seed: 42
device: cuda:0
n_epochs: 50
batch_size: 16
gradient_accumulation_steps: 1
mixed_precision: 'no'
learning_rate: 8.0e-05
lr_scheduler: constant
ent_coef: 0.01
l2_coef: 0.01
weight_decay: 0
max_grad_norm: 1.0
agent_class: sl_agent.SL
policy: ARMYPolicy
policy_kwargs:
  dim_decoder: 512
  n_decoder_layers: 4
  n_decoder_heads: 8
  use_rmsnorm: true
  pre_norm: true
  sa_att: true
  ca_att: true
  army_att: false
  army_residual: false
  army_norm: math
  dropout: 0.0
  dropout_residual: 0.0
  action_horizon: 16
  optimizer_kwargs:
    eps: 1.0e-08
  noise: false
preprocessor_class: Lerobot_2_Army
preprocessor_kwargs:
  resize_images: !!python/tuple
  - 94
  - 94
  normalize_images: true
  tokenizer_kwargs:
    model_template: qwen2-chat
    num_image_token: 9
    padding_max_length: 60
    padding_side: right
  prompt_generator: FixedPromptGenerator("Insert the peg into the hole.")
  observation_modes:
    observation.state: min_max
  action_modes:
    action: min_max
  overwrite_stats: false
  stats: null
  data_key_map:
    observation.images.base_camera: cam1
    observation.images.hand_camera: cam2
    observation.images.head_camera: cam2
    observation.images.image_side_1: cam1
    observation.images.image_side_2: cam2
    observation.images.image_wrist_1: cam3
    observation.images.image_wrist_2: cam4
    observation.top: cam1
    image: cam1
    wrist_image: cam2
    observation.state: state
    state: state
    action: action
    actions: action
  task: PegInsertionSide-v1
env_cfg:
  name: maniskill
  task: PegInsertionSide-v1
  dataset_repo_id: johnMinelli/ManiSkill_PegInsertionSide-v1_recovery
  episodes: null
  dataset_revision: v2.0
  dataset_root: null
  video_backend: pyav
  input_shapes:
    observation.images.base_camera: !!python/tuple
    - 480
    - 640
    - 3
    observation.images.hand_camera: !!python/tuple
    - 480
    - 640
    - 3
    observation.state: !!python/tuple
    - 9
    observation.privileged: !!python/tuple
    - 27
  output_shapes:
    action: !!python/tuple
    - 16
    - 8
  normalization:
    observation_modes:
      observation.state: min_max
    action_modes:
      action: min_max
  fps: 20
  delta_timestamps:
    expert_mask: '[i / 20 for i in range(16)]'
    observation.images.base_camera: '[i / 20 for i in range(16)]'
    observation.images.hand_camera: '[i / 20 for i in range(16)]'
    observation.state: '[i / 20 for i in range(16)]'
    action: '[i / 20 for i in range(16)]'
  image_transforms:
    enable: false