The trainer read a stale USD after the URDF mass update - regenerate derived assets and gate on an automated equality instrument

Symptom

Measured link masses had been committed to URDF/MJCF (total 9.58 -> 9.792 kg, weighed values), but Isaac reads the derived USD asset - which still carried the old masses: a silent 2.2% mass fork between the training plant and the evaluation plant.

Context

The v12 checklist made USD regeneration a hard precondition ("硬性 前置") and, crucially, backed it with an instrument: check_usd_mass.py compares USD vs URDF per-link mass AND inertia trace, validated by showing it FAILs the stale asset naming 9 offending links, then PASSes after re-conversion (13/13 links consistent, total 9.7920). The self-collision filter audit was re-run after the regeneration (three poses, 0.00 N) because a regenerated asset invalidates physical audits done on the old one. This milestone was also where the three plants first aligned: "三边 plant 首次对齐(armature+摩擦+ 实称质量)就在这一代".

Change

convert_urdf re-run on the training machine, regenerated asset committed, check_usd_mass.py PASS required as an acceptance gate for the generation; dependent audits repeated post-regeneration.

Outcome

The 2.2% plant fork was closed before it could distort a generation's acceptance numbers; the staleness class of bug now has a permanent detector instead of a memory.

Mechanism

Source-of-truth edits do not propagate to derived binary assets by themselves; any consumer reading the derivative silently trains or evaluates on the old plant. An automated equality check between source and derivative - proven able to fail - turns an invisible staleness into a red gate, and regeneration invalidates every audit performed on the old artifact.

Applies when

• editing masses/inertia/geometry in URDF or MJCF sources
• a trainer or evaluator consumes converted/derived assets
• plant numbers differ between simulators for no visible reason