run_id: 0617_robotwin32_5task_qwenpiv4_la_unit_detach
run_root_dir: ./results/Checkpoints2
seed: 42
wandb_project: starVLA_Robotwin
is_debug: false
version_id: '0.21'
framework:
  name: QwenPI_v4_LA
  qwenvl:
    base_vlm: playground/Pretrained_models/Qwen/Qwen3-VL-2B-Instruct
    attn_implementation: flash_attention_2
    vl_hidden_dim: 2048
    num_vl_layers: 28
  action_model:
    action_dim: 14
    state_dim: 14
    action_horizon: 32
    add_pos_embed: true
    max_seq_len: 1024
    num_target_vision_tokens: 0
    repeated_diffusion_steps: 8
    noise_beta_alpha: 1.5
    noise_beta_beta: 1.0
    noise_s: 0.999
    num_timestep_buckets: 1000
    num_inference_timesteps: 20
    diffusion_model_cfg:
      input_embedding_dim: 1024
      attention_head_dim: 32
      num_attention_heads: 32
      norm_type: norm
      dropout: 0.2
      final_dropout: true
      interleave_self_attention: true
      positional_embeddings: null
      cross_attention_dim: 2048
      num_layers: 28
      output_dim: 14
    future_action_window_size: 31
    past_action_window_size: 0
  latent_action:
    enabled: true
    backend: unit
    loss_type: ce
    train_latent: true
    train_action: true
    latent_loss_weight: 0.1
    action_loss_weight: 1.0
    detach_vl_embs_for_action_head: true
    action_train_robot_types:
    - aloha-agilex
    num_bridge_tokens: 8
    num_codebooks: 2
    codebook_size: 128
    label_smoothing: 0.0
    ce_loss_weights:
    - 1.0
    - 0.5
    strict_num_bridge_tokens: true
    groot_tokenizer_path: /inspire/qb-ilm/project/qproject-fundationmodel/public/zzt/starVLA/third_party/UniT/outputs/robotwin_tokenizer/checkpoint-80000
    dinov2_path_override: /inspire/qb-ilm/project/qproject-fundationmodel/public/zzt/models/dinov2-large
    image_size:
    - 224
    - 224
    qwenpi_v4_la:
      token_format: <latent_action_{i}>
      head_dropout: 0.0
    univla:
      ckpt_path: null
      model_dim: 768
      latent_dim: 128
      patch_size: 14
      enc_blocks: 12
      dec_blocks: 12
      num_heads: 12
      dropout: 0.0
      image_size:
      - 224
      - 224
      strict_load: true
    softvq:
      config_path: starVLA/model/modules/latent_action/softvq_joint/0614_joint_aloha_xrecon_scratch.yaml
      ckpt_path: /inspire/qb-ilm/project/qproject-fundationmodel/public/jjc/exp/softla/0614_joint_aloha_xrecon_scratch/checkpoints/partial_step_100000.pt
      image_size:
      - 224
      - 224
      strict_load: true
      kl_eps: 1.0e-08
    villax:
      ckpt_path: null
      image_size:
      - 224
      - 224
      num_learned_tokens: 4
datasets:
  vla_data:
    dataset_py: lerobot_la_datasets
    latent_action:
      enabled: true
      stride: 16
      horizon: 32
      video_key: cam_high
      include_terminal_frame: true
      image_size:
      - 224
      - 224
    include_state: true
    data_root_dir: /inspire/qb-ilm/project/qproject-fundationmodel/public/zzt/data/RoboTwin2.0/lerobot
    data_mix: robotwin32_5task
    action_type: abs_qpos
    per_device_batch_size: 16
    load_all_data_for_training: true
    obs_image_size:
    - 256
    - 256
    video_backend: torchcodec
trainer:
  max_train_steps: 50000
  num_warmup_steps: 1000
  save_interval: 5000
  eval_interval: 500
  eval_num_samples: 512
  eval_batch_size: 16
  logging_frequency: 50
  learning_rate:
    base: 5.0e-05
  lr_scheduler_type: cosine_with_min_lr
  scheduler_specific_kwargs:
    min_lr: 5.0e-06
  freeze_modules: null
  loss_scale:
    vla: 1.0
  max_grad_norm: 1.0
  gradient_clipping: 1.0
  gradient_accumulation_steps: 2
  optimizer:
    name: AdamW
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    weight_decay: 1.0e-08
config_yaml: ./examples/Robotwin/new_train/starvla_qwenpiv4_la_unit_robotwin.yaml
output_dir: ./results/Checkpoints2/0617_robotwin32_5task_qwenpiv4_la_unit_detach