← Back to project

X2 SONIC Audited im_eval: 512-Motion Scorecard

Eight historical policies, one frozen GMR v4 ruler | 2026-07-25
SONICX2im_evalcheckpoint comparison

1 | Primary result

Best source-audited primary success: 480/512 (93.75%). Progress is 95.32%, successful-only MPJPE-L 44.29 mm, all-motion MPJPE-L 47.56 mm, and all-motion MPJPE-G 249.64 mm.

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.

2 | Primary checkpoint ranking

CheckpointSuccessRateProgressSucc MPJPE-LAll MPJPE-LAll MPJPE-GCommon-126
Mainline 50k480/51293.75%95.32%44.2947.56249.64118/126
Mainline 26k477/51293.16%94.54%41.9045.79426.18117/126
Mainline 34k477/51293.16%94.96%43.5946.89372.22118/126
Mainline 42k476/51292.97%94.83%42.9346.92286.69117/126
Baseline no-rescale 15.9k465/51290.82%92.49%40.7045.38429.05114/126
Anchor-v2 8k465/51290.82%93.15%46.3151.08299.83115/126
G1 warm-start 12.3k464/51290.62%93.60%43.4548.04501.43116/126
v4 scratch 2.2k412/51280.47%87.13%57.3963.67419.75104/126

3 | What the open-source eval actually does

ComponentObserved source behavior
Seed/policySeed 0; deterministic act_inference; CUDA bitwise determinism not enabled
Startup eventsFriction/restitution, default joint position, torso COM, and wrist/torso mass are randomized
Per-step observationsPolicy and tokenizer corruption remain enabled
TerrainRandom 1-5 mm boxes/roughness from inherited trimesh
Disabled for evalPushes, 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.

4 | 50k runtime ablation

RuntimeSuccessRateProgressSucc MPJPE-LAll MPJPE-L
Released runtime (primary)480/51293.75%95.32%44.2947.56
No startup physics DR483/51294.34%95.72%43.2646.00
No startup DR or observation noise484/51294.53%95.89%43.4446.09

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.

5 | Metric scope and paper context

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.

6 | Secondary reproductions

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.

CheckpointSuccessRateProgressSucc MPJPE-LAll MPJPE-LAll MPJPE-G
Mainline 50k415/51281.05%86.30%44.2748.40190.50
Mainline 26k414/51280.86%86.38%42.2746.05300.76
Mainline 42k407/51279.49%85.83%43.5847.19222.04
Mainline 34k406/51279.30%86.29%42.9847.17267.74
Baseline no-rescale 15.9k403/51278.71%84.85%41.8545.26295.85
Anchor-v2 8k377/51273.63%82.31%45.0949.76222.03
G1 warm-start 12.3k359/51270.12%79.22%42.4646.37315.75
v4 scratch 2.2k332/51264.84%75.07%54.7759.21296.77

Paper terminations + mixed encoder. This isolates the termination difference but still randomly mixes three encoders.

CheckpointSuccessRateProgressSucc MPJPE-LAll MPJPE-LAll MPJPE-G
Mainline 50k478/51293.36%94.87%46.8750.04236.87
Mainline 42k475/51292.77%94.53%45.0849.00261.20
Mainline 26k474/51292.58%94.34%44.4148.23394.29
Mainline 34k472/51292.19%94.28%45.1649.07332.47
G1 warm-start 12.3k468/51291.41%93.79%46.6550.77456.03
Baseline no-rescale 15.9k458/51289.45%91.93%43.1147.94350.55
Anchor-v2 8k458/51289.45%92.25%49.1352.90268.93
v4 scratch 2.2k407/51279.49%86.37%58.8264.59416.47

7 | Ruler and implementation provenance

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.

This is an internal X2 regression ruler, not a SONIC paper split. It supports controlled comparisons among our checkpoints, not a claim of matching paper-scale G1 performance.
Source-audited SONIC im_eval; frozen internal v4 512-motion ruler