89 per-task distilled-steering BC policies, one .pt per LIBERO-90 task (tasks with at least one successful demo). Each predicts a 7-D chunk-averaged initial flow noise for the frozen pi0.5-LIBERO VLA from 84x84 base+wrist images (DSRL-style CNN encoders, latent 50, MLP 128x3, Gaussian head, mean = 5*tanh) — the offline DSBC protocol of Flow Reversal Steering (FRS, arXiv:2606.13675, App. E.2.2): 12,500 steps, batch 128, Adam lr 1e-4, NLL loss.

Trained on knightnemo/libero90-pi05-frs-noise-labels. At deployment, repeat the 7-D prediction across the chunk horizon (10) and pad to the 32-D model action space with i.i.d. N(0,1), then pass as initial noise to pi0.5-LIBERO. Checkpoint format: {"policy": state_dict, "noise_dim": 7}; see scripts/libero/eval_libero.py in VAM_Learn_from_Human_Video.