env:
  train:
    env_type: wan_wm
    task_suite_name: libero_spatial
    wm_env_type: libero
    total_num_envs: 128
    auto_reset: false
    ignore_terminations: false
    max_steps_per_rollout_epoch: 240
    max_episode_steps: 240
    use_rel_reward: true
    reward_coef: 5.0
    reset_gripper_open: true
    is_eval: false
    seed: 0
    group_size: 8
    use_fixed_reset_state_ids: true
    use_ordered_reset_state_ids: false
    specific_reset_id: null
    video_cfg:
      save_video: true
      info_on_video: false
      video_base_dir: '@BUNDLE@/runs/reference/video/train'
      record_every_n_episodes: 4
      reward_plot_on_video: true
      reward_plot_show_return: false
      reward_plot_save_metrics: true
      reward_plot_height: 48
      reward_plot_max_points: 30
      fps: 2
    enable_offload: true
    wan_wm_hf_ckpt_path: '@BUNDLE@/config/updated_wan_deployment'
    VAE_path: '@BUNDLE@/assets/wan_baseline/Wan2.2_VAE.pth'
    model_path: '@BUNDLE@/config/updated_wan_deployment/model-00001.safetensors'
    enable_kir: true
    initial_image_path: '@BUNDLE@/assets/wan_baseline/dataset/'
    num_inference_steps: 5
    chunk: 8
    condition_frame_length: 5
    image_size:
    - 256
    - 256
    num_frames: 13
    reward_model:
      type: ResnetRewModel
      from_pretrained: '@BUNDLE@/assets/wan_baseline/resnet_rm.pth'
    shared_wan_pool:
      enabled: true
      target_job_gib: 75.0
      startup_target_job_gib: 65.0
      device_headroom_gib: 3.5
      rpc_timeout_seconds: 1200.0
      rpc_root: ''
      log_dir: ''
    memory_guard:
      enabled: false
      poll_seconds: 5.0
      stable_checks: 2
      foreign_limit_mib: 256.0
      imagination_free_gib: 38.0
      update_free_gib: 12.0
      checkpoint_each_step: true
      retain_imagination_cache: true
    rollout_epoch: 8
    mask_after_first_success: true
    skip_successful_env_generation: true
    fast_frame_transfer: true
    success_reward_threshold: 0.9
    success_consecutive_frames: 1
    state_dim: 8
    use_libero_state_bridge: false
    use_policy_state: false
    debug_io: false
    use_kir_frame_pool: true
    kir_frame_pool_path: '@BUNDLE@/config/kir_pool'
    kir_fraction: 0.2832369942196532
    cosmos_wrist_action_horizon: 8
    use_cosmos_wrist: true
    cosmos_wrist_use_wovr_contract: true
    cosmos_gripper_action_mode: already_cosmos
    cosmos_wrist_primary_reference: true
    cosmos_wrist_input_camera_space: rlinf_runtime_rot180
    cosmos_wrist_guidance: 1.0
    cosmos_use_predicted_proprio: false
    worldpilot_root: '@BUNDLE@/project/third_party/WorldPilot'
    cosmos_host: 127.0.0.1
    cosmos_port: 8055
    cosmos_batch_size: 4
    cosmos_wrist_parallel_generation: false
    cosmos_image_size:
    - 224
    - 224
    cosmos_future_denoising_steps: 5
    cosmos_reward_mode: dense
    cosmos_value_enabled: false
    cosmos_value_host: 127.0.0.1
    cosmos_value_port: 8056
    video_show_shadow_wrist: false
    init_params:
      camera_heights: 256
      camera_widths: 256
    openpi_gripper_action_mode: libero_01_to_signed
    cosmos_wrist_cross_view_primary_history: false
    cosmos_endpoints:
    - host: 127.0.0.1
      port: 18759
    - host: 127.0.0.1
      port: 18760
    - host: 127.0.0.1
      port: 18761
    - host: 127.0.0.1
      port: 18762
    - host: 127.0.0.1
      port: 18763
    - host: 127.0.0.1
      port: 18764
    - host: 127.0.0.1
      port: 18765
    - host: 127.0.0.1
      port: 18766
  eval:
    env_type: libero
    task_suite_name: libero_spatial
    total_num_envs: 64
    auto_reset: true
    ignore_terminations: false
    max_steps_per_rollout_epoch: 240
    max_episode_steps: 240
    use_rel_reward: true
    use_step_penalty: false
    reward_coef: 1.0
    reset_gripper_open: true
    is_eval: true
    seed: 0
    group_size: 1
    use_fixed_reset_state_ids: true
    use_ordered_reset_state_ids: true
    specific_reset_id: null
    video_cfg:
      save_video: true
      info_on_video: false
      video_base_dir: '@BUNDLE@/runs/reference/video/eval'
      reward_plot_on_video: true
      reward_plot_show_return: false
      reward_plot_save_metrics: true
      reward_plot_height: 48
      reward_plot_max_points: 30
      fps: 2
    init_params:
      camera_heights: 256
      camera_widths: 256
    rollout_epoch: 8
    episodes_per_task: 50
    openpi_gripper_action_mode: libero_01_to_signed
  group_name: EnvGroup
actor:
  model:
    model_type: openpi
    model_path: '@BUNDLE@/weights/pi05/global_step_12000'
    precision: null
    num_action_chunks: 8
    action_dim: 7
    is_lora: false
    lora_rank: 32
    use_proprio: true
    num_steps: 3
    add_value_head: false
    openpi:
      config_name: pi05_libero
      num_images_in_input: 2
      noise_level: 0.5
      action_chunk: 8
      num_steps: 3
      train_expert_only: true
      action_env_dim: 7
      noise_method: flow_sde
      add_value_head: false
      value_after_vlm: false
      value_vlm_mode: mean_token
      detach_critic_input: null
      collect_online_nft_state: false
      use_dsrl: false
      dsrl_state_dim: 8
      dsrl_action_noise_dim: 32
      dsrl_num_q_heads: 10
      dsrl_agg_q: mean
      dsrl_image_latent_dim: 64
      dsrl_state_latent_dim: 64
      dsrl_hidden_dims:
      - 128
      - 128
      - 128
      checkpoint_format: native
      action_horizon: 10
      discrete_state_input: false
    openpi_data:
      norm_stats_path: '@BUNDLE@/weights/pi05/global_step_12000/physical-intelligence/libero/norm_stats.json'
  fsdp_config:
    strategy: fsdp
    sharding_strategy: no_shard
    gradient_checkpointing: false
    cpu_offload: false
    offload_pin_memory: false
    reshard_after_forward: true
    enable_gradient_accumulation: true
    forward_prefetch: false
    limit_all_gathers: false
    backward_prefetch: null
    use_orig_params: false
    use_liger_kernel: false
    mixed_precision:
      param_dtype: null
      reduce_dtype: null
      buffer_dtype: null
    amp_autocast:
      enabled: false
      precision: bf16
    grad_scaler:
      enabled: false
    save_full_model_weights: true
    save_training_state_at_steps:
    - 50
    - 60
  group_name: ActorGroup
  shared_wan_pool:
    enabled: true
    target_job_gib: 75.0
    startup_target_job_gib: 65.0
    device_headroom_gib: 3.5
    rpc_timeout_seconds: 1200.0
    rpc_root: ''
    log_dir: ''
  gpu_memory_profile:
    enabled: false
    expected_optimizer_updates: null
  training_backend: fsdp
  micro_batch_size: 32
  global_batch_size: 2048
  seed: 42
  enable_offload: false
  optim:
    lr: 5.0e-06
    adam_beta1: 0.9
    adam_beta2: 0.95
    adam_eps: 1.0e-08
    weight_decay: 0.01
    clip_grad: 1.0
  memory_pool:
    enabled: true
    target_fraction: 0.8
    device_headroom_gib: 12.0
weight_syncer:
  type: patch
  patch:
    snapshot_device: cpu
    delta_encoding: true
    compression: none
    init_sync:
      enabled: true
      prefixes: null
      bucket_size: 134217728
cluster:
  num_nodes: 1
  component_placement:
    actor,env,rollout: 0-7
runner:
  task_type: embodied
  logger:
    log_path: '@BUNDLE@/runs/reference'
    project_name: rlinf
    experiment_name: spatial_followup_primary
    logger_backends:
    - tensorboard
  max_epochs: 60
  max_steps: 60
  only_eval: false
  val_check_interval: 5
  save_interval: 5
  resume_dir: '@BUNDLE@/weights/policy/global_step_40'
  ckpt_path: null
  overlap_env_bootstrap: false
  memory_guard:
    enabled: false
    poll_seconds: 5.0
    stable_checks: 2
    foreign_limit_mib: 256.0
    imagination_free_gib: 38.0
    update_free_gib: 12.0
    checkpoint_each_step: true
    retain_imagination_cache: true
pi05_checkpoint_root: '@BUNDLE@/weights/pi05/global_step_12000'
algorithm:
  normalize_advantages: true
  kl_penalty: kl
  group_size: 8
  reward_type: action_level
  logprob_type: token_level
  entropy_type: token_level
  adv_type: grpo
  loss_type: actor
  loss_agg_func: token-mean
  kl_beta: 0.0
  entropy_bonus: 0.0
  clip_ratio_high: 0.28
  clip_ratio_low: 0.2
  clip_ratio_c: 3.0
  gamma: 0.99
  gae_lambda: 0.95
  reward_coef: 5.0
  filter_rewards: true
  rewards_lower_bound: 0.5
  rewards_upper_bound: 4.5
  update_epoch: 1
rollout:
  sampling_params:
    do_sample: true
    temperature_train: 1.6
    temperature_eval: 1.6
    top_k: -1
    top_p: 1.0
    repetition_penalty: 1.0
  group_name: RolloutGroup
  generation_backend: huggingface
  recompute_logprobs: false
  enable_offload: false
  pipeline_stage_num: 1
  model:
    model_path: '@BUNDLE@/weights/pi05/global_step_12000'
    precision: null
    model_type: openpi
    num_action_chunks: 8
    num_steps: 3
    openpi:
      config_name: pi05_libero
      num_images_in_input: 2
      action_chunk: 8
      action_env_dim: 7
      noise_level: 0.5
      num_steps: 3
      noise_method: flow_sde
      checkpoint_format: native
      action_horizon: 10
      discrete_state_input: false
    openpi_data:
      norm_stats_path: '@BUNDLE@/weights/pi05/global_step_12000/physical-intelligence/libero/norm_stats.json'
    is_lora: false
reward:
  use_reward_model: false
critic:
  use_critic_model: false