X2 GMR v4 Toe-Target Corpus & SONIC Relaunch

Retargeting milestone, audited baseline, and 32-GPU from-scratch launch | 2026-07-24

X2GMR v4SONICretargetingevaluation

1 | Milestone

121,486 / 121,486 motions completed through GMR v4 and MotionLib. The frozen manifest SHA256 is a8b28f625576be099abc4658ed3e4298dfa1047c092291fcf0659e951291dfbe.

Layer	Contract
Robot	29 actuated DOFs; fixed neck; physical head retained
Retarget only	Massless, jointless, geometry-free toe IK markers at `[0.10, 0, -0.073]` in ankle-roll frames
Training	Same 29-DOF physical robot, no toe bodies or toe state slots
Hand endpoint	X2 `wrist_roll`, replacing G1 `wrist_yaw` consistently in reward, termination, eval, and mass randomization

2 | New v4 reference gallery

Representative MotionLib references rendered directly on the canonical fixed-neck, physical-head X2 MJCF. These are kinematic references, not policy rollouts.

Forward walk loop

Backward jog loop

Turning jump

Latino chase / mambo

3 | Robot movement speed

v4 is slower, not faster: median horizontal root-speed ratio is 0.883, or 11.7% below the previous production-like 0.9-scale corpus. The rejected blanket 0.7 corpus is not the baseline.

This measures reference robot root displacement per second with source FPS preserved. It is not video duration, playback speed, or trained-policy speed. A seed-0 reproducibility pass selected 512 clips and retained 452 moving clips above 0.02 m/s; its median was 0.884.

The v4-to-old speed ratio distribution. The old comparison corpus is the prior production-like 0.9 lineage.

4 | Historical replay and primary baseline

A historical released-default replay on common-v4 scored 104/126 (82.5%), but it mixed encoders, retained the hidden ankle-XYZ termination, and inherited the G1 wrist-mass target. The audited primary 512-motion 50k 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. This is the checkpoint baseline the new v4-trained policy must beat.

The 126-clip ruler is an internal X2 regression set, not SONIC's private paper split. Null-termination speed/drift probes remain diagnostics and are not reported as official success.

Historical released-default callback metrics for the pre-v4 50k checkpoint on common-v4; see the audited 512-motion scorecard for the primary protocol.

5 | SONIC from-scratch launch and caveats

Item	Audited state
Paper/source	arXiv 2511.07820v3; official NVlabs source `4141c34`
Initialization	Random policy initialization through the official SONIC code path; `checkpoint: null` and no `pretrained_model`
Recipe	50k iterations, 4096 env/GPU, 24 rollout steps, 32 GPUs; controlled X2 anchor/linear-velocity overrides retained
Launch	Saturn A100 Beaker experiment `01KYBHC11H7HP9GYX1YKRS8RP7`, four synchronized 8-GPU replicas
Startup validation	32/32 ranks loaded all 121,486 motions; PPO entered normally; first scratch `last.pt` written at iteration 50
Early throughput	About 7.8 seconds/iteration, approximately 112 hours to 50k without preemption

No G1 or historical X2 checkpoint is loaded; warm-start runs were stopped.
The run is an X2-adapted recipe, not a zero-override paper reproduction.
32 GPUs provide one quarter of the paper's largest 128-GPU per-update batch; historical single-GPU scratch runs stayed flat, so convergence remains a real caveat.
High-priority replicas are interruptible. Replacements resume only from this scratch lineage's own last.pt, saved every 50 iterations.
Final audited primary callback eval and X2 motion diagnostics run after the 50k checkpoint.

6 | Retargeter decision

GMR and SOMA are both kinematic retargeters: neither contains robot dynamics or learns physical feasibility. GMR v4 is selected because its X2 mapping is stable and the toe targets improve ankle/foot geometry without changing the training robot. SOMA and SOMA/GMR grafts are closed for this project after sufficient negative validation.

X2 v4 milestone | references: SONIC paper | official source