RoboColosseum/droid-close-laptop-pi0.5

π0.5 — DROID CloseLaptop (10 epochs)

• Task: "close laptop" — 100 episodes, 27,649 frames @ 30 fps (aarontung/droid-close-laptop)
• Inputs/outputs: cameras left exterior / wrist + instruction; 8-D absolute joint state and action (7 Franka joints + gripper)
• Base: gs://openpi-assets/checkpoints/pi05_base (openpi, JAX)
• Trained: Full fine-tune, EMA 0.99 (config pi05_droid_joint, openpi DroidInputs, joints as deltas)
• Optim: AdamW, cosine, peak LR 2.5e-5 → 2.5e-6 (same as BimanualYAM), 1k warmup; GBS 32 (32x1 micro-batch x accumulation, 1 GPU); 10 epochs total: 5 in the DROID-models run + 5 more here; fp32 master weights, bf16 compute; 1x H200; wall time 1h47m (the 5 extra epochs)
• Files: Final checkpoint in openpi format: params/ (EMA weights) and assets/<dataset>/norm_stats.json. Optimizer state not included.
• Load: openpi.policies.policy_config.create_trained_policy(get_config("pi05_droid_joint"), <this folder>) (config: configs/openpi/yam.py (PI05_DROID))
• Code: RoboColosseum harness (scripts/train/pi05.sh); logs in W&B project RoboColosseum
• Continued from: RoboColosseum/DROID-models/pi05/closelaptop (5 epochs). The 5 extra epochs restart the same schedule (same peak LR, warmup and decay) from those final weights with a fresh optimizer, at the same global batch size; logged to the same W&B run.