X2 SONIC · embodiment audit: hardcode bugs, vendor-file conflicts, ruled-out hypotheses

4 torque typos fixed · 29/29 torque transcription certified · KP/KD = official SOP mimic profile · pelvis-mass resolved (MJCF artifact) - ALL closed · 2026-07-12

auditbugfixhumanoidsim2real

1 · TL;DR

Deep audit of our X2 stack vs official SONIC: after fixing 4 transcription bugs in x2.py torque limits, the codebase differs from official ONLY in the G1→X2-Ultra swap. No hidden physics or hyper-parameter deviations. All vendor-data questions RESOLVED: KP/KD = AgiBot official SOP mimic profile; the pelvis-mass conflict was a MuJoCo density-fallback artifact.

Trigger: policy videos showed laggy wrists and slow locomotion. We chased it through data (GMR), hardware limits, and configs — the trail ended at hand-transcribed constants and two AgiBot vendor-file inconsistencies, NOT at retargeting or embodiment ceilings as first assumed.

2 · Bugs found & fixed (x2.py hardcoded torque limits)

constant	was	real (URDF & MJCF agree)	impact
`EFF_WRIST_PR`	2.2	4.8	wrist torque halved — the main wrist-lag culprit
`EFF_HIP`	118	120	1.7%
`EFF_KNEE`	118	120	1.7%
`EFF_WAIST_YAW`	118	120	1.7%

x2.py's own comment says these were copied from the MJCF ctrlrange — they were mistyped. Both vendor files agree on the correct values, so the fixes are unambiguous. They take effect on the next training run (running arms loaded the old values at start).

3 · What was ruled OUT (each with hard evidence)

GMR retargeting bug — ruled out: on a super-fast vogue clip, human-source wrist peaks at 33.9 rad/s while our X2 retarget outputs 11.9 (GMR compresses, never amplifies); fps source==output; and the pipeline reproduces ToberJ's production v3 to 1.28°/joint.
'X2 wrist motor 4.2 rad/s hardware ceiling' — ruled out: probe rollouts show the trained policy actually reaches 25-50 rad/s at the wrist in sim (IsaacLab ignores URDF velocity for implicit actuators; x2.py sets no velocity_limit_sim). The 4.2 URDF field is also physically inconsistent with its own 4.8 Nm torque (G1 wrist: ~5 Nm & 22-37 rad/s).
DOF/body mapping errors — ruled out: all 29/29 mappings verify by name and are exact inverses.
Elbow double sign-flip — ruled out: dof_sign correctly disabled for GMR X2-native data.

4 · Codebase diff vs official SONIC (exhaustive)

Recursive diff of gear_sonic/ and the full config tree:

layer	result
config/** (rewards, PPO, obs, terminations, events)	byte-identical; only addition = sonic_x2.yaml
sonic_x2.yaml vs sonic_release.yaml	robot=x2, data paths, warm-start block, critic dims (= official default, no-op), removed upper_body_augment_prefixes (official-internal clip names)
14 changed shared .py files	X2 registration (4), dof_sign infra (no-op for G1), /dev/shm fix; the rest = upstream is newer than fork base (teleop tooling refactors, not on the training path)
rigid/articulation props vs g1.py	12/12 identical (solver 8/4, self-collision, damping…)
action_scale formula	same as G1 (0.25·effort/stiffness)

Guarantee: with the fixes applied, our stack = official SONIC + (G1→X2 Ultra embodiment, X2-native corpus, warm-start source, encoder-routing choice). Nothing else.

5 · Open vendor questions (blocked on external ground truth)

#	question	evidence	owner
1	Pelvis mass — RESOLVED ✓	The MJCF's 5.03 kg is NOT an authored value: pelvis is the only body in the vendor MJCF lacking an explicit inertial, so MuJoCo auto-computes it from collision-geom volume × default density. URDF's 3.52 kg is the only authored value (byte-identical in third-party ioai-tech/robot_description redistribution). Training uses URDF — correct all along; MJCF serves kinematics only.	closed
2	KP/KD provenance — RESOLVED ✓	x2.py gains = AgiBot official SOP 'AgiBot X2 motion-control deployment SOP (vendor PDF)' §3.1 mimic profile verbatim (the profile AgiBot pre-tuned FOR motion-mimic deployment — exactly our use case). Cross-checked vs the AimDK SDK joint-control example: wrist 20/2 identical in both profiles (official choice, likely wrist-motor protection); legs stiffer in mimic (120/150) vs safe-demo (40/80) as expected. SDK example joint ranges match our Ultra URDF digit-for-digit → same robot variant confirmed.	closed
3	URDF velocity field (wrist 4.19 rad/s) inconsistent with torque; unused by training	Parked — zero impact on current runs.	(later)

6 · Corrected understanding of the video observations

Wrist lag → halved torque (bug, fixed) + soft KP (pending verification) + 40%-trained checkpoint. NOT a hardware ceiling, NOT data, NOT GMR.
Sideway walk at 52% of reference speed → under-training at 8k/20k iters (leg torque is strong and correct); re-evaluate at 14k+.
Official G1 convergence (0.98 succ / 29mm mpjpe) is achievable in principle — our references are X2-native and the embodiment is not the bottleneck we thought.

auto-generated · X2 SONIC · full evidence chain in fork docs/perm_alignment_verified.md + this audit