configs/deploy_policy.yml

# Basic experiment configuration (overridable via eval.sh --overrides)
policy_name: Mem-0
task_name: x # TODO
task_config: demo_clean
ckpt_setting: default
seed: 0
instruction_type: unseen

# Runtime/inference settings
device: cuda:0
camera_key: head_camera
image_size: [224, 224]
execution_ckpt: ./policy/Mem-0/checkpoints/x.pt # TODO
state_stats_path: ./policy/Mem-0/assets/x/norm_stats.json # TODO

planning_module_config_path: ./policy/Mem-0/source/config/planning_module_inference.yaml
vllm_url: http://localhost:8000 # TODO

action_horizon: 30
threshold: 2 # when subtask end signal count reaches this value, we consider the subtask ended and move to the next one
global_task: null

execution_module:
  # Qwen Model
  qwen_vl:
    model_path: ./policy/Mem-0/checkpoints/Qwen3-VL-2B-Instruct
    system_prompt: null
  # Memory Bank
  memory_bank:
    window_size: 30
    initial_anchor_size: 1
    num_heads: 8
    dropout: 0.1
    memory_accumulation: 8
  # DiT Action Model
  action_model:
    action_model_type: DiT-B
    hidden_size: 2048
    add_pos_embed: True
    max_seq_len: 1024
    action_dim: 16
    state_dim: 16
    action_horizon: 30
    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: 8
    diffusion_model_cfg:    # DiT transformer parameters
      cross_attention_dim: 2048 # VLM hidden dim
      dropout: 0.2
      final_dropout: true
      interleave_self_attention: true
      norm_type: "ada_norm"
      num_layers: 16
      output_dim: 2048
      positional_embeddings: null
  # Subtask End Classifier
  classifier:
    hidden_sizes: [6144, 2048, 512]
    dropout: 0.1
    pos_weight: 10.0   # counter class imbalance (false-dominant)
    focal_gamma: 1.0   # optional focal term; set 0.0 to disable
    threshold: 0.5   # probability threshold for positive subtask prediction
  # Loss Weights
  loss_weights:
    lambda_action: 1.0
    lambda_classifier: 0.2