swordswoman/pi0.5-34bZGrtATgJs

pi0.5 checkpoint for OpenRoboto real-track season 1

Source

• Base checkpoint: 0xjulius/pi05-3cYofFKyNaP2 (OpenRoboto simulation season 1 champion), revision b75495c2b3c4c93e18dc91f2fd24d6a8c1270dbc.
• Fine-tuning: 100 steps of LoRA (backbone rank 16, action expert rank 32) on the LIBERO demonstration data with the base checkpoint's own normalization statistics, adapters merged back into the dense weights.
• Files: model.safetensors (bfloat16, openpi PyTorch layout), config.json, round_info.json (pi05_libero), and assets/physical-intelligence/libero/norm_stats.json (unchanged from the base).

The local pi0.5 checkpoint format check passed. No real-robot performance is claimed.