pi0.5 LoRA fine-tune — Franka pick-and-place on a photo-matched replica of a real lab rig.

- step 3999, full checkpoint: base weights + LoRA factors kept separate (NOT merged)
- 3.4B params, trained on 495 sim demonstrations, joint delta actions + abs gripper
- load with openpi config `pi05_realrig_pp_lora_h32_20k` (H=32), which applies the adapters