This ranking fixes the robot-motion encoder, uses the paper's termination thresholds, retains the actual released eval runtime, and applies the correct X2 wrist mapping.
| Checkpoint | Success | Rate | Progress | Succ MPJPE-L | All MPJPE-L | All MPJPE-G | Common-126 |
|---|---|---|---|---|---|---|---|
| Mainline 50k | 480/512 | 93.75% | 95.32% | 44.29 | 47.56 | 249.64 | 118/126 |
| Mainline 26k | 477/512 | 93.16% | 94.54% | 41.90 | 45.79 | 426.18 | 117/126 |
| Mainline 34k | 477/512 | 93.16% | 94.96% | 43.59 | 46.89 | 372.22 | 118/126 |
| Mainline 42k | 476/512 | 92.97% | 94.83% | 42.93 | 46.92 | 286.69 | 117/126 |
| Baseline no-rescale 15.9k | 465/512 | 90.82% | 92.49% | 40.70 | 45.38 | 429.05 | 114/126 |
| Anchor-v2 8k | 465/512 | 90.82% | 93.15% | 46.31 | 51.08 | 299.83 | 115/126 |
| G1 warm-start 12.3k | 464/512 | 90.62% | 93.60% | 43.45 | 48.04 | 501.43 | 116/126 |
| v4 scratch 2.2k | 412/512 | 80.47% | 87.13% | 57.39 | 63.67 | 419.75 | 104/126 |
| Component | Observed source behavior |
|---|---|
| Seed/policy | Seed 0; deterministic act_inference; CUDA bitwise determinism not enabled |
| Startup events | Friction/restitution, default joint position, torso COM, and wrist/torso mass are randomized |
| Per-step observations | Policy and tokenizer corruption remain enabled |
| Terrain | Random 1-5 mm boxes/roughness from inherited trimesh |
| Disabled for eval | Pushes, motion reset pose/velocity/joint jitter, and motion augmentation |
Startup parameters are sampled once per environment and reused across the four 128-motion loops. With identical seed, ordering, and batching, checkpoints face the same indexed perturbation setup.
| Runtime | Success | Rate | Progress | Succ MPJPE-L | All MPJPE-L |
|---|---|---|---|---|---|
| Released runtime (primary) | 480/512 | 93.75% | 95.32% | 44.29 | 47.56 |
| No startup physics DR | 483/512 | 94.34% | 95.72% | 43.26 | 46.00 |
| No startup DR or observation noise | 484/512 | 94.53% | 95.89% | 43.44 | 46.09 |
The nominal rows are audit ablations, not the primary ranking. The released-runtime row follows the open-source execution path after the required X2 endpoint correction.
The callback exposes both eval/success/* (successful motions only) and eval/all/* (all 512). The paper says MPJPE-L over 14 links but does not name the callback scope, so both are shown. MPJPE-G is useful for internal drift analysis but has no paper counterpart.
SONIC reports 23.2 mm MPJPE-L in its main comparison; the 32-GPU FSQ-32-32 ablation reports 27.5/26.3 mm, and the 128-GPU robot encoder reports 23.8/22.5 mm. Lower is better. Paper tables and protocol.
Released merge + mixed encoder. This reproduces the open release defaults, including the extra ankle XYZ termination and random encoder selection. It is historical reproducibility, not the primary checkpoint comparison.
| Checkpoint | Success | Rate | Progress | Succ MPJPE-L | All MPJPE-L | All MPJPE-G |
|---|---|---|---|---|---|---|
| Mainline 50k | 415/512 | 81.05% | 86.30% | 44.27 | 48.40 | 190.50 |
| Mainline 26k | 414/512 | 80.86% | 86.38% | 42.27 | 46.05 | 300.76 |
| Mainline 42k | 407/512 | 79.49% | 85.83% | 43.58 | 47.19 | 222.04 |
| Mainline 34k | 406/512 | 79.30% | 86.29% | 42.98 | 47.17 | 267.74 |
| Baseline no-rescale 15.9k | 403/512 | 78.71% | 84.85% | 41.85 | 45.26 | 295.85 |
| Anchor-v2 8k | 377/512 | 73.63% | 82.31% | 45.09 | 49.76 | 222.03 |
| G1 warm-start 12.3k | 359/512 | 70.12% | 79.22% | 42.46 | 46.37 | 315.75 |
| v4 scratch 2.2k | 332/512 | 64.84% | 75.07% | 54.77 | 59.21 | 296.77 |
Paper terminations + mixed encoder. This isolates the termination difference but still randomly mixes three encoders.
| Checkpoint | Success | Rate | Progress | Succ MPJPE-L | All MPJPE-L | All MPJPE-G |
|---|---|---|---|---|---|---|
| Mainline 50k | 478/512 | 93.36% | 94.87% | 46.87 | 50.04 | 236.87 |
| Mainline 42k | 475/512 | 92.77% | 94.53% | 45.08 | 49.00 | 261.20 |
| Mainline 26k | 474/512 | 92.58% | 94.34% | 44.41 | 48.23 | 394.29 |
| Mainline 34k | 472/512 | 92.19% | 94.28% | 45.16 | 49.07 | 332.47 |
| G1 warm-start 12.3k | 468/512 | 91.41% | 93.79% | 46.65 | 50.77 | 456.03 |
| Baseline no-rescale 15.9k | 458/512 | 89.45% | 91.93% | 43.11 | 47.94 | 350.55 |
| Anchor-v2 8k | 458/512 | 89.45% | 92.25% | 49.13 | 52.90 | 268.93 |
| v4 scratch 2.2k | 407/512 | 79.49% | 86.37% | 58.82 | 64.59 | 416.47 |
The ruler contains common-126 plus 386 motions selected by stable SHA-256 ranking from 121,486 GMR v4 motions. Manifest SHA256: 05c108fdf04bb98b0dd4b1a6f99cc29776437d70c43b7b419769a9829e3d2a11.
The X2 callback replaces the upstream hardcoded G1 wrist subset with configured embodiment endpoints. Its local metric formulas match SMPLSim commit b5c0872 exactly on deterministic random-array parity tests.