GR00T-N1.7 finetuned on super_chatton

Training

• 20,000 steps, batch 64, seed 42, adamw lr 1e-4 cosine (500 warmup), bf16
• chunk_size 16, n_action_steps 16
• use_relative_actions=true, relative_exclude_joints=['gripper']
• embodiment_tag=new_embodiment, policy training preset
• Final loss 0.016 (from 1.06), 17.3 epochs, single H100, 5h49m

Usage