pi05_yubi_ee_pickplace_reviewed

pi05 on the human-reviewed pick-and-place set — the cleanest pickplace checkpoint. Recommended pickplace model (supersedes pi05_yubi_ee_pickplace and _v2).

• Data: OmakaseAI/yubi_ee_pickplace_v2 train split, 651 episodes (708 minus 57 excluded = 45 manually excluded in the omakase-annotation gallery [failed grasps / non-pickplace / bad framing] + 14 with a frozen right-wrist tail [stream cut out over the place/release]). Applied via --dataset.episodes subset (exact, no rebuild). Exclusion list on 6000pro ~/pickplace_review_exclusions.json + the 14 frozen indices.
• Actions: ABSOLUTE (relative=true + all-20-dim exclude). 20-dim bimanual EE, 3 cams, 7500 steps, full-param (vision frozen), 8x H100.
• wandb: omakase-robotics/galaxea_pi05_ee/runs/iddr6wtc

Caveats: right-hand-only collection (left-arm dims idle); base_0_rgb is a chest fisheye. Open-loop MSE != closed-loop success — on-robot eval required. Serve via galaxea_lerobot serve_pi05_yubi_ee_bimanual.sh (MODEL_ID/DATASET_ID = this pair); client pi05_ee_inference.

Open-loop eval (clean disjoint holdout)

OmakaseAI/yubi_ee_pickplace_v2_holdout (82 eps, disjoint from the 651 train), 40 samples, seed 0, normalized:

metric	value	gate ≤0.25
first-step MAE	0.060	✅
overall MAE	0.128
final-step MAE	0.146

Far better than the full-episode bimanual A/B (overall ~0.246 / first ~0.215) — pickplace-specialized + human-reviewed. (Holdout itself is un-reviewed, so a few noisy frames are included; true quality is likely slightly better.) Open-loop MSE is not closed-loop success.