pi05_obs_only_first3mid3_last6_injection_allenv_2ep

2-epoch JAX/orbax fine-tune of RoboPRO obstacle-attention π₀.₅ (pi05_obs_only_first3mid3_last6_injection_allenv), obstacle-only attention (target and destination attention off). Supervises first-3 plus middle-3 layers (0,1,2,7,8,9) and injects into last-6 (12,13,14,15,16,17).

• 230934 steps (2 × 115467), batch 32, 2-GPU data parallel
• Saves every 25k steps plus epoch ends (115467, 230934)
• Init: mzxuan/robopro_jax_30000

Each step directory contains everything needed to resume training or run eval:

• params/ — model weights
• train_state/ — optimizer / train state
• assets/ — norm_stats.json
• _CHECKPOINT_METADATA — orbax item handlers (params, train_state, assets)
• train_config.py — TrainConfig snippet to paste into src/openpi/training/config.py (same pattern as mzxuan/robopro_jax_30000)