datasets:
  vla_data:
    action_mode: abs
    data_mix: robotwin
    data_root_dir: /mnt/data/milu/robotwin
    dataset_py: lerobot_datasets
    image_size:
    - 224
    - 224
    interval: 10
    max_step: 5
    memory: true
    per_device_batch_size: 4
    sequential_step_sampling: false
    video_backend: torchvision_av
framework:
  action_model:
    action_dim: 14
    action_hidden_dim: 4096
    action_model_type: L1RegressionActionHead
    future_action_window_size: 7
    past_action_window_size: 0
  name: RynnBrainOFT
  qwenvl:
    base_vlm: playground/Pretrained_models/RynnBrain-CoP-8B
    max_memory_step: 5
    memory: true
output_dir: ./results/Checkpoints/0605_robotwin_RynnBrainOFT
run_id: 0605_robotwin_RynnBrainOFT
run_root_dir: ./results/Checkpoints
seed: 42
trainer:
  eval_interval: 1000
  freeze_modules: true
  gradient_accumulation_steps: 1
  gradient_clipping: 1.0
  is_resume: false
  learning_rate:
    action_model: 0.0001
    base: 1.0e-05
    qwen_vl_interface: 1.0e-05
  logging_frequency: 100
  lr_scheduler_type: cosine_with_min_lr
  max_train_steps: 100000
  num_warmup_steps: 5000
  optimizer:
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  save_interval: 5000
  scheduler_specific_kwargs:
    min_lr: 5.0e-07
wandb_entity: seramasumi-south-china-university-of-technology
wandb_project: starVLA_Robotwin