madokalif/pi05-robotwin2-clean50-sft

pi0.5 RoboTwin 2.0 clean50 50-task SFT

Full fine-tune of OpenPI pi0.5 on the RoboTwin 2.0 clean50 50-task aloha suite.

• Base: motus-robotics/pi0.5_robotwin2 (pi0.5, aloha 14-dim state/action)
• Training: full fine-tune (PyTorch/DDP), global batch 64, 20,000 steps, cosine LR
• Final train loss ~0.0017 (flow-matching velocity)
• Config: pi05_robotwin2_clean50_full / exp rtw_sft_full_v1

Files

• model.safetensors — 812-key pi0.5 weights (drop-in for the openpi/RoboTwin pi0.5 eval server)
• assets/robotwin2/aloha_clean50_50task/norm_stats.json — state/action norm stats (14-dim)
• metadata.pt — training metadata

optimizer.pt is intentionally omitted (inference does not need it).