action_chunk_size: 51
vla_prepend_rows: 21
trajectory_chunk_size: 30
model:
  action_chunk_size: 51
  learnable_token_count: 0
  use_cognition_token: false
  vlm_feature_layer:
  - 3
  - 6
  - 12
  - 18
  first_frame:
    enable: true
    mode: prepend
    t: relative
  state_encoder:
    use_state_in_llm: false
    use_state_in_act_head: true
    state_norm: false
    set_state_to_zero: false
  extra_encoders:
    fov:
      dim: 2
      enabled: true
      use_norm: false
      concat_mask: false
      inject_to: act_head
      num_tokens: 3
  lora:
    enabled: false
  vl_encoders:
  - name: Qwen35VLEncoder
    model_name_or_path: Alibaba-DAMO-Academy/RynnBrain1.1-2B
    use_system_prompt: true
  act_head:
    state_fov_injection: adaln
    type: flow_matching
    action_feature_length: 1000
    only_use_cognition_token_as_input: false
    position_embedding:
      pe_type: hwt_rope
      rope:
        base: 10000.0
        action_start: 512
        history_t: -1
    flow_matching_config:
      net:
        model_type: MoT
        depth: 24
        hidden_size: 1536
        num_heads: 12
        mlp_ratio: 4.0
        qk_norm: true
        stream_order:
        - condition
        - action
        streams:
          condition:
            class_name: TransformerStream
            args:
              input_norm: false
          action:
            class_name: TransformerStream
            args:
              input_norm: false
              causal: false
        reads_from:
          condition:
          - condition
          action:
          - condition
          - action
        loop_config:
          max_repeats: 1
        lora:
          condition:
            enabled: false
          action:
            enabled: false
      predict_x0: true
      predict_x0_v: false
train_dataset:
  collator:
    vla_prepend_rows: 21
  frame_dataset:
    action_chunk_size: 30
  post_processor:
    instruction_lower: true
  datasets_params:
    common:
      action_chunk_size: 30
      action_space: camera
      state_space: camera
      rotation_type: 6d
      state_rotation_type: null
      translation_action_from_state_type: abs
      rotation_action_from_state_type: abs
      state_include_gripper: true
      view_number: 3
      force_gripper_identity_norm: true
    UnifiedV2EpisodicDataset:
      add_state_gripper: true
      separate_robot_action_features: false
      target_fps: null
      resize_short_edge: 256
      enable_history_image: false
      instruction_prefix: ''
      image_history:
        enable: true
        offsets:
        - -25