RoboSyn-pi05-weak4

pi0.5 (openpi) fine-tuned on four RoboSynChallenge simulation tasks (sample_loading, items_handover, handle_basket, click_bell; official released sim datasets, 1000 episodes each), 12k steps from pi05_base. Inference-only checkpoint: bf16 params + norm stats (assets/). Train config: pi05_robosyn_all10 in https://github.com/QuixoteH/RoboSyn (policy/pi05/src/openpi/training/config.py); used through policy/robosyn_mix (team LBWSquare, RoboSynChallenge NeurIPS 2026).