format_version: 1

artifacts:
  policy: policy.pt
  vae: Wan2.2_VAE.pth
  prompt_contexts: prompt_contexts.pt

model:
  class: RgbActionMoT
  dtype: bfloat16
  cross_modal_attention: bidirectional
  video:
    hidden_dim: 3072
    in_dim: 48
    ffn_dim: 14336
    out_dim: 48
    text_dim: 4096
    freq_dim: 256
    eps: 1.0e-6
    patch_size: [1, 2, 2]
    num_heads: 24
    attn_head_dim: 128
    num_layers: 30
    video_attention_mask_mode: bidirectional
  action:
    action_dim: 14
    state_dim: 14
    hidden_dim: 576
    ffn_dim: 2304
    text_dim: 4096
    freq_dim: 256
    eps: 1.0e-6
    num_heads: 24
    attn_head_dim: 128
    num_layers: 30

conditioning:
  image: cam_high_rgb_frame0
  state: raw_qpos_at_frame0
  text: frozen_prompt_context

image:
  input_layout: HWC
  input_dtype: uint8
  training_color_order: rgb
  preprocessing: center_square_crop_then_resize
  resolution: 256
  normalized_range: [-1.0, 1.0]

action:
  horizon: 32
  frequency_hz: 25.0
  dimension: 14
  representation: dual_arm_joint_position_plus_gripper
  layout: left_joint6_gripper_right_joint6_gripper
  semantics: absolute_recorded_joint_position_command
  integration: none
  normalization: piper_joint_action_q01_q99_v1
  implicit_clip: false

diffusion:
  video_train_timesteps: 1000
  video_infer_shift: 5.0
  action_train_timesteps: 1000
  action_infer_shift: 5.0
  inference_steps: 20

tasks:
  assemble_battery_long: Perform the long-horizon battery assembly task.
  battery_assemble: Assemble the battery components.
  pack_3_objects_plus: Pack the three objects into the container.
  stack_3_cups_gen: Stack the three cups.