Collections

OpenPI checkpoints for wash_dish, organized by model version, demo count, and training step.

HumanoidMimicGen G1 success/failure splits for reward modeling. Group-level 90/10 train/test, no source-demo leakage.