datasets:
  vla_data:
    CoT_prompt: Your task is {instruction}. To identify the key objects for your task.
      Locate their bounding boxes in [x1,y1,x2,y2] format.
    action_chunk_offset: 1
    action_chunk_size: 16
    action_dim: 14
    action_mode: abs
    current_image_color_jitter:
      brightness: 0.3
      contrast: 0.4
      hue: 0.08
      saturation: 0.5
    data_mix: xpolicylab
    data_root_dir: /mnt/project/world_model/data/sim_benchmark
    dataset_py: lerobot_datasets
    delete_pause_frame: false
    image_only_mode: true
    img_interval: 1
    jitter_current_image_only: true
    num_chunk: 1
    per_device_batch_size: 6
    state_dim: 14
    stats_filename: meta/stats.json
    training_task_weights:
    - 1.0
    training_tasks:
    - policy
    use_dummy_dataset: 0
    val_ratio: 0.0
    video_chunk_offset: 16
    video_chunk_size: 1
    vlm_block_condition_mode: causal_shared
framework:
  action_model:
    action_dim: 14
    action_horizon: 16
    action_model_type: DiT-L
    add_pos_embed: true
    diffusion_model_cfg:
      cross_attention_dim: 2560
      dropout: 0.0
      final_dropout: false
      interleave_self_attention: true
      norm_type: ada_norm
      num_layers: 16
      output_dim: 2560
      positional_embeddings: learnable
    future_action_window_size: 15
    future_obs_index: 16
    hidden_size: 2560
    logit_mean: 0.0
    logit_std: 1.0
    max_num_embodiments: 1
    max_seq_len: 4096
    noise_beta_alpha: 1.5
    noise_beta_beta: 1.0
    noise_s: 0.999
    num_inference_timesteps: 4
    num_target_vision_tokens: 32
    num_timestep_buckets: 1000
    num_views: 1
    obs_loss_weight: 1.0
    only_policy: false
    only_wo_video_gen: false
    past_action_window_size: 0
    policy_and_video_gen: false
    state_dim: 14
    use_attn_mask: true
    vision_encoder_path: /mnt/home/liukai/World-Action-Model/pretrained
    vision_encoder_size: s
    vision_encoder_type: dinov3
    visual_token_fusion_mode: direct_tokens
  name: QwenMMDiT
  qwenvl:
    base_vlm: /mnt/home/liukai/starVLA/playground/pretrained/vlm/Qwen3-VL-4B-Instruct
    pad_to: 16
output_dir: /workplace/world_model/checkpoints/lda_v2/post-train/robodojo/RoboDojo-cotrain-robodojo-multiview-0
run_id: RoboDojo-cotrain-robodojo-multiview-0
run_root_dir: /workplace/world_model/checkpoints/lda_v2/post-train/robodojo
seed: 0
swanlab_mode: cloud
swanlab_project: xpolicy_lab
swanlab_workspace: Onway
trackers:
- jsonl
- swanlab
trainer:
  enable_gradient_checkpointing: false
  eval_interval: 5000
  freeze_modules: action_model.vision_encoder
  gradient_warn: 100.0
  is_resume: false
  learning_rate:
    action_model: 0.0001
    base: 4.0e-05
    qwen_vl_interface: 1.0e-05
  lerobot_eval:
    action_horizon: 16
    enabled: true
    end_traj: null
    max_eval_steps: 720
    start_traj: 0
    traj_count: 1
  logging_frequency: 100
  lr_scheduler_type: cosine_with_min_lr
  max_checkpoints: null
  max_train_steps: 160000
  num_warmup_steps: 5000
  optimizer:
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
  pretrained_checkpoint: null
  repeated_diffusion_steps: 2
  save_interval: 5000
  scheduler_specific_kwargs:
    min_lr: 5.0e-07
  torch_compile: 0
wandb_project: LDA