add mapfly_agent_oft_p1r0/config.yaml

mapfly_agent_oft_p1r0/config.yaml

@@ -0,0 +1,63 @@
+datasets:
+  vla_data:
+    CoT_prompt: Use both images. The top-down map is north-up and stays fixed; it
+      does not rotate with you. The blue marker is your current position and the red
+      marker is the goal. Your proprioceptive state is the pose relative to the episode
+      start in the start-body frame, [dx, dy, dz, dyaw]; use dyaw with this fixed
+      map to infer which way you are currently facing. Turn toward the goal, do not
+      fly past it, and stop only when you are already next to it. Use the first-person
+      view to keep the path ahead clear. {instruction}
+    data_mix: uav_mapfly_goalgeo_seen12_tau05
+    data_root_dir: ./playground/Datasets/uav
+    dataset_py: lerobot_datasets
+    include_state: true
+    obs_image_size:
+    - 224
+    - 224
+    per_device_batch_size: 4
+    video_backend: torchvision_av
+framework:
+  action_model:
+    action_dim: 4
+    action_hidden_dim: 2560
+    action_horizon: 8
+    action_model_type: MLP
+  goal_head:
+    hidden_dim: 512
+    stop_bias_init: -2.8
+    stop_pos_weight: 10.0
+    stop_radius_m: 8.0
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: Qwen/Qwen3-VL-4B-Instruct
+output_dir: ./playground/Checkpoints/mapfly_agent_oft_p1r0
+run_id: mapfly_agent_oft_p1r0
+run_root_dir: ./playground/Checkpoints
+seed: 42
+trainer:
+  eval_interval: 10000
+  freeze_modules: ''
+  gradient_clipping: 1.0
+  learning_rate:
+    action_model: 0.0001
+    base: 2.5e-05
+    goal_head: 0.0001
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 10
+  loss_weight:
+    progress: 0.5
+    stop: 0.05
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 80000
+  num_warmup_steps: 4000
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+  save_interval: 10000
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_project: starVLA_uav