knightnemo/libero90-pi05-frs-noise-labels

LIBERO-90 flow-reversal noise labels for pi0.5 (FRS-style offline DSBC)

Per-frame latent-noise labels for the success-filtered LIBERO-90 demos (Embodied-CoT/embodied_features_and_demos_libero), computed by reverse-integrating the frozen pi0.5-LIBERO flow ODE (plain reverse Euler, 10 steps, mirroring openpi's forward discretization) at each demo action chunk — the offline labeling step of Flow Reversal Steering (FRS)-style distilled noise-space BC.

Files

One .npz per task ({task_slug}.npz):

tasks.json maps task slugs to language instructions; summary.json has global stats. Labels were produced on 12 parallel GPU workers and merged per task.

Upstream data note: 3 of the 3,917 source trajectories (train indices 578-580, inside libero_lm_90_openpi-train.tfrecord-00018-of-00128) are unreadable — the shard is corrupted at byte 107834118 in the upstream HF repository itself (our mirror's sha256 matches HF's linked etag), so those episodes are excluded; 3,914 trajectories are labeled.

Task accounting: LIBERO-90 has 90 tasks; the success filter leaves 89 with at least one successful demo (FRS's trainable set, 3,917 trajectories in total). Shards are grouped by LIBERO task identity (scene + task name, from episode_metadata/file_path), giving one shard set per trainable task — matching FRS's one-policy-per-task protocol. Note 16 task pairs share a language instruction across scenes; the slug keeps them separate.

Provenance / reproduction

Generated with scripts/libero/label_noises.py from VAM_Learn_from_Human_Video (branch feat/efficient-ft): pi05_libero checkpoint (JAX -> PyTorch via openpi's converter), quantile action normalization from the checkpoint's own norm stats, actions normalized before zero-padding to the 32-dim model space.