# @package _global_
defaults:
  - train
  - override /data: robotwin
  - override /model: w0
  - _self_

posttrain_required: true
posttrain_stats: ${oc.env:W0_ROBOTWIN_STATS,assets/robotwin/normalization.json}
init_checkpoint: ${oc.env:W0_ROBOTWIN_CHECKPOINT,checkpoints/robotwin-clean.pt}
resume: null
seed: 42
model:
  proprio_dim: 37
  action_horizon: 64
  action_chunk_size: 16
  max_action_offset: 0
  num_history_frames: 0
  video_plan_horizon: 96
  video_action_stride: 8
  action_dit_config:
    action_dim: 37
    num_soft_prompts: 8
    soft_prompt_domains: [robotwin_wam/aloha]
    domain_configs:
      - name: robotwin_wam/aloha
        dataset: robotwin_wam
        embodiment: aloha
        action_dim: 37
        proprio_dim: 37
        action_valid_indices: [0, 1, 2, 3, 4, 5, 16, 17, 18, 19, 20, 21, 22, 33]
        proprio_valid_indices: [0, 1, 2, 3, 4, 5, 16, 17, 18, 19, 20, 21, 22, 33]
        raw_joint_names: [left_waist, left_shoulder, left_elbow, left_forearm_roll, left_wrist_angle, left_wrist_rotate, left_gripper, right_waist, right_shoulder, right_elbow, right_forearm_roll, right_wrist_angle, right_wrist_rotate, right_gripper]
        action_semantics: native_absolute_joint_command_interval_end

batch_size: 12
gradient_accumulation_steps: 1
learning_rate: 3.0e-5
soft_prompt_lr_scale: 1.0
num_epochs: 15
num_workers: 4
history_aware_batching: false
optimizer_group_max_numel: 200000000
save_every: 2500
eval_every: 500
log_every: 10
keep_training_states: 1
checkpoint_min_free_gib: 128
freeze_video_dit: false
output_dir: ${oc.env:W0_OUTPUT_DIR,outputs}/robotwin/${now:%Y%m%d_%H%M%S}
wandb:
  name: internw0_robotwin_clean_c2r