Fix MJCF D6 collapse dropping distinct axes and leaving unused D6 axes free

[aidaodedjl]

Description

Fixes the PhysX D6 collapse step of the MJCF import pipeline (Isaac Sim 6.0, mjc_to_physx_conversion_utils) dropping distinct joint DOFs and leaving unused D6 axes free. Reported in isaac-sim/IsaacLab#6854.

Problem

When several single-axis joints connect the same body pair (e.g. a humanoid hip/shoulder/spine authored as three hinge joints — the standard authoring style in MJCF), _convert_overconstrained_group_to_d6() collapses them into one D6 joint. The axis of each source joint was derived from physics:axis alone.

But mujoco-usd-converter always exports single-axis joints x-aligned and encodes the real MJCF axis direction in localRot0/localRot1. So physics:axis is X for every joint in such a group. Three consequences:

1. Lost DOFs: every joint after the first collided on the rotX token and was dropped with a "duplicates D6 axis" warning — even when its physical axis was distinct and exactly representable by a D6. A 3-hinge hip collapsed to a 1-DOF joint.
2. Wrong rotation directions: the D6 frame was left unrotated, so surviving axes rotated about the wrong directions whenever the MJCF axes were not cardinal.
3. Spurious free DOFs: the unused D6 axes were never locked, so the collapsed joint gained free translational DOFs in PhysX (the "robot explodes / drifts away" symptom in the report).

Fix

• Recover the physical axis of each source joint from localRot0/localRot1 (_joint_axis_directions()).
• Build a right-handed orthonormal D6 basis from the group's axes via Gram–Schmidt, and assign each joint to a basis column by dot product (_assign_axes_to_d6_basis()). The D6 frame now encodes the real axis directions.
• Joints that genuinely cannot be represented (duplicate / non-orthogonal axes) are still dropped, but with a physically accurate warning.
• Explicitly lock all six D6 axes except those backed by a source joint.
• Record the source MJCF joint name on each used axis as mjcf:<axis>:name for traceability.

Validation

Unit tests (test_mjc_d6_collapse.py, plain pxr, no Kit): 8 tests covering 3-distinct-axes groups, duplicate-axis drops, mixed revolute/prismatic groups, 2-joint groups, per-axis limits/drives landing on the correct axis, unused-axis locking, and mjcf:<axis>:name metadata.

End-to-end (Linux, real pipeline: mujoco-usd-converter 0.5.0 → convert_mjc_to_physx → combine_overconstrained_joints_to_d6), original vs patched module on the issue's own repro models (ProtoMotions soma23_humanoid.xml / smpl_humanoid.xml) plus a minimal rotated-triad case:

Metric	before	after
soma23: D6 joints with per-axis mjcf:<axis>:name metadata	0/22	22/22
soma23: D6 joints with a free translational axis	22/22	0/22
smpl: D6 joints with axis metadata	0/23	23/23
smpl: D6 joints with a free translational axis	23/23	0/23
rotated triad (two 45° diagonal hinges): preserved rotational DOFs	1/3	3/3

In the rotated-triad case the old code logged j_b duplicates D6 axis 'rotX' ... DOF will be lost for a joint whose physical axis was perfectly distinct — with the fix all three DOFs survive with correct per-axis limits ([-45,45], [-30,60], [-90,90]), the D6 frame reproduces the diagonal world directions, and all translation axes are locked.

Checklist

• Bug fix (non-breaking change which fixes an issue)
• Regression tests added (standalone_tests/test_mjc_d6_collapse.py, runs with plain pxr)
• Validated end-to-end against the issue's repro models

Refs isaac-sim/IsaacLab#6854