RoboColosseum/droid-close-laptop-molmoact2

MolmoAct2 — DROID CloseLaptop (10 epochs)

• Task: "close laptop" — 100 episodes, 27,649 frames @ 30 fps (aarontung/droid-close-laptop)
• Inputs/outputs: cameras left exterior / right exterior / wrist + instruction; 8-D absolute joint state and action (7 Franka joints + gripper)
• Base: allenai/MolmoAct2 (base)
• Trained: Full fine-tune (VLM + action expert)
• Optim: LLM LR 2e-5, ViT/connector 1e-5, action expert 1e-4 (same as BimanualYAM), warmup, decay to 0.1x; GBS 64 (32x2 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 5h08m (the 5 extra epochs)
• Files: Final checkpoint converted to HF format (model-*.safetensors, processor, norm_stats.json).
• Load: HF trust_remote_code model; see inference.py in this folder
• Code: RoboColosseum harness (scripts/train/molmoact2.sh); logs in W&B project RoboColosseum
• Continued from: RoboColosseum/DROID-models/molmoact2/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.