rayaanoidpr/vjepa-wm-bigrun

V-JEPA quadruped world model - big run checkpoints

Frozen V-JEPA 2.1 ViT-B/384 encoder + causal tubelet masking (boundary {6,8,10})

• motion-weighted LN-L1 + per-window command tokens + velocity-consistency aux head (lambda 0.5) + command dropout 25%. fp32, batch 4, ~40 missions, stride 4.

• big_run_step4000.pt (16-mission phase)
• big_run_step8000.pt (final, ~40 missions)

Held-out eval (ICE-1 / SPX-2), step-8000:

• feature cosine: 0.890 / 0.854
• velocity readout RMSE (m/s) true vs shuffled commands: 0.061 vs 0.097 / 0.052 vs 0.092 (control signal +58% / +79%)
• far-horizon (3 s) displacement causal vs copy: beats copy in all cells (steady-far 0.34 / 0.49 m vs 0.61-0.78 m)

Files: pred/mlp/vh state dicts + step + mets. See big_run_log.txt for the full run. Full report in the repo (experiments branch molab/big_run_results.md).