Bimanual FK audit: world-coordinate precision¶

This is a disclosed amendment to a numerical diagnostic after the two-hand and support test runs. It does not change the controller, physical model, recorded states, 170-degree success predicate, test resets or 100 micrometre tolerance. The original freeze and failed audit remain available.

Observed failure¶

The 30 independent simulations occupy origins separated by 12 m. Their global coordinates therefore extend to 348 m, although each robot moves only near its own local origin. PhysX and recorded tool coordinates use float32. The original CPU audit compares double-precision local forward kinematics with those recorded world transforms after subtracting the environment origin.

The support run completed 30/30 tasks and passed every sampled CAD, native joint limit and motor-force check. Two individual tool samples exceeded the original absolute-world FK bound: left hand, seed 92023 at 90.5667 s, 107.179 micrometres; right hand, seed 92028 at 59.1 s, 104.556 micrometres. The final queue stopped and did not open the third condition until this discrepancy was investigated.

Translation counterfactual¶

The diagnostic assigns each offending recorded joint configuration and base orientation to the same native articulation at several world translations. It calls native forward kinematics without integrating physics, applying controller commands or adding task contacts. Joint coordinates and orientations agree exactly across translations.

The discrepancy is almost entirely along the translated axis. This reproduces the audit failure by changing world origin alone; it is not evidence of a controller failure or a physical limit violation. The diagnostic source, assigned poses and raw transforms are retained with the evidence archive.

Supplemental audit¶

scripts/audit_bimanual_coordinate_precision.py checks every powered 30 Hz sample, with no phase or outcome exclusions. For each recorded pose it obtains native TCP coordinates at the original world origin and near zero, holding the recorded joint coordinates and orientation fixed. The difference measures the world-coordinate rounding contribution. It retains three separate residuals:

1. Recorded TCP versus native FK at the original world origin.
2. Native FK near zero versus independent double-precision CPU FK.
3. Recorded TCP after removing the measured coordinate contribution versus CPU FK.

Each residual must remain below the original 100 µm threshold. All original CAD, joint-position and motor-force checks still apply. The support audit checks 162,000 poses (both hands at each pose); its largest corrected TCP discrepancy is 61.182 µm. Reports preserve the uncorrected errors and hashes of the original trace, original audit, native URDF, supplemental audit source and residual arrays.

scripts/report_bimanual_coordinate_study.py uses this supplemental verification and otherwise retains the frozen result aggregation. It also checks equality of the actual initial-offset arrays across the three conditions. The original report_bimanual_study.py remains unchanged and still rejects the original support audit. The amendment is applied consistently to all modes and separate demo/reproduction trajectories; completed physical episodes are not rerun or selected again.

This establishes consistency of sampled recorded kinematics to the stated precision. It is not a proof of collision freedom between samples or calibrated hardware accuracy. The native kinematic reconstruction is an audit operation, not part of the controller or the rendered task dynamics.