Finetuned from lerobot/pi0_base on collected-ai/lehome-four-types-merged (dual-arm SO101 garment folding; 3 RGB cameras 480x640 + 12-dim joint state; task "fold the garment on the table").

Training: 30k steps, effective batch 64 (bs 32 x 2 GPU DDP), bf16 + gradient checkpointing, pi0 training preset (AdamW lr 2.5e-5, cosine w/ 1k warmup). ~7.2 epochs; final flow-matching loss ~0.013.

Checkpoints (every 5k steps): main is the final step-30000 model; intermediates are branches: step-5000, step-10000, step-15000, step-20000, step-25000.