X2 SONIC 50k: Corrected Evaluation

The 38% diagnosis is withdrawn; evaluation semantics are now source-audited | 2026-07-25

SONICX2evaluationcorrection

1 | Correction

The historical 50k policy was substantially better than the former 38% report. The corrected 512-motion primary result is 480/512 (93.75%), with 95.32% progress, successful-only MPJPE-L 44.29 mm, and all-motion MPJPE-L 47.56 mm.

The deleted project-local evaluator checked root XY and full 3D endpoint error. SONIC success checks root/end-effector height and root orientation. The old number cannot be reused as SONIC success, a framework ceiling, or evidence that this checkpoint failed.

2 | Final primary contract

Layer	Setting
Data	Frozen 512-motion GMR v4 ruler
Encoder	Fixed robot-motion encoder (internal key `g1`)
Termination	Paper thresholds: root/end-point height 0.25 m, root orientation 1 rad, timeout succeeds
Runtime	Released-code startup randomization, observation noise, and random micro-terrain retained
X2 mapping	Hand endpoint and hand-mass target use `wrist_roll`, not G1 `wrist_yaw`

3 | Two upstream mismatches

The released config merge retains the training foot_pos_xyz guard, adding a 0.2 m full ankle XYZ termination absent from the paper prose. It also leaves startup domain randomization, observation corruption, and rough terrain active even though the paper describes domain randomization as a training setting. We report these facts instead of calling release-code and paper protocols identical.

4 | What remains valid

The anchor recipe improved horizontal displacement attainment from 0.70 to 0.91; this remains a locomotion diagnostic, not success.
The v4 corpus fixes the ankle-as-foot endpoint and uses measured X2 leg scale plus retarget-only sole markers.
SOMA remains rejected after sufficient X2 validation.

5 | Paper context

Lower MPJPE is better. SONIC reports 23.2 mm in its main comparison, 27.5/26.3 mm for the 32-GPU FSQ-32-32 ablation, and 23.8/22.5 mm for the 128-GPU robot encoder.

The paper reports MPJPE-L but not MPJPE-G and does not identify whether its number maps to the callback's successful-only or all-motion aggregate. See SONIC arXiv v3.

Source-audited SONIC im_eval; frozen internal X2 v4 ruler