Consolidate paper experiment assets by task and run (301-450)

latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/checkpoints/best_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:630d55bc258d986e0de4b04d9223f418855bb94c2d310c13ae808912553d6c7f
+size 9138230516

latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/checkpoints/steps_500_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad5f732f19bbfda68d6686e39e875f07957d08980bf6325dea7f85eccb4abc70
+size 9138230516

latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/config.full.yaml

@@ -0,0 +1,191 @@
+framework:
+  name: QwenOFT
+  qwenvl:
+    base_vlm: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    action_hidden_dim: 2560
+    future_action_window_size: 0
+    past_action_window_size: 0
+    loss_type: discrete_ce
+    state_dim: 7
+    action_horizon: 1
+    action_env_dim: 7
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2
+    data_mix: deadly_corridor_train__bridge
+    eval_data_mix: deadly_corridor_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  converted_name: deadly_corridor_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: deadly_corridor_train
+  mixed_converted_name: deadly_corridor_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+initialization:
+  checkpoint_local_dir: /inspire/hdd/project/spatialintelligence/public/lzj/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 500
+  num_warmup_steps: 100
+  save_interval: 50
+  eval_interval: 100
+  eval_num_batches: 20
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 16
+  distributed_backend: none
+  is_resume: false
+  pretrained_checkpoint: null
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: /inspire/hdd/project/spatialintelligence/public/lzj/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 4
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: You are playing Deadly Corridor in VizDoom. Choose actions from
+      MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.
+    deadly:
+      action_layout: multibinary_7
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2/deadly_corridor_train__bridge/latency_prompt_map.json
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: true
+      interval_steps: 50
+      latencies:
+      - 2
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: true
+      latencies:
+      - 0
+      - 2
+      - 4
+      - 6
+      - 8
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  task: deadly_corridor
+  deadly_corridor_loss_type: null
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: deadly_corridor
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: /inspire/hdd/project/spatialintelligence/public/lzj/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: true
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: openvla_deadly_corridor_fix_latency_2
+output_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/results/Checkpoints/openvla_deadly_corridor_fix_latency_2
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/config.yaml

@@ -0,0 +1,92 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: true
+  save_pt_file: false
+  sync:
+    enabled: false
+    keep_last_n: 0
+    repo_id: null
+datasets:
+  vla_data:
+    data_mix: deadly_corridor_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: deadly_corridor_train__bridge__val
+    include_state: true
+    obs_image_size: null
+    per_device_batch_size: 32
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 7
+    action_hidden_dim: 2560
+    action_horizon: 1
+    action_model_type: MLP
+    loss_type: discrete_ce
+    state_dim: 7
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+output_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/results/Checkpoints/openvla_deadly_corridor_fix_latency_2
+rl_games:
+  env_eval:
+    deadly:
+      action_layout: multibinary_7
+    distributed_mode: none
+    enabled: true
+    fixed_episode_seeds: true
+    frameskip: 4
+    image_size: 224
+    latency:
+      prompt_map_path: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2/deadly_corridor_train__bridge/latency_prompt_map.json
+    latency_seed_stride: 0
+    mid_train:
+      enabled: true
+      interval_steps: 50
+      latencies:
+      - 2
+      max_steps_per_episode: 3600
+      num_episodes: 5
+    seed: 42
+    task_description: You are playing Deadly Corridor in VizDoom. Choose actions from
+      MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.
+    task_seed_stride: 0
+    vectorized:
+      enabled: false
+  model_alias: openvla
+  task: deadly_corridor
+run_id: openvla_deadly_corridor_fix_latency_2
+run_root_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/results/Checkpoints
+seed: 42
+trainer:
+  distributed_backend: none
+  eval_interval: 100
+  eval_num_batches: 20
+  freeze_modules: ''
+  gradient_accumulation_steps: 16
+  gradient_clipping: 1.0
+  is_resume: false
+  learning_rate:
+    action_model: 0.0001
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 500
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: false
+    weight_decay: 1.0e-08
+  pretrained_checkpoint: null
+  reload_modules: null
+  save_interval: 50
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games

latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.9800952076911926,
+        0.00043271310278214514,
+        0.04197317361831665,
+        0.853310227394104,
+        0.33102554082870483,
+        0.3102552890777588,
+        0.8277801871299744
+      ],
+      "std": [
+        0.13967378437519073,
+        0.02079719677567482,
+        0.2005293071269989,
+        0.3537939190864563,
+        0.47058752179145813,
+        0.4625938832759857,
+        0.3775743842124939
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        0.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 2311,
+    "num_trajectories": 90
+  }
+}

latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.9674796462059021,
+        0.0,
+        0.05284553021192551,
+        0.8211382031440735,
+        0.3211382031440735,
+        0.32926830649375916,
+        0.8373983502388
+      ],
+      "std": [
+        0.17737756669521332,
+        0.0,
+        0.22372478246688843,
+        0.38323649764060974,
+        0.4669141173362732,
+        0.46994760632514954,
+        0.36900219321250916
+      ],
+      "max": [
+        1.0,
+        0.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        0.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 246,
+    "num_trajectories": 10
+  }
+}

latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/hydra/.hydra/config.yaml

@@ -0,0 +1,189 @@
+framework:
+  qwenvl:
+    base_vlm: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    state_dim: 7
+    loss_type: discrete_ce
+    action_horizon: 1
+    future_action_window_size: 0
+    past_action_window_size: 0
+    action_dim: 7
+    action_env_dim: 7
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2
+    data_mix: deadly_corridor_train__bridge
+    eval_data_mix: deadly_corridor_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ${dataset.single_source_hf}
+  converted_name: ${dataset.single_converted_name}
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: deadly_corridor_train
+  mixed_converted_name: deadly_corridor_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: ${dataset.single_latency_filter}
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: ${bridge_base_model.repo_id.${model}}
+initialization:
+  checkpoint_local_dir: /inspire/hdd/project/spatialintelligence/public/lzj/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 500
+  num_warmup_steps: 100
+  save_interval: 50
+  eval_interval: 100
+  eval_num_batches: 20
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 16
+  distributed_backend: none
+  is_resume: false
+  pretrained_checkpoint: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: ${bridge_base_model.local_dir.${model}}
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 4
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: You are playing Deadly Corridor in VizDoom. Choose actions from
+      MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.
+    deadly:
+      action_layout: multibinary_7
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2/deadly_corridor_train__bridge/latency_prompt_map.json
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: true
+      interval_steps: 50
+      latencies:
+      - 2
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: true
+      latencies:
+      - 0
+      - 2
+      - 4
+      - 6
+      - 8
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  task: deadly_corridor
+  deadly_corridor_loss_type: null
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: deadly_corridor
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: /inspire/hdd/project/spatialintelligence/public/lzj/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: true
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: openvla_deadly_corridor_fix_latency_2
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/hydra/.hydra/hydra.yaml

@@ -0,0 +1,252 @@
+hydra:
+  run:
+    dir: ${run_root_dir}/${run_id}/hydra
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - model=openvla
+    - env=deadly_corridor
+    - init=bridge
+    - mode=single
+    - run_id=openvla_deadly_corridor_fix_latency_2
+    - run_root_dir=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/results/Checkpoints
+    - seed=42
+    - wandb_entity=dongqianyu99-zhejiang-university
+    - wandb_project=starVLA_rl_games
+    - rl_games.env_eval.enabled=true
+    - checkpoint.sync.enabled=false
+    - checkpoint.sync.keep_last_n=0
+    - checkpoint.local.keep_last_n=1
+    - checkpoint.save_best_model=true
+    - checkpoint.save_pt_file=false
+    - trainer.is_resume=false
+    - trainer.pretrained_checkpoint=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+    - trainer.resume_step=0
+    - trainer.max_train_steps=500
+    - trainer.num_warmup_steps=100
+    - trainer.save_interval=50
+    - trainer.eval_interval=100
+    - trainer.eval_num_batches=20
+    - trainer.logging_frequency=1
+    - trainer.gradient_accumulation_steps=16
+    - trainer.distributed_backend=none
+    - trainer.learning_rate.base=2e-05
+    - trainer.learning_rate.qwen_vl_interface=1e-05
+    - trainer.learning_rate.action_model=0.0001
+    - trainer.lr_scheduler_type=cosine_with_min_lr
+    - trainer.scheduler_specific_kwargs.min_lr=1e-06
+    - trainer.loss_scale.vla=1.0
+    - trainer.loss_scale.vlm=0.1
+    - trainer.max_grad_norm=1.0
+    - trainer.weight_decay=0.0
+    - trainer.gradient_clipping=1.0
+    - trainer.optimizer.name=AdamW
+    - trainer.optimizer.betas=[0.9, 0.95]
+    - trainer.optimizer.eps=1e-08
+    - trainer.optimizer.weight_decay=1e-08
+    - trainer.optimizer.fused=false
+    - trainer.save_format=pt
+    - framework.name=QwenOFT
+    - framework.qwenvl.attn_implementation=flash_attention_2
+    - framework.qwenvl.enable_gradient_checkpointing=true
+    - framework.action_model.action_dim=7
+    - framework.action_model.action_env_dim=7
+    - framework.action_model.state_dim=7
+    - framework.action_model.loss_type=discrete_ce
+    - framework.action_model.action_horizon=1
+    - framework.action_model.future_action_window_size=0
+    - framework.action_model.past_action_window_size=0
+    - datasets.vla_data.include_state=true
+    - datasets.vla_data.action_type=discrete
+    - datasets.vla_data.sequential_step_sampling=false
+    - datasets.vla_data.per_device_batch_size=32
+    - datasets.vla_data.load_all_data_for_training=true
+    - datasets.vla_data.video_backend=torchvision_av
+    - datasets.vla_data.data_root_dir=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2
+    - datasets.vla_data.data_mix=deadly_corridor_train__bridge
+    - datasets.vla_data.eval_data_mix=deadly_corridor_train__bridge__val
+    - framework.qwenvl.base_vlm=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - rl_games.task=deadly_corridor
+    - rl_games.model_alias=openvla
+    - rl_games.initialization_mode=bridge
+    - rl_games.action_carrier=bridge
+    - rl_games.env_eval.distributed_mode=none
+    - rl_games.env_eval.latency.mode=single
+    - rl_games.env_eval.frameskip=4
+    - rl_games.env_eval.image_size=224
+    - rl_games.env_eval.seed=42
+    - rl_games.env_eval.fixed_episode_seeds=true
+    - rl_games.env_eval.latency_seed_stride=0
+    - rl_games.env_eval.task_seed_stride=0
+    - rl_games.env_eval.task_description='You are playing Deadly Corridor in VizDoom.
+      Choose actions from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT,
+      TURN_RIGHT, ATTACK.'
+    - rl_games.env_eval.latency.values=[0]
+    - rl_games.env_eval.latency.prompt_map_path=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2/deadly_corridor_train__bridge/latency_prompt_map.json
+    - rl_games.env_eval.mid_train.enabled=true
+    - rl_games.env_eval.mid_train.interval_steps=50
+    - rl_games.env_eval.mid_train.num_episodes=5
+    - rl_games.env_eval.mid_train.max_steps_per_episode=3600
+    - rl_games.env_eval.mid_train.latencies=[2]
+    - rl_games.env_eval.post_train.enabled=true
+    - rl_games.env_eval.post_train.num_episodes=5
+    - rl_games.env_eval.post_train.max_steps_per_episode=3600
+    - rl_games.env_eval.post_train.latencies=[0,2,4,6,8]
+    - rl_games.env_eval.deadly.action_layout=multibinary_7
+  job:
+    name: train_starvla_hydra
+    chdir: false
+    override_dirname: checkpoint.local.keep_last_n=1,checkpoint.save_best_model=true,checkpoint.save_pt_file=false,checkpoint.sync.enabled=false,checkpoint.sync.keep_last_n=0,datasets.vla_data.action_type=discrete,datasets.vla_data.data_mix=deadly_corridor_train__bridge,datasets.vla_data.data_root_dir=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2,datasets.vla_data.eval_data_mix=deadly_corridor_train__bridge__val,datasets.vla_data.include_state=true,datasets.vla_data.load_all_data_for_training=true,datasets.vla_data.per_device_batch_size=32,datasets.vla_data.sequential_step_sampling=false,datasets.vla_data.video_backend=torchvision_av,env=deadly_corridor,framework.action_model.action_dim=7,framework.action_model.action_env_dim=7,framework.action_model.action_horizon=1,framework.action_model.future_action_window_size=0,framework.action_model.loss_type=discrete_ce,framework.action_model.past_action_window_size=0,framework.action_model.state_dim=7,framework.name=QwenOFT,framework.qwenvl.attn_implementation=flash_attention_2,framework.qwenvl.base_vlm=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/playground/Pretrained_models/Qwen3-VL-4B-Instruct,framework.qwenvl.enable_gradient_checkpointing=true,init=bridge,mode=single,model=openvla,rl_games.action_carrier=bridge,rl_games.env_eval.deadly.action_layout=multibinary_7,rl_games.env_eval.distributed_mode=none,rl_games.env_eval.enabled=true,rl_games.env_eval.fixed_episode_seeds=true,rl_games.env_eval.frameskip=4,rl_games.env_eval.image_size=224,rl_games.env_eval.latency.mode=single,rl_games.env_eval.latency.prompt_map_path=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2/deadly_corridor_train__bridge/latency_prompt_map.json,rl_games.env_eval.latency.values=[0],rl_games.env_eval.latency_seed_stride=0,rl_games.env_eval.mid_train.enabled=true,rl_games.env_eval.mid_train.interval_steps=50,rl_games.env_eval.mid_train.latencies=[2],rl_games.env_eval.mid_train.max_steps_per_episode=3600,rl_games.env_eval.mid_train.num_episodes=5,rl_games.env_eval.post_train.enabled=true,rl_games.env_eval.post_train.latencies=[0,2,4,6,8],rl_games.env_eval.post_train.max_steps_per_episode=3600,rl_games.env_eval.post_train.num_episodes=5,rl_games.env_eval.seed=42,rl_games.env_eval.task_description='You
+      are playing Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD,
+      MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.',rl_games.env_eval.task_seed_stride=0,rl_games.initialization_mode=bridge,rl_games.model_alias=openvla,rl_games.task=deadly_corridor,run_id=openvla_deadly_corridor_fix_latency_2,run_root_dir=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/results/Checkpoints,seed=42,trainer.distributed_backend=none,trainer.eval_interval=100,trainer.eval_num_batches=20,trainer.gradient_accumulation_steps=16,trainer.gradient_clipping=1.0,trainer.is_resume=false,trainer.learning_rate.action_model=0.0001,trainer.learning_rate.base=2e-05,trainer.learning_rate.qwen_vl_interface=1e-05,trainer.logging_frequency=1,trainer.loss_scale.vla=1.0,trainer.loss_scale.vlm=0.1,trainer.lr_scheduler_type=cosine_with_min_lr,trainer.max_grad_norm=1.0,trainer.max_train_steps=500,trainer.num_warmup_steps=100,trainer.optimizer.betas=[0.9,
+      0.95],trainer.optimizer.eps=1e-08,trainer.optimizer.fused=false,trainer.optimizer.name=AdamW,trainer.optimizer.weight_decay=1e-08,trainer.pretrained_checkpoint=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt,trainer.resume_step=0,trainer.save_format=pt,trainer.save_interval=50,trainer.scheduler_specific_kwargs.min_lr=1e-06,trainer.weight_decay=0.0,wandb_entity=dongqianyu99-zhejiang-university,wandb_project=starVLA_rl_games
+    id: ???
+    num: ???
+    config_name: train
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.1'
+    cwd: /inspire/hdd/project/spatialintelligence/public/lzj/starVLA
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /inspire/hdd/project/spatialintelligence/public/lzj/starVLA/examples/rl_games/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/results/Checkpoints/openvla_deadly_corridor_fix_latency_2/hydra
+    choices:
+      cross_task_setup: null
+      checkpoint: default
+      mode: single
+      init: bridge
+      env: deadly_corridor
+      model: openvla
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/hydra/.hydra/overrides.yaml

@@ -0,0 +1,89 @@
+- model=openvla
+- env=deadly_corridor
+- init=bridge
+- mode=single
+- run_id=openvla_deadly_corridor_fix_latency_2
+- run_root_dir=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/results/Checkpoints
+- seed=42
+- wandb_entity=dongqianyu99-zhejiang-university
+- wandb_project=starVLA_rl_games
+- rl_games.env_eval.enabled=true
+- checkpoint.sync.enabled=false
+- checkpoint.sync.keep_last_n=0
+- checkpoint.local.keep_last_n=1
+- checkpoint.save_best_model=true
+- checkpoint.save_pt_file=false
+- trainer.is_resume=false
+- trainer.pretrained_checkpoint=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+- trainer.resume_step=0
+- trainer.max_train_steps=500
+- trainer.num_warmup_steps=100
+- trainer.save_interval=50
+- trainer.eval_interval=100
+- trainer.eval_num_batches=20
+- trainer.logging_frequency=1
+- trainer.gradient_accumulation_steps=16
+- trainer.distributed_backend=none
+- trainer.learning_rate.base=2e-05
+- trainer.learning_rate.qwen_vl_interface=1e-05
+- trainer.learning_rate.action_model=0.0001
+- trainer.lr_scheduler_type=cosine_with_min_lr
+- trainer.scheduler_specific_kwargs.min_lr=1e-06
+- trainer.loss_scale.vla=1.0
+- trainer.loss_scale.vlm=0.1
+- trainer.max_grad_norm=1.0
+- trainer.weight_decay=0.0
+- trainer.gradient_clipping=1.0
+- trainer.optimizer.name=AdamW
+- trainer.optimizer.betas=[0.9, 0.95]
+- trainer.optimizer.eps=1e-08
+- trainer.optimizer.weight_decay=1e-08
+- trainer.optimizer.fused=false
+- trainer.save_format=pt
+- framework.name=QwenOFT
+- framework.qwenvl.attn_implementation=flash_attention_2
+- framework.qwenvl.enable_gradient_checkpointing=true
+- framework.action_model.action_dim=7
+- framework.action_model.action_env_dim=7
+- framework.action_model.state_dim=7
+- framework.action_model.loss_type=discrete_ce
+- framework.action_model.action_horizon=1
+- framework.action_model.future_action_window_size=0
+- framework.action_model.past_action_window_size=0
+- datasets.vla_data.include_state=true
+- datasets.vla_data.action_type=discrete
+- datasets.vla_data.sequential_step_sampling=false
+- datasets.vla_data.per_device_batch_size=32
+- datasets.vla_data.load_all_data_for_training=true
+- datasets.vla_data.video_backend=torchvision_av
+- datasets.vla_data.data_root_dir=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2
+- datasets.vla_data.data_mix=deadly_corridor_train__bridge
+- datasets.vla_data.eval_data_mix=deadly_corridor_train__bridge__val
+- framework.qwenvl.base_vlm=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- rl_games.task=deadly_corridor
+- rl_games.model_alias=openvla
+- rl_games.initialization_mode=bridge
+- rl_games.action_carrier=bridge
+- rl_games.env_eval.distributed_mode=none
+- rl_games.env_eval.latency.mode=single
+- rl_games.env_eval.frameskip=4
+- rl_games.env_eval.image_size=224
+- rl_games.env_eval.seed=42
+- rl_games.env_eval.fixed_episode_seeds=true
+- rl_games.env_eval.latency_seed_stride=0
+- rl_games.env_eval.task_seed_stride=0
+- rl_games.env_eval.task_description='You are playing Deadly Corridor in VizDoom.
+  Choose actions from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT,
+  TURN_RIGHT, ATTACK.'
+- rl_games.env_eval.latency.values=[0]
+- rl_games.env_eval.latency.prompt_map_path=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/deadly_corridor_fix_latency_2/deadly_corridor_train__bridge/latency_prompt_map.json
+- rl_games.env_eval.mid_train.enabled=true
+- rl_games.env_eval.mid_train.interval_steps=50
+- rl_games.env_eval.mid_train.num_episodes=5
+- rl_games.env_eval.mid_train.max_steps_per_episode=3600
+- rl_games.env_eval.mid_train.latencies=[2]
+- rl_games.env_eval.post_train.enabled=true
+- rl_games.env_eval.post_train.num_episodes=5
+- rl_games.env_eval.post_train.max_steps_per_episode=3600
+- rl_games.env_eval.post_train.latencies=[0,2,4,6,8]
+- rl_games.env_eval.deadly.action_layout=multibinary_7

latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/summary.jsonl

@@ -0,0 +1,10 @@
+{"steps": 50}
+{"steps": 100}
+{"steps": 150}
+{"steps": 200}
+{"steps": 250}
+{"steps": 300}
+{"steps": 350}
+{"steps": 400}
+{"steps": 450}
+{"steps": 500}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/checkpoints/best_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adb853d16890002501675633ddefc2126fd3f102a63cf6ece7a5e56d0c8c662c
+size 9138230516

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/checkpoints/steps_2000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adb853d16890002501675633ddefc2126fd3f102a63cf6ece7a5e56d0c8c662c
+size 9138230516

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/config.full.yaml

@@ -0,0 +1,190 @@
+framework:
+  name: QwenOFT
+  qwenvl:
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    action_hidden_dim: 2560
+    future_action_window_size: 0
+    past_action_window_size: 0
+    loss_type: discrete_ce
+    state_dim: 7
+    action_horizon: 1
+    action_env_dim: 6
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/demon_attack_fix_latency_2_small
+    data_mix: demon_attack_train__bridge
+    eval_data_mix: demon_attack_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  converted_name: demon_attack_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: demon_attack_train
+  mixed_converted_name: demon_attack_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 2000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 20
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  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: 16
+  distributed_backend: none
+  is_resume: true
+  pretrained_checkpoint: /workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2_small/checkpoints/steps_600_state
+  resume_step: 600
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 4
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: 'You are playing Demon Attack from a single game image. Choose
+      exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'
+    demon_attack:
+      noop_max: 30
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: /workspace/starVLA/data/demon_attack_fix_latency_2_small/demon_attack_train__bridge/latency_prompt_map.json
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: true
+      interval_steps: 100
+      latencies:
+      - 2
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: true
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  task: demon_attack
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: demon_attack
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: true
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: openvla_demon_attack_fix_latency_2_small
+output_dir: /workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2_small
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/config.yaml

@@ -0,0 +1,92 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: true
+  save_pt_file: false
+  sync:
+    enabled: false
+    keep_last_n: 0
+    repo_id: null
+datasets:
+  vla_data:
+    data_mix: demon_attack_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: demon_attack_train__bridge__val
+    include_state: true
+    obs_image_size: null
+    per_device_batch_size: 32
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 6
+    action_hidden_dim: 2560
+    action_horizon: 1
+    action_model_type: MLP
+    loss_type: discrete_ce
+    state_dim: 7
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+output_dir: /workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2_small
+rl_games:
+  env_eval:
+    demon_attack:
+      noop_max: 30
+    distributed_mode: none
+    enabled: true
+    fixed_episode_seeds: true
+    frameskip: 4
+    image_size: 224
+    latency:
+      prompt_map_path: /workspace/starVLA/data/demon_attack_fix_latency_2_small/demon_attack_train__bridge/latency_prompt_map.json
+    latency_seed_stride: 0
+    mid_train:
+      enabled: true
+      interval_steps: 100
+      latencies:
+      - 2
+      max_steps_per_episode: 3600
+      num_episodes: 5
+    seed: 42
+    task_description: 'You are playing Demon Attack from a single game image. Choose
+      exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'
+    task_seed_stride: 0
+    vectorized:
+      enabled: false
+  model_alias: openvla
+  task: demon_attack
+run_id: openvla_demon_attack_fix_latency_2_small
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  distributed_backend: none
+  eval_interval: 100
+  eval_num_batches: 20
+  freeze_modules: ''
+  gradient_accumulation_steps: 16
+  gradient_clipping: 1.0
+  is_resume: true
+  learning_rate:
+    action_model: 0.0001
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 10
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 2000
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: false
+    weight_decay: 1.0e-08
+  pretrained_checkpoint: /workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2_small/checkpoints/steps_600_state
+  resume_step: 600
+  save_interval: 100
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.12531481683254242,
+        0.3168888986110687,
+        0.1617870330810547,
+        0.12757407128810883,
+        0.1149444431066513,
+        0.15349073708057404,
+        0.0
+      ],
+      "std": [
+        0.33097583055496216,
+        0.46518969535827637,
+        0.36836159229278564,
+        0.33368170261383057,
+        0.31906843185424805,
+        0.36036959290504456,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 108000,
+    "num_trajectories": 30
+  }
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.125083327293396,
+        0.3141111135482788,
+        0.16361111402511597,
+        0.12720833718776703,
+        0.11493055522441864,
+        0.15505555272102356,
+        0.0
+      ],
+      "std": [
+        0.3307972252368927,
+        0.4641607403755188,
+        0.36990460753440857,
+        0.33313921093940735,
+        0.31889820098876953,
+        0.3619128167629242,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 72000,
+    "num_trajectories": 20
+  }
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/hydra/.hydra/config.yaml

@@ -0,0 +1,188 @@
+framework:
+  qwenvl:
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    state_dim: 7
+    loss_type: discrete_ce
+    action_horizon: 1
+    future_action_window_size: 0
+    past_action_window_size: 0
+    action_dim: 7
+    action_env_dim: 6
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/demon_attack_fix_latency_2
+    data_mix: demon_attack_train__bridge
+    eval_data_mix: demon_attack_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ${dataset.single_source_hf}
+  converted_name: ${dataset.single_converted_name}
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: demon_attack_train
+  mixed_converted_name: demon_attack_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: ${dataset.single_latency_filter}
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: ${bridge_base_model.repo_id.${model}}
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 2000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 20
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  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: 16
+  distributed_backend: none
+  is_resume: true
+  pretrained_checkpoint: /workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2/checkpoints/steps_600_state
+  resume_step: 600
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: ${bridge_base_model.local_dir.${model}}
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 4
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: 'You are playing Demon Attack from a single game image. Choose
+      exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'
+    demon_attack:
+      noop_max: 30
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: /workspace/starVLA/data/demon_attack_fix_latency_2/demon_attack_train__bridge/latency_prompt_map.json
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: true
+      interval_steps: 100
+      latencies:
+      - 2
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: true
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  task: demon_attack
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: demon_attack
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: true
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: openvla_demon_attack_fix_latency_2
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/hydra/.hydra/hydra.yaml

@@ -0,0 +1,251 @@
+hydra:
+  run:
+    dir: ${run_root_dir}/${run_id}/hydra
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - model=openvla
+    - env=demon_attack
+    - init=bridge
+    - mode=single
+    - run_id=openvla_demon_attack_fix_latency_2
+    - run_root_dir=/workspace/starVLA/results/Checkpoints
+    - seed=42
+    - wandb_entity=dongqianyu99-zhejiang-university
+    - wandb_project=starVLA_rl_games
+    - rl_games.env_eval.enabled=true
+    - checkpoint.sync.enabled=false
+    - checkpoint.sync.keep_last_n=0
+    - checkpoint.local.keep_last_n=1
+    - checkpoint.save_best_model=true
+    - checkpoint.save_pt_file=false
+    - trainer.is_resume=true
+    - trainer.pretrained_checkpoint=/workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2/checkpoints/steps_600_state
+    - trainer.resume_step=600
+    - trainer.max_train_steps=2000
+    - trainer.num_warmup_steps=100
+    - trainer.save_interval=100
+    - trainer.eval_interval=100
+    - trainer.eval_num_batches=20
+    - trainer.logging_frequency=10
+    - trainer.gradient_accumulation_steps=16
+    - trainer.distributed_backend=none
+    - trainer.learning_rate.base=2e-05
+    - trainer.learning_rate.qwen_vl_interface=1e-05
+    - trainer.learning_rate.action_model=0.0001
+    - trainer.lr_scheduler_type=cosine_with_min_lr
+    - trainer.scheduler_specific_kwargs.min_lr=1e-06
+    - trainer.loss_scale.vla=1.0
+    - trainer.loss_scale.vlm=0.1
+    - trainer.max_grad_norm=1.0
+    - trainer.weight_decay=0.0
+    - trainer.gradient_clipping=1.0
+    - trainer.optimizer.name=AdamW
+    - trainer.optimizer.betas=[0.9, 0.95]
+    - trainer.optimizer.eps=1e-08
+    - trainer.optimizer.weight_decay=1e-08
+    - trainer.optimizer.fused=false
+    - trainer.save_format=pt
+    - framework.name=QwenOFT
+    - framework.qwenvl.attn_implementation=flash_attention_2
+    - framework.qwenvl.enable_gradient_checkpointing=true
+    - framework.action_model.action_dim=7
+    - framework.action_model.action_env_dim=6
+    - framework.action_model.state_dim=7
+    - framework.action_model.loss_type=discrete_ce
+    - framework.action_model.action_horizon=1
+    - framework.action_model.future_action_window_size=0
+    - framework.action_model.past_action_window_size=0
+    - datasets.vla_data.include_state=true
+    - datasets.vla_data.action_type=discrete
+    - datasets.vla_data.sequential_step_sampling=false
+    - datasets.vla_data.per_device_batch_size=32
+    - datasets.vla_data.load_all_data_for_training=true
+    - datasets.vla_data.video_backend=torchvision_av
+    - datasets.vla_data.data_root_dir=/workspace/starVLA/data/demon_attack_fix_latency_2
+    - datasets.vla_data.data_mix=demon_attack_train__bridge
+    - datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val
+    - framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - rl_games.task=demon_attack
+    - rl_games.model_alias=openvla
+    - rl_games.initialization_mode=bridge
+    - rl_games.action_carrier=bridge
+    - rl_games.env_eval.distributed_mode=none
+    - rl_games.env_eval.latency.mode=single
+    - rl_games.env_eval.frameskip=4
+    - rl_games.env_eval.image_size=224
+    - rl_games.env_eval.seed=42
+    - rl_games.env_eval.fixed_episode_seeds=true
+    - rl_games.env_eval.latency_seed_stride=0
+    - rl_games.env_eval.task_seed_stride=0
+    - 'rl_games.env_eval.task_description=''You are playing Demon Attack from a single
+      game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE,
+      LEFTFIRE.'''
+    - rl_games.env_eval.latency.values=[0]
+    - rl_games.env_eval.latency.prompt_map_path=/workspace/starVLA/data/demon_attack_fix_latency_2/demon_attack_train__bridge/latency_prompt_map.json
+    - rl_games.env_eval.mid_train.enabled=true
+    - rl_games.env_eval.mid_train.interval_steps=100
+    - rl_games.env_eval.mid_train.num_episodes=5
+    - rl_games.env_eval.mid_train.max_steps_per_episode=3600
+    - rl_games.env_eval.mid_train.latencies=[2]
+    - rl_games.env_eval.post_train.enabled=true
+    - rl_games.env_eval.post_train.num_episodes=5
+    - rl_games.env_eval.post_train.max_steps_per_episode=3600
+    - rl_games.env_eval.post_train.latencies=[0,1,2,3,4]
+  job:
+    name: train_starvla_hydra
+    chdir: false
+    override_dirname: 'checkpoint.local.keep_last_n=1,checkpoint.save_best_model=true,checkpoint.save_pt_file=false,checkpoint.sync.enabled=false,checkpoint.sync.keep_last_n=0,datasets.vla_data.action_type=discrete,datasets.vla_data.data_mix=demon_attack_train__bridge,datasets.vla_data.data_root_dir=/workspace/starVLA/data/demon_attack_fix_latency_2,datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val,datasets.vla_data.include_state=true,datasets.vla_data.load_all_data_for_training=true,datasets.vla_data.per_device_batch_size=32,datasets.vla_data.sequential_step_sampling=false,datasets.vla_data.video_backend=torchvision_av,env=demon_attack,framework.action_model.action_dim=7,framework.action_model.action_env_dim=6,framework.action_model.action_horizon=1,framework.action_model.future_action_window_size=0,framework.action_model.loss_type=discrete_ce,framework.action_model.past_action_window_size=0,framework.action_model.state_dim=7,framework.name=QwenOFT,framework.qwenvl.attn_implementation=flash_attention_2,framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct,framework.qwenvl.enable_gradient_checkpointing=true,init=bridge,mode=single,model=openvla,rl_games.action_carrier=bridge,rl_games.env_eval.distributed_mode=none,rl_games.env_eval.enabled=true,rl_games.env_eval.fixed_episode_seeds=true,rl_games.env_eval.frameskip=4,rl_games.env_eval.image_size=224,rl_games.env_eval.latency.mode=single,rl_games.env_eval.latency.prompt_map_path=/workspace/starVLA/data/demon_attack_fix_latency_2/demon_attack_train__bridge/latency_prompt_map.json,rl_games.env_eval.latency.values=[0],rl_games.env_eval.latency_seed_stride=0,rl_games.env_eval.mid_train.enabled=true,rl_games.env_eval.mid_train.interval_steps=100,rl_games.env_eval.mid_train.latencies=[2],rl_games.env_eval.mid_train.max_steps_per_episode=3600,rl_games.env_eval.mid_train.num_episodes=5,rl_games.env_eval.post_train.enabled=true,rl_games.env_eval.post_train.latencies=[0,1,2,3,4],rl_games.env_eval.post_train.max_steps_per_episode=3600,rl_games.env_eval.post_train.num_episodes=5,rl_games.env_eval.seed=42,rl_games.env_eval.task_description=''You
+      are playing Demon Attack from a single game image. Choose exactly one action
+      from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'',rl_games.env_eval.task_seed_stride=0,rl_games.initialization_mode=bridge,rl_games.model_alias=openvla,rl_games.task=demon_attack,run_id=openvla_demon_attack_fix_latency_2,run_root_dir=/workspace/starVLA/results/Checkpoints,seed=42,trainer.distributed_backend=none,trainer.eval_interval=100,trainer.eval_num_batches=20,trainer.gradient_accumulation_steps=16,trainer.gradient_clipping=1.0,trainer.is_resume=true,trainer.learning_rate.action_model=0.0001,trainer.learning_rate.base=2e-05,trainer.learning_rate.qwen_vl_interface=1e-05,trainer.logging_frequency=10,trainer.loss_scale.vla=1.0,trainer.loss_scale.vlm=0.1,trainer.lr_scheduler_type=cosine_with_min_lr,trainer.max_grad_norm=1.0,trainer.max_train_steps=2000,trainer.num_warmup_steps=100,trainer.optimizer.betas=[0.9,
+      0.95],trainer.optimizer.eps=1e-08,trainer.optimizer.fused=false,trainer.optimizer.name=AdamW,trainer.optimizer.weight_decay=1e-08,trainer.pretrained_checkpoint=/workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2/checkpoints/steps_600_state,trainer.resume_step=600,trainer.save_format=pt,trainer.save_interval=100,trainer.scheduler_specific_kwargs.min_lr=1e-06,trainer.weight_decay=0.0,wandb_entity=dongqianyu99-zhejiang-university,wandb_project=starVLA_rl_games'
+    id: ???
+    num: ???
+    config_name: train
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.1'
+    cwd: /workspace/starVLA
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/starVLA/examples/rl_games/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2/hydra
+    choices:
+      cross_task_setup: null
+      checkpoint: default
+      mode: single
+      init: bridge
+      env: demon_attack
+      model: openvla
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/hydra/.hydra/overrides.yaml

@@ -0,0 +1,88 @@
+- model=openvla
+- env=demon_attack
+- init=bridge
+- mode=single
+- run_id=openvla_demon_attack_fix_latency_2
+- run_root_dir=/workspace/starVLA/results/Checkpoints
+- seed=42
+- wandb_entity=dongqianyu99-zhejiang-university
+- wandb_project=starVLA_rl_games
+- rl_games.env_eval.enabled=true
+- checkpoint.sync.enabled=false
+- checkpoint.sync.keep_last_n=0
+- checkpoint.local.keep_last_n=1
+- checkpoint.save_best_model=true
+- checkpoint.save_pt_file=false
+- trainer.is_resume=true
+- trainer.pretrained_checkpoint=/workspace/starVLA/results/Checkpoints/openvla_demon_attack_fix_latency_2/checkpoints/steps_600_state
+- trainer.resume_step=600
+- trainer.max_train_steps=2000
+- trainer.num_warmup_steps=100
+- trainer.save_interval=100
+- trainer.eval_interval=100
+- trainer.eval_num_batches=20
+- trainer.logging_frequency=10
+- trainer.gradient_accumulation_steps=16
+- trainer.distributed_backend=none
+- trainer.learning_rate.base=2e-05
+- trainer.learning_rate.qwen_vl_interface=1e-05
+- trainer.learning_rate.action_model=0.0001
+- trainer.lr_scheduler_type=cosine_with_min_lr
+- trainer.scheduler_specific_kwargs.min_lr=1e-06
+- trainer.loss_scale.vla=1.0
+- trainer.loss_scale.vlm=0.1
+- trainer.max_grad_norm=1.0
+- trainer.weight_decay=0.0
+- trainer.gradient_clipping=1.0
+- trainer.optimizer.name=AdamW
+- trainer.optimizer.betas=[0.9, 0.95]
+- trainer.optimizer.eps=1e-08
+- trainer.optimizer.weight_decay=1e-08
+- trainer.optimizer.fused=false
+- trainer.save_format=pt
+- framework.name=QwenOFT
+- framework.qwenvl.attn_implementation=flash_attention_2
+- framework.qwenvl.enable_gradient_checkpointing=true
+- framework.action_model.action_dim=7
+- framework.action_model.action_env_dim=6
+- framework.action_model.state_dim=7
+- framework.action_model.loss_type=discrete_ce
+- framework.action_model.action_horizon=1
+- framework.action_model.future_action_window_size=0
+- framework.action_model.past_action_window_size=0
+- datasets.vla_data.include_state=true
+- datasets.vla_data.action_type=discrete
+- datasets.vla_data.sequential_step_sampling=false
+- datasets.vla_data.per_device_batch_size=32
+- datasets.vla_data.load_all_data_for_training=true
+- datasets.vla_data.video_backend=torchvision_av
+- datasets.vla_data.data_root_dir=/workspace/starVLA/data/demon_attack_fix_latency_2
+- datasets.vla_data.data_mix=demon_attack_train__bridge
+- datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val
+- framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- rl_games.task=demon_attack
+- rl_games.model_alias=openvla
+- rl_games.initialization_mode=bridge
+- rl_games.action_carrier=bridge
+- rl_games.env_eval.distributed_mode=none
+- rl_games.env_eval.latency.mode=single
+- rl_games.env_eval.frameskip=4
+- rl_games.env_eval.image_size=224
+- rl_games.env_eval.seed=42
+- rl_games.env_eval.fixed_episode_seeds=true
+- rl_games.env_eval.latency_seed_stride=0
+- rl_games.env_eval.task_seed_stride=0
+- 'rl_games.env_eval.task_description=''You are playing Demon Attack from a single
+  game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE,
+  LEFTFIRE.'''
+- rl_games.env_eval.latency.values=[0]
+- rl_games.env_eval.latency.prompt_map_path=/workspace/starVLA/data/demon_attack_fix_latency_2/demon_attack_train__bridge/latency_prompt_map.json
+- rl_games.env_eval.mid_train.enabled=true
+- rl_games.env_eval.mid_train.interval_steps=100
+- rl_games.env_eval.mid_train.num_episodes=5
+- rl_games.env_eval.mid_train.max_steps_per_episode=3600
+- rl_games.env_eval.mid_train.latencies=[2]
+- rl_games.env_eval.post_train.enabled=true
+- rl_games.env_eval.post_train.num_episodes=5
+- rl_games.env_eval.post_train.max_steps_per_episode=3600
+- rl_games.env_eval.post_train.latencies=[0,1,2,3,4]

latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/summary.jsonl

@@ -0,0 +1,20 @@
+{"steps": 100}
+{"steps": 200}
+{"steps": 300}
+{"steps": 400}
+{"steps": 500}
+{"steps": 600}
+{"steps": 700}
+{"steps": 800}
+{"steps": 900}
+{"steps": 1000}
+{"steps": 1100}
+{"steps": 1200}
+{"steps": 1300}
+{"steps": 1400}
+{"steps": 1500}
+{"steps": 1600}
+{"steps": 1700}
+{"steps": 1800}
+{"steps": 1900}
+{"steps": 2000}

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/checkpoints/steps_5000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cff93451f827160bf275aa8bc0bcab83cbd71645ef0605b2e8869bd3691a3bb6
+size 9138230516

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/config.full.yaml

@@ -0,0 +1,198 @@
+framework:
+  name: QwenOFT
+  qwenvl:
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    action_hidden_dim: 2560
+    future_action_window_size: 0
+    past_action_window_size: 0
+    loss_type: discrete_ce
+    state_dim: 7
+    action_horizon: 1
+    action_env_dim: 2
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/flappy_fix_latency_1_200ep
+    data_mix: flappy_train__bridge
+    eval_data_mix: flappy_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  config_name: null
+  converted_name: flappy_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train
+  mixed_converted_name: flappy_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  eval_action_classification: true
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 4
+  distributed_backend: none
+  is_resume: false
+  pretrained_checkpoint: null
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: ''
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: null
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: false
+      interval_steps: 100
+      latencies:
+      - 0
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: false
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  task: flappy
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: flappy
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: flappy_fix_latency_1_200ep
+output_dir: /workspace/starVLA/results/Checkpoints/flappy_fix_latency_1_200ep
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/config.yaml

@@ -0,0 +1,71 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: false
+  save_pt_file: false
+  sync:
+    enabled: false
+    keep_last_n: 0
+    repo_id: null
+datasets:
+  vla_data:
+    data_mix: flappy_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: flappy_train__bridge__val
+    obs_image_size: null
+    per_device_batch_size: 32
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 2
+    action_hidden_dim: 2560
+    action_horizon: 1
+    action_model_type: MLP
+    loss_type: discrete_ce
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+output_dir: /workspace/starVLA/results/Checkpoints/flappy_fix_latency_1_200ep
+rl_games:
+  env_eval:
+    enabled: true
+    mid_train:
+      enabled: false
+      interval_steps: 100
+  task: flappy
+run_id: flappy_fix_latency_1_200ep
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  distributed_backend: none
+  eval_action_classification: true
+  eval_interval: 100
+  eval_num_batches: 100
+  freeze_modules: ''
+  gradient_accumulation_steps: 4
+  gradient_clipping: 1.0
+  is_resume: false
+  learning_rate:
+    action_model: 0.0001
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: false
+    weight_decay: 1.0e-08
+  pretrained_checkpoint: null
+  reload_modules: null
+  save_interval: 100
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.6979423761367798,
+        0.3020575940608978,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.4588429629802704,
+        0.4588429629802704,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 297483,
+    "num_trajectories": 180
+  }
+}

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.8606944680213928,
+        0.13930556178092957,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.3463059663772583,
+        0.3463059663772583,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 72000,
+    "num_trajectories": 20
+  }
+}

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/hydra/.hydra/config.yaml

@@ -0,0 +1,196 @@
+framework:
+  qwenvl:
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    state_dim: 7
+    loss_type: discrete_ce
+    action_horizon: 1
+    future_action_window_size: 0
+    past_action_window_size: 0
+    action_dim: 7
+    action_env_dim: 2
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/flappy_fix_latency_1_200ep
+    data_mix: flappy_train__bridge
+    eval_data_mix: flappy_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ${dataset.single_source_hf}
+  config_name: null
+  converted_name: ${dataset.single_converted_name}
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train
+  mixed_converted_name: flappy_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: ${dataset.single_latency_filter}
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: ${bridge_base_model.repo_id.${model}}
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  eval_action_classification: true
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 4
+  distributed_backend: none
+  is_resume: false
+  pretrained_checkpoint: /workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: ${bridge_base_model.local_dir.${model}}
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: ''
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: null
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: false
+      interval_steps: 100
+      latencies:
+      - 0
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: false
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  task: flappy
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: flappy
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: flappy_fix_latency_1_200ep
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/hydra/.hydra/hydra.yaml

@@ -0,0 +1,252 @@
+hydra:
+  run:
+    dir: ${run_root_dir}/${run_id}/hydra
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - model=openvla
+    - env=flappy
+    - init=bridge
+    - mode=single
+    - run_id=flappy_fix_latency_1_200ep
+    - run_root_dir=/workspace/starVLA/results/Checkpoints
+    - seed=42
+    - wandb_entity=dongqianyu99-zhejiang-university
+    - wandb_project=starVLA_rl_games
+    - rl_games.env_eval.enabled=true
+    - checkpoint.sync.enabled=false
+    - checkpoint.sync.keep_last_n=0
+    - checkpoint.local.keep_last_n=1
+    - checkpoint.save_best_model=false
+    - checkpoint.save_pt_file=false
+    - trainer.is_resume=false
+    - trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+    - trainer.resume_step=0
+    - trainer.max_train_steps=5000
+    - trainer.num_warmup_steps=100
+    - trainer.save_interval=100
+    - trainer.eval_interval=100
+    - trainer.eval_num_batches=100
+    - trainer.logging_frequency=1
+    - trainer.gradient_accumulation_steps=4
+    - trainer.distributed_backend=none
+    - trainer.learning_rate.base=2e-05
+    - trainer.learning_rate.qwen_vl_interface=1e-05
+    - trainer.learning_rate.action_model=0.0001
+    - trainer.lr_scheduler_type=cosine_with_min_lr
+    - trainer.scheduler_specific_kwargs.min_lr=1e-06
+    - trainer.loss_scale.vla=1.0
+    - trainer.loss_scale.vlm=0.1
+    - trainer.max_grad_norm=1.0
+    - trainer.weight_decay=0.0
+    - trainer.gradient_clipping=1.0
+    - trainer.optimizer.name=AdamW
+    - trainer.optimizer.betas=[0.9, 0.95]
+    - trainer.optimizer.eps=1e-08
+    - trainer.optimizer.weight_decay=1e-08
+    - trainer.optimizer.fused=false
+    - trainer.save_format=pt
+    - framework.name=QwenOFT
+    - framework.qwenvl.attn_implementation=flash_attention_2
+    - framework.qwenvl.enable_gradient_checkpointing=true
+    - framework.action_model.action_dim=7
+    - framework.action_model.action_env_dim=2
+    - framework.action_model.state_dim=7
+    - framework.action_model.loss_type=discrete_ce
+    - framework.action_model.action_horizon=1
+    - framework.action_model.future_action_window_size=0
+    - framework.action_model.past_action_window_size=0
+    - datasets.vla_data.include_state=true
+    - datasets.vla_data.action_type=discrete
+    - datasets.vla_data.sequential_step_sampling=false
+    - datasets.vla_data.shuffle=true
+    - datasets.vla_data.action_balance.enabled=false
+    - datasets.vla_data.action_balance.strategy=balanced_epoch
+    - datasets.vla_data.action_balance.action_key=action_id
+    - datasets.vla_data.action_balance.target_flap_fraction=0.3
+    - datasets.vla_data.action_balance.noop_id=0
+    - datasets.vla_data.action_balance.flap_id=1
+    - datasets.vla_data.per_device_batch_size=32
+    - datasets.vla_data.load_all_data_for_training=true
+    - datasets.vla_data.video_backend=torchvision_av
+    - datasets.vla_data.data_root_dir=/workspace/starVLA/data/flappy_fix_latency_1_200ep
+    - datasets.vla_data.data_mix=flappy_train__bridge
+    - datasets.vla_data.eval_data_mix=flappy_train__bridge__val
+    - framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - rl_games.task=flappy
+    - rl_games.model_alias=openvla
+    - rl_games.initialization_mode=bridge
+    - rl_games.action_carrier=bridge
+    - rl_games.env_eval.distributed_mode=none
+    - rl_games.env_eval.latency.mode=single
+    - rl_games.env_eval.frameskip=1
+    - rl_games.env_eval.image_size=224
+    - rl_games.env_eval.seed=42
+    - rl_games.env_eval.fixed_episode_seeds=true
+    - rl_games.env_eval.latency_seed_stride=0
+    - rl_games.env_eval.task_seed_stride=0
+    - rl_games.env_eval.latency.values=[0]
+    - rl_games.env_eval.mid_train.enabled=false
+    - rl_games.env_eval.mid_train.interval_steps=100
+    - rl_games.env_eval.mid_train.num_episodes=5
+    - rl_games.env_eval.mid_train.max_steps_per_episode=3600
+    - rl_games.env_eval.mid_train.latencies=[0]
+    - rl_games.env_eval.post_train.enabled=false
+    - rl_games.env_eval.post_train.num_episodes=5
+    - rl_games.env_eval.post_train.max_steps_per_episode=3600
+    - rl_games.env_eval.post_train.latencies=[0,1,2,3,4]
+  job:
+    name: train_starvla_hydra
+    chdir: false
+    override_dirname: checkpoint.local.keep_last_n=1,checkpoint.save_best_model=false,checkpoint.save_pt_file=false,checkpoint.sync.enabled=false,checkpoint.sync.keep_last_n=0,datasets.vla_data.action_balance.action_key=action_id,datasets.vla_data.action_balance.enabled=false,datasets.vla_data.action_balance.flap_id=1,datasets.vla_data.action_balance.noop_id=0,datasets.vla_data.action_balance.strategy=balanced_epoch,datasets.vla_data.action_balance.target_flap_fraction=0.3,datasets.vla_data.action_type=discrete,datasets.vla_data.data_mix=flappy_train__bridge,datasets.vla_data.data_root_dir=/workspace/starVLA/data/flappy_fix_latency_1_200ep,datasets.vla_data.eval_data_mix=flappy_train__bridge__val,datasets.vla_data.include_state=true,datasets.vla_data.load_all_data_for_training=true,datasets.vla_data.per_device_batch_size=32,datasets.vla_data.sequential_step_sampling=false,datasets.vla_data.shuffle=true,datasets.vla_data.video_backend=torchvision_av,env=flappy,framework.action_model.action_dim=7,framework.action_model.action_env_dim=2,framework.action_model.action_horizon=1,framework.action_model.future_action_window_size=0,framework.action_model.loss_type=discrete_ce,framework.action_model.past_action_window_size=0,framework.action_model.state_dim=7,framework.name=QwenOFT,framework.qwenvl.attn_implementation=flash_attention_2,framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct,framework.qwenvl.enable_gradient_checkpointing=true,init=bridge,mode=single,model=openvla,rl_games.action_carrier=bridge,rl_games.env_eval.distributed_mode=none,rl_games.env_eval.enabled=true,rl_games.env_eval.fixed_episode_seeds=true,rl_games.env_eval.frameskip=1,rl_games.env_eval.image_size=224,rl_games.env_eval.latency.mode=single,rl_games.env_eval.latency.values=[0],rl_games.env_eval.latency_seed_stride=0,rl_games.env_eval.mid_train.enabled=false,rl_games.env_eval.mid_train.interval_steps=100,rl_games.env_eval.mid_train.latencies=[0],rl_games.env_eval.mid_train.max_steps_per_episode=3600,rl_games.env_eval.mid_train.num_episodes=5,rl_games.env_eval.post_train.enabled=false,rl_games.env_eval.post_train.latencies=[0,1,2,3,4],rl_games.env_eval.post_train.max_steps_per_episode=3600,rl_games.env_eval.post_train.num_episodes=5,rl_games.env_eval.seed=42,rl_games.env_eval.task_seed_stride=0,rl_games.initialization_mode=bridge,rl_games.model_alias=openvla,rl_games.task=flappy,run_id=flappy_fix_latency_1_200ep,run_root_dir=/workspace/starVLA/results/Checkpoints,seed=42,trainer.distributed_backend=none,trainer.eval_interval=100,trainer.eval_num_batches=100,trainer.gradient_accumulation_steps=4,trainer.gradient_clipping=1.0,trainer.is_resume=false,trainer.learning_rate.action_model=0.0001,trainer.learning_rate.base=2e-05,trainer.learning_rate.qwen_vl_interface=1e-05,trainer.logging_frequency=1,trainer.loss_scale.vla=1.0,trainer.loss_scale.vlm=0.1,trainer.lr_scheduler_type=cosine_with_min_lr,trainer.max_grad_norm=1.0,trainer.max_train_steps=5000,trainer.num_warmup_steps=100,trainer.optimizer.betas=[0.9,
+      0.95],trainer.optimizer.eps=1e-08,trainer.optimizer.fused=false,trainer.optimizer.name=AdamW,trainer.optimizer.weight_decay=1e-08,trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt,trainer.resume_step=0,trainer.save_format=pt,trainer.save_interval=100,trainer.scheduler_specific_kwargs.min_lr=1e-06,trainer.weight_decay=0.0,wandb_entity=dongqianyu99-zhejiang-university,wandb_project=starVLA_rl_games
+    id: ???
+    num: ???
+    config_name: train
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.3
+    version_base: '1.1'
+    cwd: /workspace/starVLA
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/starVLA/examples/rl_games/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/starVLA/results/Checkpoints/flappy_fix_latency_1_200ep/hydra
+    choices:
+      cross_task_setup: null
+      checkpoint: default
+      mode: single
+      init: bridge
+      env: flappy
+      model: openvla
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/hydra/.hydra/overrides.yaml

@@ -0,0 +1,91 @@
+- model=openvla
+- env=flappy
+- init=bridge
+- mode=single
+- run_id=flappy_fix_latency_1_200ep
+- run_root_dir=/workspace/starVLA/results/Checkpoints
+- seed=42
+- wandb_entity=dongqianyu99-zhejiang-university
+- wandb_project=starVLA_rl_games
+- rl_games.env_eval.enabled=true
+- checkpoint.sync.enabled=false
+- checkpoint.sync.keep_last_n=0
+- checkpoint.local.keep_last_n=1
+- checkpoint.save_best_model=false
+- checkpoint.save_pt_file=false
+- trainer.is_resume=false
+- trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+- trainer.resume_step=0
+- trainer.max_train_steps=5000
+- trainer.num_warmup_steps=100
+- trainer.save_interval=100
+- trainer.eval_interval=100
+- trainer.eval_num_batches=100
+- trainer.logging_frequency=1
+- trainer.gradient_accumulation_steps=4
+- trainer.distributed_backend=none
+- trainer.learning_rate.base=2e-05
+- trainer.learning_rate.qwen_vl_interface=1e-05
+- trainer.learning_rate.action_model=0.0001
+- trainer.lr_scheduler_type=cosine_with_min_lr
+- trainer.scheduler_specific_kwargs.min_lr=1e-06
+- trainer.loss_scale.vla=1.0
+- trainer.loss_scale.vlm=0.1
+- trainer.max_grad_norm=1.0
+- trainer.weight_decay=0.0
+- trainer.gradient_clipping=1.0
+- trainer.optimizer.name=AdamW
+- trainer.optimizer.betas=[0.9, 0.95]
+- trainer.optimizer.eps=1e-08
+- trainer.optimizer.weight_decay=1e-08
+- trainer.optimizer.fused=false
+- trainer.save_format=pt
+- framework.name=QwenOFT
+- framework.qwenvl.attn_implementation=flash_attention_2
+- framework.qwenvl.enable_gradient_checkpointing=true
+- framework.action_model.action_dim=7
+- framework.action_model.action_env_dim=2
+- framework.action_model.state_dim=7
+- framework.action_model.loss_type=discrete_ce
+- framework.action_model.action_horizon=1
+- framework.action_model.future_action_window_size=0
+- framework.action_model.past_action_window_size=0
+- datasets.vla_data.include_state=true
+- datasets.vla_data.action_type=discrete
+- datasets.vla_data.sequential_step_sampling=false
+- datasets.vla_data.shuffle=true
+- datasets.vla_data.action_balance.enabled=false
+- datasets.vla_data.action_balance.strategy=balanced_epoch
+- datasets.vla_data.action_balance.action_key=action_id
+- datasets.vla_data.action_balance.target_flap_fraction=0.3
+- datasets.vla_data.action_balance.noop_id=0
+- datasets.vla_data.action_balance.flap_id=1
+- datasets.vla_data.per_device_batch_size=32
+- datasets.vla_data.load_all_data_for_training=true
+- datasets.vla_data.video_backend=torchvision_av
+- datasets.vla_data.data_root_dir=/workspace/starVLA/data/flappy_fix_latency_1_200ep
+- datasets.vla_data.data_mix=flappy_train__bridge
+- datasets.vla_data.eval_data_mix=flappy_train__bridge__val
+- framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- rl_games.task=flappy
+- rl_games.model_alias=openvla
+- rl_games.initialization_mode=bridge
+- rl_games.action_carrier=bridge
+- rl_games.env_eval.distributed_mode=none
+- rl_games.env_eval.latency.mode=single
+- rl_games.env_eval.frameskip=1
+- rl_games.env_eval.image_size=224
+- rl_games.env_eval.seed=42
+- rl_games.env_eval.fixed_episode_seeds=true
+- rl_games.env_eval.latency_seed_stride=0
+- rl_games.env_eval.task_seed_stride=0
+- rl_games.env_eval.latency.values=[0]
+- rl_games.env_eval.mid_train.enabled=false
+- rl_games.env_eval.mid_train.interval_steps=100
+- rl_games.env_eval.mid_train.num_episodes=5
+- rl_games.env_eval.mid_train.max_steps_per_episode=3600
+- rl_games.env_eval.mid_train.latencies=[0]
+- rl_games.env_eval.post_train.enabled=false
+- rl_games.env_eval.post_train.num_episodes=5
+- rl_games.env_eval.post_train.max_steps_per_episode=3600
+- rl_games.env_eval.post_train.latencies=[0,1,2,3,4]

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/summary.jsonl

@@ -0,0 +1,50 @@
+{"steps": 100}
+{"steps": 200}
+{"steps": 300}
+{"steps": 400}
+{"steps": 500}
+{"steps": 600}
+{"steps": 700}
+{"steps": 800}
+{"steps": 900}
+{"steps": 1000}
+{"steps": 1100}
+{"steps": 1200}
+{"steps": 1300}
+{"steps": 1400}
+{"steps": 1500}
+{"steps": 1600}
+{"steps": 1700}
+{"steps": 1800}
+{"steps": 1900}
+{"steps": 2000}
+{"steps": 2100}
+{"steps": 2200}
+{"steps": 2300}
+{"steps": 2400}
+{"steps": 2500}
+{"steps": 2600}
+{"steps": 2700}
+{"steps": 2800}
+{"steps": 2900}
+{"steps": 3000}
+{"steps": 3100}
+{"steps": 3200}
+{"steps": 3300}
+{"steps": 3400}
+{"steps": 3500}
+{"steps": 3600}
+{"steps": 3700}
+{"steps": 3800}
+{"steps": 3900}
+{"steps": 4000}
+{"steps": 4100}
+{"steps": 4200}
+{"steps": 4300}
+{"steps": 4400}
+{"steps": 4500}
+{"steps": 4600}
+{"steps": 4700}
+{"steps": 4800}
+{"steps": 4900}
+{"steps": 5000}

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/checkpoints/steps_5000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aabcc8e3a22947e22d161d324de0e613a03300f39fa1dd28bffa2df927e994cf
+size 9138230516

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/config.full.yaml

@@ -0,0 +1,199 @@
+framework:
+  name: QwenOFT
+  qwenvl:
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    action_hidden_dim: 2560
+    future_action_window_size: 0
+    past_action_window_size: 0
+    loss_type: discrete_ce
+    state_dim: 7
+    action_horizon: 1
+    action_env_dim: 2
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/flappy_fix_latency_2_200ep
+    data_mix: flappy_train__bridge
+    eval_data_mix: flappy_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  config_name: null
+  source_subdir: null
+  converted_name: flappy_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train
+  mixed_converted_name: flappy_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  eval_action_classification: true
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 4
+  distributed_backend: none
+  is_resume: false
+  pretrained_checkpoint: null
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: ''
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: null
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: false
+      interval_steps: 100
+      latencies:
+      - 0
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: false
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  task: flappy
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: flappy
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: flappy_fix_latency_2_200ep
+output_dir: /workspace/starVLA/results/Checkpoints/flappy_fix_latency_2_200ep
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/config.yaml

@@ -0,0 +1,71 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: false
+  save_pt_file: false
+  sync:
+    enabled: false
+    keep_last_n: 0
+    repo_id: null
+datasets:
+  vla_data:
+    data_mix: flappy_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: flappy_train__bridge__val
+    obs_image_size: null
+    per_device_batch_size: 32
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 2
+    action_hidden_dim: 2560
+    action_horizon: 1
+    action_model_type: MLP
+    loss_type: discrete_ce
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+output_dir: /workspace/starVLA/results/Checkpoints/flappy_fix_latency_2_200ep
+rl_games:
+  env_eval:
+    enabled: true
+    mid_train:
+      enabled: false
+      interval_steps: 100
+  task: flappy
+run_id: flappy_fix_latency_2_200ep
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  distributed_backend: none
+  eval_action_classification: true
+  eval_interval: 100
+  eval_num_batches: 100
+  freeze_modules: ''
+  gradient_accumulation_steps: 4
+  gradient_clipping: 1.0
+  is_resume: false
+  learning_rate:
+    action_model: 0.0001
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: false
+    weight_decay: 1.0e-08
+  pretrained_checkpoint: null
+  reload_modules: null
+  save_interval: 100
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.6028500199317932,
+        0.3971499800682068,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.4890792667865753,
+        0.4890792667865753,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 330734,
+    "num_trajectories": 180
+  }
+}

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.7959861159324646,
+        0.2040138840675354,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.4030573070049286,
+        0.4030573070049286,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 72000,
+    "num_trajectories": 20
+  }
+}

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/hydra/.hydra/config.yaml

@@ -0,0 +1,197 @@
+framework:
+  qwenvl:
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    state_dim: 7
+    loss_type: discrete_ce
+    action_horizon: 1
+    future_action_window_size: 0
+    past_action_window_size: 0
+    action_dim: 7
+    action_env_dim: 2
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/flappy_fix_latency_2_200ep
+    data_mix: flappy_train__bridge
+    eval_data_mix: flappy_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ${dataset.single_source_hf}
+  config_name: null
+  source_subdir: null
+  converted_name: ${dataset.single_converted_name}
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train
+  mixed_converted_name: flappy_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: ${dataset.single_latency_filter}
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: ${bridge_base_model.repo_id.${model}}
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  eval_action_classification: true
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 4
+  distributed_backend: none
+  is_resume: false
+  pretrained_checkpoint: /workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: ${bridge_base_model.local_dir.${model}}
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: ''
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: null
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: false
+      interval_steps: 100
+      latencies:
+      - 0
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: false
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  task: flappy
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: flappy
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: flappy_fix_latency_2_200ep
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/hydra/.hydra/hydra.yaml

@@ -0,0 +1,252 @@
+hydra:
+  run:
+    dir: ${run_root_dir}/${run_id}/hydra
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - model=openvla
+    - env=flappy
+    - init=bridge
+    - mode=single
+    - run_id=flappy_fix_latency_2_200ep
+    - run_root_dir=/workspace/starVLA/results/Checkpoints
+    - seed=42
+    - wandb_entity=dongqianyu99-zhejiang-university
+    - wandb_project=starVLA_rl_games
+    - rl_games.env_eval.enabled=true
+    - checkpoint.sync.enabled=false
+    - checkpoint.sync.keep_last_n=0
+    - checkpoint.local.keep_last_n=1
+    - checkpoint.save_best_model=false
+    - checkpoint.save_pt_file=false
+    - trainer.is_resume=false
+    - trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+    - trainer.resume_step=0
+    - trainer.max_train_steps=5000
+    - trainer.num_warmup_steps=100
+    - trainer.save_interval=100
+    - trainer.eval_interval=100
+    - trainer.eval_num_batches=100
+    - trainer.logging_frequency=1
+    - trainer.gradient_accumulation_steps=4
+    - trainer.distributed_backend=none
+    - trainer.learning_rate.base=2e-05
+    - trainer.learning_rate.qwen_vl_interface=1e-05
+    - trainer.learning_rate.action_model=0.0001
+    - trainer.lr_scheduler_type=cosine_with_min_lr
+    - trainer.scheduler_specific_kwargs.min_lr=1e-06
+    - trainer.loss_scale.vla=1.0
+    - trainer.loss_scale.vlm=0.1
+    - trainer.max_grad_norm=1.0
+    - trainer.weight_decay=0.0
+    - trainer.gradient_clipping=1.0
+    - trainer.optimizer.name=AdamW
+    - trainer.optimizer.betas=[0.9, 0.95]
+    - trainer.optimizer.eps=1e-08
+    - trainer.optimizer.weight_decay=1e-08
+    - trainer.optimizer.fused=false
+    - trainer.save_format=pt
+    - framework.name=QwenOFT
+    - framework.qwenvl.attn_implementation=flash_attention_2
+    - framework.qwenvl.enable_gradient_checkpointing=true
+    - framework.action_model.action_dim=7
+    - framework.action_model.action_env_dim=2
+    - framework.action_model.state_dim=7
+    - framework.action_model.loss_type=discrete_ce
+    - framework.action_model.action_horizon=1
+    - framework.action_model.future_action_window_size=0
+    - framework.action_model.past_action_window_size=0
+    - datasets.vla_data.include_state=true
+    - datasets.vla_data.action_type=discrete
+    - datasets.vla_data.sequential_step_sampling=false
+    - datasets.vla_data.shuffle=true
+    - datasets.vla_data.action_balance.enabled=false
+    - datasets.vla_data.action_balance.strategy=balanced_epoch
+    - datasets.vla_data.action_balance.action_key=action_id
+    - datasets.vla_data.action_balance.target_flap_fraction=0.3
+    - datasets.vla_data.action_balance.noop_id=0
+    - datasets.vla_data.action_balance.flap_id=1
+    - datasets.vla_data.per_device_batch_size=32
+    - datasets.vla_data.load_all_data_for_training=true
+    - datasets.vla_data.video_backend=torchvision_av
+    - datasets.vla_data.data_root_dir=/workspace/starVLA/data/flappy_fix_latency_2_200ep
+    - datasets.vla_data.data_mix=flappy_train__bridge
+    - datasets.vla_data.eval_data_mix=flappy_train__bridge__val
+    - framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - rl_games.task=flappy
+    - rl_games.model_alias=openvla
+    - rl_games.initialization_mode=bridge
+    - rl_games.action_carrier=bridge
+    - rl_games.env_eval.distributed_mode=none
+    - rl_games.env_eval.latency.mode=single
+    - rl_games.env_eval.frameskip=1
+    - rl_games.env_eval.image_size=224
+    - rl_games.env_eval.seed=42
+    - rl_games.env_eval.fixed_episode_seeds=true
+    - rl_games.env_eval.latency_seed_stride=0
+    - rl_games.env_eval.task_seed_stride=0
+    - rl_games.env_eval.latency.values=[0]
+    - rl_games.env_eval.mid_train.enabled=false
+    - rl_games.env_eval.mid_train.interval_steps=100
+    - rl_games.env_eval.mid_train.num_episodes=5
+    - rl_games.env_eval.mid_train.max_steps_per_episode=3600
+    - rl_games.env_eval.mid_train.latencies=[0]
+    - rl_games.env_eval.post_train.enabled=false
+    - rl_games.env_eval.post_train.num_episodes=5
+    - rl_games.env_eval.post_train.max_steps_per_episode=3600
+    - rl_games.env_eval.post_train.latencies=[0,1,2,3,4]
+  job:
+    name: train_starvla_hydra
+    chdir: false
+    override_dirname: checkpoint.local.keep_last_n=1,checkpoint.save_best_model=false,checkpoint.save_pt_file=false,checkpoint.sync.enabled=false,checkpoint.sync.keep_last_n=0,datasets.vla_data.action_balance.action_key=action_id,datasets.vla_data.action_balance.enabled=false,datasets.vla_data.action_balance.flap_id=1,datasets.vla_data.action_balance.noop_id=0,datasets.vla_data.action_balance.strategy=balanced_epoch,datasets.vla_data.action_balance.target_flap_fraction=0.3,datasets.vla_data.action_type=discrete,datasets.vla_data.data_mix=flappy_train__bridge,datasets.vla_data.data_root_dir=/workspace/starVLA/data/flappy_fix_latency_2_200ep,datasets.vla_data.eval_data_mix=flappy_train__bridge__val,datasets.vla_data.include_state=true,datasets.vla_data.load_all_data_for_training=true,datasets.vla_data.per_device_batch_size=32,datasets.vla_data.sequential_step_sampling=false,datasets.vla_data.shuffle=true,datasets.vla_data.video_backend=torchvision_av,env=flappy,framework.action_model.action_dim=7,framework.action_model.action_env_dim=2,framework.action_model.action_horizon=1,framework.action_model.future_action_window_size=0,framework.action_model.loss_type=discrete_ce,framework.action_model.past_action_window_size=0,framework.action_model.state_dim=7,framework.name=QwenOFT,framework.qwenvl.attn_implementation=flash_attention_2,framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct,framework.qwenvl.enable_gradient_checkpointing=true,init=bridge,mode=single,model=openvla,rl_games.action_carrier=bridge,rl_games.env_eval.distributed_mode=none,rl_games.env_eval.enabled=true,rl_games.env_eval.fixed_episode_seeds=true,rl_games.env_eval.frameskip=1,rl_games.env_eval.image_size=224,rl_games.env_eval.latency.mode=single,rl_games.env_eval.latency.values=[0],rl_games.env_eval.latency_seed_stride=0,rl_games.env_eval.mid_train.enabled=false,rl_games.env_eval.mid_train.interval_steps=100,rl_games.env_eval.mid_train.latencies=[0],rl_games.env_eval.mid_train.max_steps_per_episode=3600,rl_games.env_eval.mid_train.num_episodes=5,rl_games.env_eval.post_train.enabled=false,rl_games.env_eval.post_train.latencies=[0,1,2,3,4],rl_games.env_eval.post_train.max_steps_per_episode=3600,rl_games.env_eval.post_train.num_episodes=5,rl_games.env_eval.seed=42,rl_games.env_eval.task_seed_stride=0,rl_games.initialization_mode=bridge,rl_games.model_alias=openvla,rl_games.task=flappy,run_id=flappy_fix_latency_2_200ep,run_root_dir=/workspace/starVLA/results/Checkpoints,seed=42,trainer.distributed_backend=none,trainer.eval_interval=100,trainer.eval_num_batches=100,trainer.gradient_accumulation_steps=4,trainer.gradient_clipping=1.0,trainer.is_resume=false,trainer.learning_rate.action_model=0.0001,trainer.learning_rate.base=2e-05,trainer.learning_rate.qwen_vl_interface=1e-05,trainer.logging_frequency=1,trainer.loss_scale.vla=1.0,trainer.loss_scale.vlm=0.1,trainer.lr_scheduler_type=cosine_with_min_lr,trainer.max_grad_norm=1.0,trainer.max_train_steps=5000,trainer.num_warmup_steps=100,trainer.optimizer.betas=[0.9,
+      0.95],trainer.optimizer.eps=1e-08,trainer.optimizer.fused=false,trainer.optimizer.name=AdamW,trainer.optimizer.weight_decay=1e-08,trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt,trainer.resume_step=0,trainer.save_format=pt,trainer.save_interval=100,trainer.scheduler_specific_kwargs.min_lr=1e-06,trainer.weight_decay=0.0,wandb_entity=dongqianyu99-zhejiang-university,wandb_project=starVLA_rl_games
+    id: ???
+    num: ???
+    config_name: train
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.3
+    version_base: '1.1'
+    cwd: /workspace/starVLA
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/starVLA/examples/rl_games/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/starVLA/results/Checkpoints/flappy_fix_latency_2_200ep/hydra
+    choices:
+      cross_task_setup: null
+      checkpoint: default
+      mode: single
+      init: bridge
+      env: flappy
+      model: openvla
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/hydra/.hydra/overrides.yaml

@@ -0,0 +1,91 @@
+- model=openvla
+- env=flappy
+- init=bridge
+- mode=single
+- run_id=flappy_fix_latency_2_200ep
+- run_root_dir=/workspace/starVLA/results/Checkpoints
+- seed=42
+- wandb_entity=dongqianyu99-zhejiang-university
+- wandb_project=starVLA_rl_games
+- rl_games.env_eval.enabled=true
+- checkpoint.sync.enabled=false
+- checkpoint.sync.keep_last_n=0
+- checkpoint.local.keep_last_n=1
+- checkpoint.save_best_model=false
+- checkpoint.save_pt_file=false
+- trainer.is_resume=false
+- trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+- trainer.resume_step=0
+- trainer.max_train_steps=5000
+- trainer.num_warmup_steps=100
+- trainer.save_interval=100
+- trainer.eval_interval=100
+- trainer.eval_num_batches=100
+- trainer.logging_frequency=1
+- trainer.gradient_accumulation_steps=4
+- trainer.distributed_backend=none
+- trainer.learning_rate.base=2e-05
+- trainer.learning_rate.qwen_vl_interface=1e-05
+- trainer.learning_rate.action_model=0.0001
+- trainer.lr_scheduler_type=cosine_with_min_lr
+- trainer.scheduler_specific_kwargs.min_lr=1e-06
+- trainer.loss_scale.vla=1.0
+- trainer.loss_scale.vlm=0.1
+- trainer.max_grad_norm=1.0
+- trainer.weight_decay=0.0
+- trainer.gradient_clipping=1.0
+- trainer.optimizer.name=AdamW
+- trainer.optimizer.betas=[0.9, 0.95]
+- trainer.optimizer.eps=1e-08
+- trainer.optimizer.weight_decay=1e-08
+- trainer.optimizer.fused=false
+- trainer.save_format=pt
+- framework.name=QwenOFT
+- framework.qwenvl.attn_implementation=flash_attention_2
+- framework.qwenvl.enable_gradient_checkpointing=true
+- framework.action_model.action_dim=7
+- framework.action_model.action_env_dim=2
+- framework.action_model.state_dim=7
+- framework.action_model.loss_type=discrete_ce
+- framework.action_model.action_horizon=1
+- framework.action_model.future_action_window_size=0
+- framework.action_model.past_action_window_size=0
+- datasets.vla_data.include_state=true
+- datasets.vla_data.action_type=discrete
+- datasets.vla_data.sequential_step_sampling=false
+- datasets.vla_data.shuffle=true
+- datasets.vla_data.action_balance.enabled=false
+- datasets.vla_data.action_balance.strategy=balanced_epoch
+- datasets.vla_data.action_balance.action_key=action_id
+- datasets.vla_data.action_balance.target_flap_fraction=0.3
+- datasets.vla_data.action_balance.noop_id=0
+- datasets.vla_data.action_balance.flap_id=1
+- datasets.vla_data.per_device_batch_size=32
+- datasets.vla_data.load_all_data_for_training=true
+- datasets.vla_data.video_backend=torchvision_av
+- datasets.vla_data.data_root_dir=/workspace/starVLA/data/flappy_fix_latency_2_200ep
+- datasets.vla_data.data_mix=flappy_train__bridge
+- datasets.vla_data.eval_data_mix=flappy_train__bridge__val
+- framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- rl_games.task=flappy
+- rl_games.model_alias=openvla
+- rl_games.initialization_mode=bridge
+- rl_games.action_carrier=bridge
+- rl_games.env_eval.distributed_mode=none
+- rl_games.env_eval.latency.mode=single
+- rl_games.env_eval.frameskip=1
+- rl_games.env_eval.image_size=224
+- rl_games.env_eval.seed=42
+- rl_games.env_eval.fixed_episode_seeds=true
+- rl_games.env_eval.latency_seed_stride=0
+- rl_games.env_eval.task_seed_stride=0
+- rl_games.env_eval.latency.values=[0]
+- rl_games.env_eval.mid_train.enabled=false
+- rl_games.env_eval.mid_train.interval_steps=100
+- rl_games.env_eval.mid_train.num_episodes=5
+- rl_games.env_eval.mid_train.max_steps_per_episode=3600
+- rl_games.env_eval.mid_train.latencies=[0]
+- rl_games.env_eval.post_train.enabled=false
+- rl_games.env_eval.post_train.num_episodes=5
+- rl_games.env_eval.post_train.max_steps_per_episode=3600
+- rl_games.env_eval.post_train.latencies=[0,1,2,3,4]

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/summary.jsonl

@@ -0,0 +1,50 @@
+{"steps": 100}
+{"steps": 200}
+{"steps": 300}
+{"steps": 400}
+{"steps": 500}
+{"steps": 600}
+{"steps": 700}
+{"steps": 800}
+{"steps": 900}
+{"steps": 1000}
+{"steps": 1100}
+{"steps": 1200}
+{"steps": 1300}
+{"steps": 1400}
+{"steps": 1500}
+{"steps": 1600}
+{"steps": 1700}
+{"steps": 1800}
+{"steps": 1900}
+{"steps": 2000}
+{"steps": 2100}
+{"steps": 2200}
+{"steps": 2300}
+{"steps": 2400}
+{"steps": 2500}
+{"steps": 2600}
+{"steps": 2700}
+{"steps": 2800}
+{"steps": 2900}
+{"steps": 3000}
+{"steps": 3100}
+{"steps": 3200}
+{"steps": 3300}
+{"steps": 3400}
+{"steps": 3500}
+{"steps": 3600}
+{"steps": 3700}
+{"steps": 3800}
+{"steps": 3900}
+{"steps": 4000}
+{"steps": 4100}
+{"steps": 4200}
+{"steps": 4300}
+{"steps": 4400}
+{"steps": 4500}
+{"steps": 4600}
+{"steps": 4700}
+{"steps": 4800}
+{"steps": 4900}
+{"steps": 5000}

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/checkpoints/steps_5000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5da01d6eeb7a46861da90b0a8dbc364a33f92d542d25079bcc86280b10617418
+size 9138230516

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/config.full.yaml

@@ -0,0 +1,199 @@
+framework:
+  name: QwenOFT
+  qwenvl:
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    action_hidden_dim: 2560
+    future_action_window_size: 0
+    past_action_window_size: 0
+    loss_type: discrete_ce
+    state_dim: 7
+    action_horizon: 1
+    action_env_dim: 2
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/flappy_fix_latency_3_200ep
+    data_mix: flappy_train__bridge
+    eval_data_mix: flappy_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  config_name: null
+  source_subdir: null
+  converted_name: flappy_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train
+  mixed_converted_name: flappy_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  eval_action_classification: true
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 4
+  distributed_backend: none
+  is_resume: false
+  pretrained_checkpoint: null
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: ''
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: null
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: false
+      interval_steps: 100
+      latencies:
+      - 0
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: false
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  task: flappy
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: flappy
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: flappy_fix_latency_3_200ep
+output_dir: /workspace/starVLA/results/Checkpoints/flappy_fix_latency_3_200ep
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/config.yaml

@@ -0,0 +1,71 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: false
+  save_pt_file: false
+  sync:
+    enabled: false
+    keep_last_n: 0
+    repo_id: null
+datasets:
+  vla_data:
+    data_mix: flappy_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: flappy_train__bridge__val
+    obs_image_size: null
+    per_device_batch_size: 32
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 2
+    action_hidden_dim: 2560
+    action_horizon: 1
+    action_model_type: MLP
+    loss_type: discrete_ce
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+output_dir: /workspace/starVLA/results/Checkpoints/flappy_fix_latency_3_200ep
+rl_games:
+  env_eval:
+    enabled: true
+    mid_train:
+      enabled: false
+      interval_steps: 100
+  task: flappy
+run_id: flappy_fix_latency_3_200ep
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  distributed_backend: none
+  eval_action_classification: true
+  eval_interval: 100
+  eval_num_batches: 100
+  freeze_modules: ''
+  gradient_accumulation_steps: 4
+  gradient_clipping: 1.0
+  is_resume: false
+  learning_rate:
+    action_model: 0.0001
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: false
+    weight_decay: 1.0e-08
+  pretrained_checkpoint: null
+  reload_modules: null
+  save_interval: 100
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.5971351861953735,
+        0.40286484360694885,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.49107804894447327,
+        0.49107804894447327,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 316528,
+    "num_trajectories": 180
+  }
+}

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.8040833473205566,
+        0.19591666758060455,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.39686188101768494,
+        0.39686188101768494,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 72000,
+    "num_trajectories": 20
+  }
+}

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/hydra/.hydra/config.yaml

@@ -0,0 +1,197 @@
+framework:
+  qwenvl:
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    state_dim: 7
+    loss_type: discrete_ce
+    action_horizon: 1
+    future_action_window_size: 0
+    past_action_window_size: 0
+    action_dim: 7
+    action_env_dim: 2
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/flappy_fix_latency_3_200ep
+    data_mix: flappy_train__bridge
+    eval_data_mix: flappy_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    per_device_batch_size: 32
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ${dataset.single_source_hf}
+  config_name: null
+  source_subdir: null
+  converted_name: ${dataset.single_converted_name}
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train
+  mixed_converted_name: flappy_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  verify_rows: 200
+  max_episodes: null
+  episodes_per_latency: null
+  latency_filter: ${dataset.single_latency_filter}
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: ${bridge_base_model.repo_id.${model}}
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-OFT-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_5000_pytorch_model.pt
+trainer:
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  eval_action_classification: true
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.0e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 4
+  distributed_backend: none
+  is_resume: false
+  pretrained_checkpoint: /workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: ${bridge_base_model.local_dir.${model}}
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: openvla
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: ''
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: null
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: false
+      interval_steps: 100
+      latencies:
+      - 0
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: false
+      latencies:
+      - 0
+      - 1
+      - 2
+      - 3
+      - 4
+      num_episodes: 5
+      max_steps_per_episode: 3600
+  task: flappy
+  initialization_mode: bridge
+  action_carrier: bridge
+model: openvla
+env: flappy
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: flappy_fix_latency_3_200ep
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/hydra/.hydra/hydra.yaml

@@ -0,0 +1,252 @@
+hydra:
+  run:
+    dir: ${run_root_dir}/${run_id}/hydra
+  sweep:
+    dir: multirun/${now:%Y-%m-%d}/${now:%H-%M-%S}
+    subdir: ${hydra.job.num}
+  launcher:
+    _target_: hydra._internal.core_plugins.basic_launcher.BasicLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - model=openvla
+    - env=flappy
+    - init=bridge
+    - mode=single
+    - run_id=flappy_fix_latency_3_200ep
+    - run_root_dir=/workspace/starVLA/results/Checkpoints
+    - seed=42
+    - wandb_entity=dongqianyu99-zhejiang-university
+    - wandb_project=starVLA_rl_games
+    - rl_games.env_eval.enabled=true
+    - checkpoint.sync.enabled=false
+    - checkpoint.sync.keep_last_n=0
+    - checkpoint.local.keep_last_n=1
+    - checkpoint.save_best_model=false
+    - checkpoint.save_pt_file=false
+    - trainer.is_resume=false
+    - trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+    - trainer.resume_step=0
+    - trainer.max_train_steps=5000
+    - trainer.num_warmup_steps=100
+    - trainer.save_interval=100
+    - trainer.eval_interval=100
+    - trainer.eval_num_batches=100
+    - trainer.logging_frequency=1
+    - trainer.gradient_accumulation_steps=4
+    - trainer.distributed_backend=none
+    - trainer.learning_rate.base=2e-05
+    - trainer.learning_rate.qwen_vl_interface=1e-05
+    - trainer.learning_rate.action_model=0.0001
+    - trainer.lr_scheduler_type=cosine_with_min_lr
+    - trainer.scheduler_specific_kwargs.min_lr=1e-06
+    - trainer.loss_scale.vla=1.0
+    - trainer.loss_scale.vlm=0.1
+    - trainer.max_grad_norm=1.0
+    - trainer.weight_decay=0.0
+    - trainer.gradient_clipping=1.0
+    - trainer.optimizer.name=AdamW
+    - trainer.optimizer.betas=[0.9, 0.95]
+    - trainer.optimizer.eps=1e-08
+    - trainer.optimizer.weight_decay=1e-08
+    - trainer.optimizer.fused=false
+    - trainer.save_format=pt
+    - framework.name=QwenOFT
+    - framework.qwenvl.attn_implementation=flash_attention_2
+    - framework.qwenvl.enable_gradient_checkpointing=true
+    - framework.action_model.action_dim=7
+    - framework.action_model.action_env_dim=2
+    - framework.action_model.state_dim=7
+    - framework.action_model.loss_type=discrete_ce
+    - framework.action_model.action_horizon=1
+    - framework.action_model.future_action_window_size=0
+    - framework.action_model.past_action_window_size=0
+    - datasets.vla_data.include_state=true
+    - datasets.vla_data.action_type=discrete
+    - datasets.vla_data.sequential_step_sampling=false
+    - datasets.vla_data.shuffle=true
+    - datasets.vla_data.action_balance.enabled=false
+    - datasets.vla_data.action_balance.strategy=balanced_epoch
+    - datasets.vla_data.action_balance.action_key=action_id
+    - datasets.vla_data.action_balance.target_flap_fraction=0.3
+    - datasets.vla_data.action_balance.noop_id=0
+    - datasets.vla_data.action_balance.flap_id=1
+    - datasets.vla_data.per_device_batch_size=32
+    - datasets.vla_data.load_all_data_for_training=true
+    - datasets.vla_data.video_backend=torchvision_av
+    - datasets.vla_data.data_root_dir=/workspace/starVLA/data/flappy_fix_latency_3_200ep
+    - datasets.vla_data.data_mix=flappy_train__bridge
+    - datasets.vla_data.eval_data_mix=flappy_train__bridge__val
+    - framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - rl_games.task=flappy
+    - rl_games.model_alias=openvla
+    - rl_games.initialization_mode=bridge
+    - rl_games.action_carrier=bridge
+    - rl_games.env_eval.distributed_mode=none
+    - rl_games.env_eval.latency.mode=single
+    - rl_games.env_eval.frameskip=1
+    - rl_games.env_eval.image_size=224
+    - rl_games.env_eval.seed=42
+    - rl_games.env_eval.fixed_episode_seeds=true
+    - rl_games.env_eval.latency_seed_stride=0
+    - rl_games.env_eval.task_seed_stride=0
+    - rl_games.env_eval.latency.values=[0]
+    - rl_games.env_eval.mid_train.enabled=false
+    - rl_games.env_eval.mid_train.interval_steps=100
+    - rl_games.env_eval.mid_train.num_episodes=5
+    - rl_games.env_eval.mid_train.max_steps_per_episode=3600
+    - rl_games.env_eval.mid_train.latencies=[0]
+    - rl_games.env_eval.post_train.enabled=false
+    - rl_games.env_eval.post_train.num_episodes=5
+    - rl_games.env_eval.post_train.max_steps_per_episode=3600
+    - rl_games.env_eval.post_train.latencies=[0,1,2,3,4]
+  job:
+    name: train_starvla_hydra
+    chdir: false
+    override_dirname: checkpoint.local.keep_last_n=1,checkpoint.save_best_model=false,checkpoint.save_pt_file=false,checkpoint.sync.enabled=false,checkpoint.sync.keep_last_n=0,datasets.vla_data.action_balance.action_key=action_id,datasets.vla_data.action_balance.enabled=false,datasets.vla_data.action_balance.flap_id=1,datasets.vla_data.action_balance.noop_id=0,datasets.vla_data.action_balance.strategy=balanced_epoch,datasets.vla_data.action_balance.target_flap_fraction=0.3,datasets.vla_data.action_type=discrete,datasets.vla_data.data_mix=flappy_train__bridge,datasets.vla_data.data_root_dir=/workspace/starVLA/data/flappy_fix_latency_3_200ep,datasets.vla_data.eval_data_mix=flappy_train__bridge__val,datasets.vla_data.include_state=true,datasets.vla_data.load_all_data_for_training=true,datasets.vla_data.per_device_batch_size=32,datasets.vla_data.sequential_step_sampling=false,datasets.vla_data.shuffle=true,datasets.vla_data.video_backend=torchvision_av,env=flappy,framework.action_model.action_dim=7,framework.action_model.action_env_dim=2,framework.action_model.action_horizon=1,framework.action_model.future_action_window_size=0,framework.action_model.loss_type=discrete_ce,framework.action_model.past_action_window_size=0,framework.action_model.state_dim=7,framework.name=QwenOFT,framework.qwenvl.attn_implementation=flash_attention_2,framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct,framework.qwenvl.enable_gradient_checkpointing=true,init=bridge,mode=single,model=openvla,rl_games.action_carrier=bridge,rl_games.env_eval.distributed_mode=none,rl_games.env_eval.enabled=true,rl_games.env_eval.fixed_episode_seeds=true,rl_games.env_eval.frameskip=1,rl_games.env_eval.image_size=224,rl_games.env_eval.latency.mode=single,rl_games.env_eval.latency.values=[0],rl_games.env_eval.latency_seed_stride=0,rl_games.env_eval.mid_train.enabled=false,rl_games.env_eval.mid_train.interval_steps=100,rl_games.env_eval.mid_train.latencies=[0],rl_games.env_eval.mid_train.max_steps_per_episode=3600,rl_games.env_eval.mid_train.num_episodes=5,rl_games.env_eval.post_train.enabled=false,rl_games.env_eval.post_train.latencies=[0,1,2,3,4],rl_games.env_eval.post_train.max_steps_per_episode=3600,rl_games.env_eval.post_train.num_episodes=5,rl_games.env_eval.seed=42,rl_games.env_eval.task_seed_stride=0,rl_games.initialization_mode=bridge,rl_games.model_alias=openvla,rl_games.task=flappy,run_id=flappy_fix_latency_3_200ep,run_root_dir=/workspace/starVLA/results/Checkpoints,seed=42,trainer.distributed_backend=none,trainer.eval_interval=100,trainer.eval_num_batches=100,trainer.gradient_accumulation_steps=4,trainer.gradient_clipping=1.0,trainer.is_resume=false,trainer.learning_rate.action_model=0.0001,trainer.learning_rate.base=2e-05,trainer.learning_rate.qwen_vl_interface=1e-05,trainer.logging_frequency=1,trainer.loss_scale.vla=1.0,trainer.loss_scale.vlm=0.1,trainer.lr_scheduler_type=cosine_with_min_lr,trainer.max_grad_norm=1.0,trainer.max_train_steps=5000,trainer.num_warmup_steps=100,trainer.optimizer.betas=[0.9,
+      0.95],trainer.optimizer.eps=1e-08,trainer.optimizer.fused=false,trainer.optimizer.name=AdamW,trainer.optimizer.weight_decay=1e-08,trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt,trainer.resume_step=0,trainer.save_format=pt,trainer.save_interval=100,trainer.scheduler_specific_kwargs.min_lr=1e-06,trainer.weight_decay=0.0,wandb_entity=dongqianyu99-zhejiang-university,wandb_project=starVLA_rl_games
+    id: ???
+    num: ???
+    config_name: train
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.3
+    version_base: '1.1'
+    cwd: /workspace/starVLA
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/starVLA/examples/rl_games/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/starVLA/results/Checkpoints/flappy_fix_latency_3_200ep/hydra
+    choices:
+      cross_task_setup: null
+      checkpoint: default
+      mode: single
+      init: bridge
+      env: flappy
+      model: openvla
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: basic
+      hydra/output: default
+  verbose: false

latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/hydra/.hydra/overrides.yaml

@@ -0,0 +1,91 @@
+- model=openvla
+- env=flappy
+- init=bridge
+- mode=single
+- run_id=flappy_fix_latency_3_200ep
+- run_root_dir=/workspace/starVLA/results/Checkpoints
+- seed=42
+- wandb_entity=dongqianyu99-zhejiang-university
+- wandb_project=starVLA_rl_games
+- rl_games.env_eval.enabled=true
+- checkpoint.sync.enabled=false
+- checkpoint.sync.keep_last_n=0
+- checkpoint.local.keep_last_n=1
+- checkpoint.save_best_model=false
+- checkpoint.save_pt_file=false
+- trainer.is_resume=false
+- trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt
+- trainer.resume_step=0
+- trainer.max_train_steps=5000
+- trainer.num_warmup_steps=100
+- trainer.save_interval=100
+- trainer.eval_interval=100
+- trainer.eval_num_batches=100
+- trainer.logging_frequency=1
+- trainer.gradient_accumulation_steps=4
+- trainer.distributed_backend=none
+- trainer.learning_rate.base=2e-05
+- trainer.learning_rate.qwen_vl_interface=1e-05
+- trainer.learning_rate.action_model=0.0001
+- trainer.lr_scheduler_type=cosine_with_min_lr
+- trainer.scheduler_specific_kwargs.min_lr=1e-06
+- trainer.loss_scale.vla=1.0
+- trainer.loss_scale.vlm=0.1
+- trainer.max_grad_norm=1.0
+- trainer.weight_decay=0.0
+- trainer.gradient_clipping=1.0
+- trainer.optimizer.name=AdamW
+- trainer.optimizer.betas=[0.9, 0.95]
+- trainer.optimizer.eps=1e-08
+- trainer.optimizer.weight_decay=1e-08
+- trainer.optimizer.fused=false
+- trainer.save_format=pt
+- framework.name=QwenOFT
+- framework.qwenvl.attn_implementation=flash_attention_2
+- framework.qwenvl.enable_gradient_checkpointing=true
+- framework.action_model.action_dim=7
+- framework.action_model.action_env_dim=2
+- framework.action_model.state_dim=7
+- framework.action_model.loss_type=discrete_ce
+- framework.action_model.action_horizon=1
+- framework.action_model.future_action_window_size=0
+- framework.action_model.past_action_window_size=0
+- datasets.vla_data.include_state=true
+- datasets.vla_data.action_type=discrete
+- datasets.vla_data.sequential_step_sampling=false
+- datasets.vla_data.shuffle=true
+- datasets.vla_data.action_balance.enabled=false
+- datasets.vla_data.action_balance.strategy=balanced_epoch
+- datasets.vla_data.action_balance.action_key=action_id
+- datasets.vla_data.action_balance.target_flap_fraction=0.3
+- datasets.vla_data.action_balance.noop_id=0
+- datasets.vla_data.action_balance.flap_id=1
+- datasets.vla_data.per_device_batch_size=32
+- datasets.vla_data.load_all_data_for_training=true
+- datasets.vla_data.video_backend=torchvision_av
+- datasets.vla_data.data_root_dir=/workspace/starVLA/data/flappy_fix_latency_3_200ep
+- datasets.vla_data.data_mix=flappy_train__bridge
+- datasets.vla_data.eval_data_mix=flappy_train__bridge__val
+- framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- rl_games.task=flappy
+- rl_games.model_alias=openvla
+- rl_games.initialization_mode=bridge
+- rl_games.action_carrier=bridge
+- rl_games.env_eval.distributed_mode=none
+- rl_games.env_eval.latency.mode=single
+- rl_games.env_eval.frameskip=1
+- rl_games.env_eval.image_size=224
+- rl_games.env_eval.seed=42
+- rl_games.env_eval.fixed_episode_seeds=true
+- rl_games.env_eval.latency_seed_stride=0
+- rl_games.env_eval.task_seed_stride=0
+- rl_games.env_eval.latency.values=[0]
+- rl_games.env_eval.mid_train.enabled=false
+- rl_games.env_eval.mid_train.interval_steps=100
+- rl_games.env_eval.mid_train.num_episodes=5
+- rl_games.env_eval.mid_train.max_steps_per_episode=3600
+- rl_games.env_eval.mid_train.latencies=[0]
+- rl_games.env_eval.post_train.enabled=false
+- rl_games.env_eval.post_train.num_episodes=5
+- rl_games.env_eval.post_train.max_steps_per_episode=3600
+- rl_games.env_eval.post_train.latencies=[0,1,2,3,4]