RadioFlow: the stats each checkpoint was trained with; dims checked for all 104, values checked against the original training file for 64. ClothCylinderWipe uses its floored policy stats. LiftSoftDex/RubikCube/RopeBridgeDex from their training runs; PlugCharger recomputed from raw. PI0.5: openpi assets of each run (identical to the copy inside the training checkpoint).
Upload DualArmOpenScissor-v1/ckpt_step_40000.pt with huggingface_hub
