); R.precision_test at shot distances; references/critique.md. GATE: frames looked at; elbows, knees, shoulders, wrists, feet read right.
10. Deliver. New version, validate(profile="rig") clean, test JSON, playblast, what was not verified.