# SO-101 PickOrange — StarVLA QwenGR00T (Qwen3-VL-4B vision + GR00T flow-matching head)
# LeIsaac SO-101 single-arm, 6-DOF joints, LeRobot v2.1 (read via v2.0 path).
# Run 1 recipe: freeze the whole VLM, train only the action head (GR00T-N1.7-style,
# fits 32G, leverages pretrained Qwen3-VL vision as-is = the "vision dividend").
run_id: so101_pickorange_qwengr00t
run_root_dir: /root/autodl-tmp/starvla-outputs
seed: 42
wandb_entity: ""
wandb_project: starVLA_SO101
is_debug: false
version_id: "0.21"

framework:
  name: QwenGR00T
  qwenvl:
    base_vlm: /root/autodl-tmp/models/Qwen3-VL-4B-Instruct
    attn_implementation: flash_attention_2
    vl_hidden_dim: 2048
  action_model:
    action_dim: 6
    state_dim: 6
    action_horizon: 16

datasets:
  vla_data:
    dataset_py: lerobot_datasets
    data_root_dir: /root/autodl-tmp/datasets
    data_mix: so101_pickorange
    lerobot_version: v2.0          # our dataset is v2.1 -> read via v2.0 file layout
    sequential_step_sampling: false
    per_device_batch_size: 8
    load_all_data_for_training: true
    video_backend: torchvision_av

trainer:
  max_train_steps: 30000
  num_warmup_steps: 1500
  save_interval: 3000
  eval_interval: 100000            # no in-train sim eval here -> effectively off
  learning_rate:
    base: 2.5e-05
    qwen_vl_interface: 1.0e-05
    action_model: 1.0e-04
  lr_scheduler_type: cosine_with_min_lr
  scheduler_specific_kwargs:
    min_lr: 1.0e-06
  freeze_modules: 'qwen_vl_interface'
  loss_scale:
    vla: 1.0
    vlm: 0.1
  max_grad_norm: 1.0
  weight_decay: 0.0
  logging_frequency: 10
  gradient_clipping: 1.0
  gradient_accumulation_steps: 1   # must match ds_config.yaml (deepspeed hardcodes 1)
  gradient_checkpointing: true
  optimizer:
    name: AdamW
    betas: [0.9, 0.95]
    eps: 1.0e-08
    weight_decay: 1.0e-08