Consolidate paper experiment assets by task and run (151-300)

latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/train/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/checkpoint_p0/checkpoint_000048836_25004032.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6786d84ce663078e7cf02e7032035c4d899843fe06ca3d6d07e666c8d606644f
+size 13352656

latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/train/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/config.json

@@ -0,0 +1,260 @@
+{
+  "help": false,
+  "algo": "APPO",
+  "env": "doom_deadly_corridor",
+  "experiment": "deadly_corridor_teacher_fixed_l6_obs8p75_seed0",
+  "train_dir": "results/checkpoints/memory/rollout_teacher",
+  "restart_behavior": "resume",
+  "device": "gpu",
+  "seed": 0,
+  "num_policies": 1,
+  "async_rl": true,
+  "serial_mode": false,
+  "batched_sampling": false,
+  "num_batches_to_accumulate": 2,
+  "worker_num_splits": 2,
+  "policy_workers_per_policy": 1,
+  "max_policy_lag": 1000,
+  "num_workers": 32,
+  "num_envs_per_worker": 4,
+  "batch_size": 1024,
+  "num_batches_per_epoch": 2,
+  "num_epochs": 2,
+  "rollout": 128,
+  "recurrence": 128,
+  "shuffle_minibatches": false,
+  "gamma": 0.99,
+  "reward_scale": 1.0,
+  "reward_clip": 1000.0,
+  "value_bootstrap": false,
+  "normalize_returns": true,
+  "exploration_loss_coeff": 0.001,
+  "value_loss_coeff": 0.5,
+  "kl_loss_coeff": 0.0,
+  "exploration_loss": "symmetric_kl",
+  "gae_lambda": 0.95,
+  "ppo_clip_ratio": 0.1,
+  "ppo_clip_value": 0.2,
+  "with_vtrace": false,
+  "vtrace_rho": 1.0,
+  "vtrace_c": 1.0,
+  "optimizer": "adam",
+  "adam_eps": 1e-06,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "max_grad_norm": 4.0,
+  "learning_rate": 0.0001,
+  "lr_schedule": "constant",
+  "lr_schedule_kl_threshold": 0.008,
+  "lr_adaptive_min": 1e-06,
+  "lr_adaptive_max": 0.01,
+  "obs_subtract_mean": 0.0,
+  "obs_scale": 255.0,
+  "normalize_input": true,
+  "normalize_input_keys": null,
+  "decorrelate_experience_max_seconds": 0,
+  "decorrelate_envs_on_one_worker": true,
+  "actor_worker_gpus": [],
+  "set_workers_cpu_affinity": true,
+  "force_envs_single_thread": false,
+  "default_niceness": 0,
+  "log_to_file": true,
+  "experiment_summaries_interval": 1,
+  "flush_summaries_interval": 30,
+  "stats_avg": 100,
+  "summaries_use_frameskip": true,
+  "heartbeat_interval": 20,
+  "heartbeat_reporting_interval": 600,
+  "train_for_env_steps": 25000000,
+  "train_for_seconds": 10000000000,
+  "save_every_sec": 600,
+  "keep_checkpoints": 5,
+  "load_checkpoint_kind": "latest",
+  "save_milestones_sec": -1,
+  "save_best_every_sec": 5,
+  "save_best_metric": "reward",
+  "save_best_after": 100000,
+  "benchmark": false,
+  "encoder_mlp_layers": [
+    512,
+    512
+  ],
+  "encoder_conv_architecture": "convnet_simple",
+  "encoder_conv_mlp_layers": [
+    512
+  ],
+  "use_rnn": true,
+  "rnn_size": 512,
+  "rnn_type": "gru",
+  "rnn_num_layers": 1,
+  "decoder_mlp_layers": [],
+  "nonlinearity": "elu",
+  "policy_initialization": "orthogonal",
+  "policy_init_gain": 1.0,
+  "actor_critic_share_weights": true,
+  "adaptive_stddev": true,
+  "continuous_tanh_scale": 0.0,
+  "initial_stddev": 1.0,
+  "use_env_info_cache": false,
+  "env_gpu_actions": false,
+  "env_gpu_observations": true,
+  "env_frameskip": 1,
+  "env_framestack": 1,
+  "pixel_format": "CHW",
+  "use_record_episode_statistics": false,
+  "with_wandb": true,
+  "wandb_user": null,
+  "wandb_project": "latency-sensitive-bench",
+  "wandb_group": "deadly-corridor-obs8p75-fixed_l6",
+  "wandb_job_type": "sample_factory",
+  "wandb_tags": [
+    "memory_rollout_teacher",
+    "deadly-corridor",
+    "teacher",
+    "fixed",
+    "fixed_l6",
+    "obs8p75",
+    "seed0"
+  ],
+  "with_pbt": false,
+  "pbt_mix_policies_in_one_env": true,
+  "pbt_period_env_steps": 5000000,
+  "pbt_start_mutation": 20000000,
+  "pbt_replace_fraction": 0.3,
+  "pbt_mutation_rate": 0.15,
+  "pbt_replace_reward_gap": 0.1,
+  "pbt_replace_reward_gap_absolute": 1e-06,
+  "pbt_optimize_gamma": false,
+  "pbt_target_objective": "true_objective",
+  "pbt_perturb_min": 1.1,
+  "pbt_perturb_max": 1.5,
+  "fps": 35,
+  "eval_env_frameskip": 1,
+  "no_render": false,
+  "save_video": false,
+  "video_frames": 1000000000.0,
+  "video_name": null,
+  "max_num_frames": 1000000000.0,
+  "max_num_episodes": 1000000000.0,
+  "push_to_hub": false,
+  "hf_repository": null,
+  "policy_index": 0,
+  "eval_deterministic": true,
+  "train_script": null,
+  "enjoy_script": null,
+  "num_agents": -1,
+  "num_humans": 0,
+  "num_bots": -1,
+  "start_bot_difficulty": null,
+  "timelimit": null,
+  "res_w": 128,
+  "res_h": 72,
+  "wide_aspect_ratio": false,
+  "record_to": null,
+  "mode": "train",
+  "env_fps": 35.0,
+  "obs_fps": 8.75,
+  "latency_type": "fixed",
+  "fixed_latency_ms": 171.42857142857144,
+  "mean_latency_ms": null,
+  "std_latency_ms": null,
+  "min_latency_ms": null,
+  "max_latency_ms": null,
+  "latency_seed": null,
+  "add_latency_info": false,
+  "max_pending_actions": null,
+  "hold_policy": "hold",
+  "ordering_policy": "latest_ready",
+  "eval_episodes": 100,
+  "eval_parallel_envs": 1,
+  "eval_latency_raw_frame_values": "0,1,2,3,4,5,10,15,40",
+  "eval_max_steps": 30000,
+  "eval_raw_reward": true,
+  "disable_reward_scaling": true,
+  "frame_stack": 4,
+  "max_episode_steps": 0,
+  "simulator": "cpu",
+  "episode_metrics_path": "results/checkpoints/memory/rollout_teacher/deadly_corridor_teacher_fixed_l6_obs8p75_seed0/episode_metrics.jsonl",
+  "vizdoom_init_max_parallel": 10,
+  "vizdoom_init_lock_timeout_sec": 5.0,
+  "vizdoom_init_lock_max_wait_sec": 60.0,
+  "vizdoom_init_retry_jitter_sec": 0.25,
+  "command_line": "--mode train --algo APPO --env doom_deadly_corridor --experiment deadly_corridor_teacher_fixed_l6_obs8p75_seed0 --train_dir results/checkpoints/memory/rollout_teacher --restart_behavior resume --device gpu --seed 0 --episode_metrics_path results/checkpoints/memory/rollout_teacher/deadly_corridor_teacher_fixed_l6_obs8p75_seed0/episode_metrics.jsonl --train_for_env_steps 25000000 --num_workers 32 --num_envs_per_worker 4 --num_policies 1 --batch_size 1024 --rollout 128 --recurrence 128 --num_epochs 2 --num_batches_per_epoch 2 --worker_num_splits 2 --learning_rate 0.0001 --gamma 0.99 --gae_lambda 0.95 --ppo_clip_ratio 0.1 --ppo_clip_value 0.2 --exploration_loss symmetric_kl --exploration_loss_coeff 0.001 --value_loss_coeff 0.5 --max_grad_norm 4.0 --rnn_type gru --rnn_size 512 --save_every_sec 600 --keep_checkpoints 5 --stats_avg 100 --experiment_summaries_interval 1 --async_rl True --use_rnn True --normalize_returns True --normalize_input True --latency-type fixed --fixed-latency-ms 171.42857142857144 --add-latency-info False --eval-episodes 100 --eval-parallel-envs 1 --eval-max-steps 30000 --eval-deterministic True --eval-raw-reward --with_wandb True --wandb_project latency-sensitive-bench --wandb_group deadly-corridor-obs8p75-fixed_l6 --wandb_job_type sample_factory --wandb_tags memory_rollout_teacher deadly-corridor teacher fixed fixed_l6 obs8p75 seed0 --env-fps 35 --obs-fps 8.75 --res_w 128 --res_h 72 --wide_aspect_ratio False --frame_stack 4 --simulator cpu --disable-reward-scaling True --hold-policy hold --ordering-policy latest_ready",
+  "cli_args": {
+    "algo": "APPO",
+    "env": "doom_deadly_corridor",
+    "experiment": "deadly_corridor_teacher_fixed_l6_obs8p75_seed0",
+    "train_dir": "results/checkpoints/memory/rollout_teacher",
+    "restart_behavior": "resume",
+    "device": "gpu",
+    "seed": 0,
+    "num_policies": 1,
+    "async_rl": true,
+    "worker_num_splits": 2,
+    "num_workers": 32,
+    "num_envs_per_worker": 4,
+    "batch_size": 1024,
+    "num_batches_per_epoch": 2,
+    "num_epochs": 2,
+    "rollout": 128,
+    "recurrence": 128,
+    "gamma": 0.99,
+    "normalize_returns": true,
+    "exploration_loss_coeff": 0.001,
+    "value_loss_coeff": 0.5,
+    "exploration_loss": "symmetric_kl",
+    "gae_lambda": 0.95,
+    "ppo_clip_ratio": 0.1,
+    "ppo_clip_value": 0.2,
+    "max_grad_norm": 4.0,
+    "learning_rate": 0.0001,
+    "normalize_input": true,
+    "experiment_summaries_interval": 1,
+    "stats_avg": 100,
+    "train_for_env_steps": 25000000,
+    "save_every_sec": 600,
+    "keep_checkpoints": 5,
+    "use_rnn": true,
+    "rnn_size": 512,
+    "rnn_type": "gru",
+    "with_wandb": true,
+    "wandb_project": "latency-sensitive-bench",
+    "wandb_group": "deadly-corridor-obs8p75-fixed_l6",
+    "wandb_job_type": "sample_factory",
+    "wandb_tags": [
+      "memory_rollout_teacher",
+      "deadly-corridor",
+      "teacher",
+      "fixed",
+      "fixed_l6",
+      "obs8p75",
+      "seed0"
+    ],
+    "eval_deterministic": true,
+    "res_w": 128,
+    "res_h": 72,
+    "wide_aspect_ratio": false,
+    "mode": "train",
+    "env_fps": 35.0,
+    "obs_fps": 8.75,
+    "latency_type": "fixed",
+    "fixed_latency_ms": 171.42857142857144,
+    "add_latency_info": false,
+    "hold_policy": "hold",
+    "ordering_policy": "latest_ready",
+    "eval_episodes": 100,
+    "eval_parallel_envs": 1,
+    "eval_max_steps": 30000,
+    "eval_raw_reward": true,
+    "disable_reward_scaling": true,
+    "frame_stack": 4,
+    "simulator": "cpu",
+    "episode_metrics_path": "results/checkpoints/memory/rollout_teacher/deadly_corridor_teacher_fixed_l6_obs8p75_seed0/episode_metrics.jsonl",
+    "env_frameskip": 1,
+    "eval_env_frameskip": 1
+  },
+  "git_hash": "bedb62f3a97309f6dacdb6a2953e3312d9776765",
+  "git_repo_name": "git@github.com:ZihanWang314/latency-sensitive-bench.git",
+  "output_dir": "outputs/memory/rollout_teacher/deadly_corridor/train/teacher/fixed_l6/obs8p75/seed_0",
+  "wandb_unique_id": "deadly-corridor-obs8p75-fixed_l6-s0"
+}

latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8bd1415990de5a8616a203b15d58f105dbcc57e5ac10ffff0ea70e2220d9c5c
+size 25211531190

latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/config.full.yaml

@@ -0,0 +1,286 @@
+framework:
+  name: WanOFT
+  world_model:
+    base_wm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    extract_layers:
+    - -1
+    num_frames: 5
+  qwenvl:
+    base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    attn_implementation: sdpa
+    flex_backend: triton
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    action_hidden_dim: 3072
+    action_horizon: 8
+    future_action_window_size: 7
+    past_action_window_size: 0
+    loss_type: current_multibinary_bce
+    class_weights: null
+    future_loss_weight: null
+    action_query_source: mean
+    action_query_num_heads: 24
+    state_dim: 7
+    action_env_dim: 7
+  kv_memory:
+    enabled: false
+    window: 4
+    rollout_len: 8
+    packed_train: false
+    rebased_sink: true
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/data/deadly_corridor_fix_latency_6_1000ep_context5
+    data_mix: deadly_corridor_train__bridge
+    eval_data_mix: null
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    eval_sequential_step_sampling: null
+    num_workers: 8
+    eval_num_workers: 8
+    prefetch_factor: 4
+    persistent_workers: true
+    pin_memory: true
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+      phase_distributions: null
+      new_latency_passes: 1.0
+      replay_passes: 0.25
+      target_total_passes: 2.0
+      final_equalization: true
+      step_budget_mode: auto
+      eval_at_phase_end: false
+      save_at_phase_end: false
+      computed_plan: null
+    per_device_batch_size: 1
+    load_all_data_for_training: true
+    num_obs_frames: 1
+    image_mode: single
+    prompt_mode: raw
+    stitch_grid:
+    - 2
+    - 2
+    obs_image_size:
+    - 224
+    - 224
+    video_backend: torchvision_av
+    observation_indices:
+    - -4
+    - -3
+    - -2
+    - -1
+    - 0
+    language_indices:
+    - 0
+    state_indices:
+    - 0
+    action_indices:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    - 5
+    - 6
+    - 7
+    pack_image_sequence: true
+    image_sequence_length: 5
+    context_images_column: observation.context_images
+dataset:
+  source_hf: latency-sensitive-bench/memory-data
+  config_name: deadly_corridor_fixed_latency_6_1000ep_7k2steps
+  source_subdir: null
+  converted_name: deadly_corridor_train__bridge
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: deadly_corridor_train__bridge
+  mixed_converted_name: deadly_corridor_mixed_latency_train__bridge
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  skip_verification: false
+  target_latency_unit: observation_steps
+  verify_rows: 200
+  max_episodes: 1000
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Wan-AI/Wan2.2-TI2V-5B-Diffusers
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/WM4A-Wan2d2-OFT-LIBERO-4in1
+  checkpoint_hf_repo_id: StarVLA/WM4A-Wan2d2-OFT-LIBERO-4in1
+  checkpoint_filename: checkpoints/steps_60000_pytorch_model.pt
+trainer:
+  max_train_steps: 1000
+  num_warmup_steps: 0
+  save_interval: 500
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  eval_action_classification: false
+  eval_action_classification_interval: null
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.5e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+    action_query_proj: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  freeze_vit: false
+  freeze_tied_embedding: false
+  freeze_llm_layers: []
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  profile_timing:
+    enabled: false
+    log_interval: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 16
+  distributed_backend: deepspeed
+  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: true
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: zihanwang-ai-northwestern-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: data/deadly_corridor_fix_latency_6_1000ep_context5
+  dataset_cache_dir: null
+  base_model_dir: playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+  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: wan_oft
+  env_eval:
+    image_size: 224
+    frameskip: 4
+    image_transform: raw_rgb
+    prompt_mode: raw
+    ghost_trail:
+      history_frames: 5
+      gamma: 1.3
+      min_alpha: 35
+      scroll_px_per_step: 4.0
+      ground_fraction: 0.22
+    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.
+    eval_parallel_envs: 5
+    action_chunk_execution:
+      enabled: false
+      chunk_size: null
+    deadly:
+      action_layout: multibinary_7
+      multibinary_threshold: 0.0
+    enabled: false
+    eval_backend: eval_core
+    distributed_mode: rank_sharded
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: data/deadly_corridor_fix_latency_6_1000ep_context5/deadly_corridor_train__bridge/latency_prompt_map.json
+      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: deadly_corridor
+  deadly_corridor_loss_type: current_multibinary_bce
+  initialization_mode: bridge
+  action_carrier: bridge
+model: wan_oft
+env: deadly_corridor
+init: wan_oft_libero
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: true
+  save_training_state: true
+  save_safetensors_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: wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/config.yaml

@@ -0,0 +1,83 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: true
+  save_safetensors_file: false
+  save_training_state: true
+  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: null
+    latency_curriculum:
+      enabled: false
+    obs_image_size:
+    - 224
+    - 224
+    per_device_batch_size: 1
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 7
+    action_hidden_dim: 3072
+    action_horizon: 8
+    action_model_type: MLP
+    action_query_source: mean
+    loss_type: current_multibinary_bce
+  name: WanOFT
+  qwenvl:
+    base_vlm: /workspace/runtime/models/Wan2.2-TI2V-5B-Diffusers
+  world_model:
+    base_wm: /workspace/runtime/models/Wan2.2-TI2V-5B-Diffusers
+    extract_layers:
+    - -1
+    num_frames: 5
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce
+rl_games:
+  env_eval:
+    enabled: false
+run_id: wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  distributed_backend: deepspeed
+  eval_action_classification: false
+  eval_action_classification_interval: null
+  eval_interval: 100
+  freeze_llm_layers: []
+  freeze_modules: ''
+  freeze_tied_embedding: false
+  freeze_vit: false
+  gradient_accumulation_steps: 16
+  is_resume: false
+  learning_rate:
+    action_model: 0.0001
+    action_query_proj: 0.0001
+    base: 2.5e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 1000
+  num_warmup_steps: 0
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: true
+    weight_decay: 1.0e-08
+  pretrained_checkpoint: null
+  profile_timing:
+    enabled: false
+  reload_modules: null
+  save_interval: 500
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: zihanwang-ai-northwestern-university
+wandb_project: starVLA_rl_games

latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.9434846043586731,
+        0.023302890360355377,
+        0.21685810387134552,
+        0.5315648317337036,
+        0.18562282621860504,
+        0.09055173397064209,
+        0.8273231983184814
+      ],
+      "std": [
+        0.2308816909790039,
+        0.15088047087192535,
+        0.4120523929595947,
+        0.4989914894104004,
+        0.388867050409317,
+        0.287028968334198,
+        0.37795618176460266
+      ],
+      "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,
+        1.0,
+        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": 42484,
+    "num_trajectories": 990
+  }
+}

latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/train/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/checkpoint_p0/checkpoint_000048880_25026560.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:131b3bc73aab996f9154c1fb40b632b77afd46fe1cf1cf9733b11d29870234bd
+size 7210813

latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/train/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/config.json

@@ -0,0 +1,255 @@
+{
+  "help": false,
+  "algo": "APPO",
+  "env": "latency_demon_attack",
+  "experiment": "demon_attack_teacher_fixed_l6_obs15_seed0",
+  "train_dir": "results/checkpoints/memory/rollout_teacher",
+  "restart_behavior": "resume",
+  "device": "gpu",
+  "seed": 0,
+  "num_policies": 1,
+  "async_rl": true,
+  "serial_mode": false,
+  "batched_sampling": true,
+  "num_batches_to_accumulate": 2,
+  "worker_num_splits": 1,
+  "policy_workers_per_policy": 1,
+  "max_policy_lag": 300,
+  "num_workers": 1,
+  "num_envs_per_worker": 1,
+  "batch_size": 1024,
+  "num_batches_per_epoch": 8,
+  "num_epochs": 2,
+  "rollout": 64,
+  "recurrence": 32,
+  "shuffle_minibatches": false,
+  "gamma": 0.99,
+  "reward_scale": 1.0,
+  "reward_clip": 1000.0,
+  "value_bootstrap": false,
+  "normalize_returns": true,
+  "exploration_loss_coeff": 0.003,
+  "value_loss_coeff": 0.5,
+  "kl_loss_coeff": 0.0,
+  "exploration_loss": "entropy",
+  "gae_lambda": 0.95,
+  "ppo_clip_ratio": 0.1,
+  "ppo_clip_value": 0.2,
+  "with_vtrace": false,
+  "vtrace_rho": 1.0,
+  "vtrace_c": 1.0,
+  "optimizer": "adam",
+  "adam_eps": 1e-05,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "max_grad_norm": 0.0,
+  "learning_rate": 0.000303,
+  "lr_schedule": "linear_decay",
+  "lr_schedule_kl_threshold": 0.008,
+  "lr_adaptive_min": 1e-06,
+  "lr_adaptive_max": 0.01,
+  "obs_subtract_mean": 0.0,
+  "obs_scale": 255.0,
+  "normalize_input": true,
+  "normalize_input_keys": null,
+  "decorrelate_experience_max_seconds": 0,
+  "decorrelate_envs_on_one_worker": true,
+  "actor_worker_gpus": [
+    0
+  ],
+  "set_workers_cpu_affinity": true,
+  "force_envs_single_thread": false,
+  "default_niceness": 0,
+  "log_to_file": true,
+  "experiment_summaries_interval": 1,
+  "flush_summaries_interval": 30,
+  "stats_avg": 100,
+  "summaries_use_frameskip": true,
+  "heartbeat_interval": 20,
+  "heartbeat_reporting_interval": 180,
+  "train_for_env_steps": 25000000,
+  "train_for_seconds": 10000000000,
+  "save_every_sec": 600,
+  "keep_checkpoints": 5,
+  "load_checkpoint_kind": "latest",
+  "save_milestones_sec": -1,
+  "save_best_every_sec": 5,
+  "save_best_metric": "reward",
+  "save_best_after": 100000,
+  "benchmark": false,
+  "encoder_mlp_layers": [
+    512,
+    512
+  ],
+  "encoder_conv_architecture": "convnet_atari",
+  "encoder_conv_mlp_layers": [
+    512
+  ],
+  "use_rnn": false,
+  "rnn_size": 512,
+  "rnn_type": "gru",
+  "rnn_num_layers": 1,
+  "decoder_mlp_layers": [],
+  "nonlinearity": "elu",
+  "policy_initialization": "orthogonal",
+  "policy_init_gain": 1.0,
+  "actor_critic_share_weights": true,
+  "adaptive_stddev": true,
+  "continuous_tanh_scale": 0.0,
+  "initial_stddev": 1.0,
+  "use_env_info_cache": false,
+  "env_gpu_actions": true,
+  "env_gpu_observations": true,
+  "env_frameskip": 1,
+  "env_framestack": 1,
+  "pixel_format": "CHW",
+  "use_record_episode_statistics": false,
+  "with_wandb": true,
+  "wandb_user": null,
+  "wandb_project": "latency-sensitive-bench",
+  "wandb_group": "demon-attack-obs15-fixed_l6",
+  "wandb_job_type": "sample_factory",
+  "wandb_tags": [
+    "memory_rollout_teacher",
+    "demon-attack",
+    "teacher",
+    "fixed",
+    "fixed_l6",
+    "obs15",
+    "seed0"
+  ],
+  "with_pbt": false,
+  "pbt_mix_policies_in_one_env": true,
+  "pbt_period_env_steps": 5000000,
+  "pbt_start_mutation": 20000000,
+  "pbt_replace_fraction": 0.3,
+  "pbt_mutation_rate": 0.15,
+  "pbt_replace_reward_gap": 0.1,
+  "pbt_replace_reward_gap_absolute": 1e-06,
+  "pbt_optimize_gamma": false,
+  "pbt_target_objective": "true_objective",
+  "pbt_perturb_min": 1.1,
+  "pbt_perturb_max": 1.5,
+  "env_id": "ALE/DemonAttack-v5",
+  "env_fps": 60.0,
+  "obs_fps": 15.0,
+  "screen_size": 84,
+  "noop_max": 30,
+  "reset_mode": "fire",
+  "frame_stack": 4,
+  "terminal_on_life_loss": true,
+  "clip_reward": true,
+  "repeat_action_probability": 0.0,
+  "full_action_space": false,
+  "max_episode_steps": 25000,
+  "simulator": "gpu",
+  "gpu_env_device": "auto",
+  "gpu_env_batch_size": 512,
+  "export_env_raw_rgb_frames": false,
+  "mode": "train",
+  "latency_type": "fixed",
+  "fixed_latency_ms": 100.0,
+  "mean_latency_ms": null,
+  "std_latency_ms": null,
+  "min_latency_ms": null,
+  "max_latency_ms": null,
+  "latency_seed": null,
+  "add_latency_info": false,
+  "max_pending_actions": null,
+  "hold_policy": "hold",
+  "ordering_policy": "latest_ready",
+  "eval_episodes": 100,
+  "eval_parallel_envs": 50,
+  "eval_latency_raw_frame_values": "0,1,2,3,4,5",
+  "eval_max_steps": 7200,
+  "eval_deterministic": true,
+  "eval_raw_reward": true,
+  "episode_metrics_path": "results/checkpoints/memory/rollout_teacher/demon_attack_teacher_fixed_l6_obs15_seed0/episode_metrics.jsonl",
+  "command_line": "--mode train --algo APPO --env latency_demon_attack --experiment demon_attack_teacher_fixed_l6_obs15_seed0 --train_dir results/checkpoints/memory/rollout_teacher --restart_behavior resume --device gpu --actor_worker_gpus 0 --env_gpu_observations True --env_gpu_actions True --gpu-env-batch-size 512 --seed 0 --episode_metrics_path results/checkpoints/memory/rollout_teacher/demon_attack_teacher_fixed_l6_obs15_seed0/episode_metrics.jsonl --train_for_env_steps 25000000 --num_workers 1 --num_envs_per_worker 1 --num_policies 1 --batch_size 1024 --rollout 64 --recurrence 32 --num_epochs 2 --num_batches_per_epoch 8 --worker_num_splits 1 --max_policy_lag 300 --learning_rate 0.000303 --gamma 0.99 --gae_lambda 0.95 --ppo_clip_ratio 0.1 --ppo_clip_value 0.2 --value_loss_coeff 0.5 --max_grad_norm 0.0 --save_every_sec 600 --keep_checkpoints 5 --stats_avg 100 --experiment_summaries_interval 1 --batched_sampling True --async_rl True --batched_sampling True --use_rnn False --normalize_returns True --normalize_input True --latency-type fixed --fixed-latency-ms 100.0 --add-latency-info False --eval-episodes 100 --eval-parallel-envs 50 --eval-max-steps 7200 --eval-deterministic True --eval-raw-reward --with_wandb True --wandb_project latency-sensitive-bench --wandb_group demon-attack-obs15-fixed_l6 --wandb_job_type sample_factory --wandb_tags memory_rollout_teacher demon-attack teacher fixed fixed_l6 obs15 seed0 --env_id ALE/DemonAttack-v5 --env-fps 60 --obs-fps 15.0 --screen_size 84 --noop_max 30 --reset-mode fire --frame_stack 4 --terminal_on_life_loss True --clip_reward True --repeat_action_probability 0.0 --full_action_space False --max_episode_steps 25000 --simulator gpu --gpu-env-device auto --hold-policy hold --ordering-policy latest_ready",
+  "cli_args": {
+    "algo": "APPO",
+    "env": "latency_demon_attack",
+    "experiment": "demon_attack_teacher_fixed_l6_obs15_seed0",
+    "train_dir": "results/checkpoints/memory/rollout_teacher",
+    "restart_behavior": "resume",
+    "device": "gpu",
+    "seed": 0,
+    "num_policies": 1,
+    "async_rl": true,
+    "batched_sampling": true,
+    "worker_num_splits": 1,
+    "max_policy_lag": 300,
+    "num_workers": 1,
+    "num_envs_per_worker": 1,
+    "batch_size": 1024,
+    "num_batches_per_epoch": 8,
+    "num_epochs": 2,
+    "rollout": 64,
+    "recurrence": 32,
+    "gamma": 0.99,
+    "normalize_returns": true,
+    "value_loss_coeff": 0.5,
+    "gae_lambda": 0.95,
+    "ppo_clip_ratio": 0.1,
+    "ppo_clip_value": 0.2,
+    "max_grad_norm": 0.0,
+    "learning_rate": 0.000303,
+    "normalize_input": true,
+    "actor_worker_gpus": [
+      0
+    ],
+    "experiment_summaries_interval": 1,
+    "stats_avg": 100,
+    "train_for_env_steps": 25000000,
+    "save_every_sec": 600,
+    "keep_checkpoints": 5,
+    "use_rnn": false,
+    "env_gpu_actions": true,
+    "env_gpu_observations": true,
+    "with_wandb": true,
+    "wandb_project": "latency-sensitive-bench",
+    "wandb_group": "demon-attack-obs15-fixed_l6",
+    "wandb_job_type": "sample_factory",
+    "wandb_tags": [
+      "memory_rollout_teacher",
+      "demon-attack",
+      "teacher",
+      "fixed",
+      "fixed_l6",
+      "obs15",
+      "seed0"
+    ],
+    "env_id": "ALE/DemonAttack-v5",
+    "env_fps": 60.0,
+    "obs_fps": 15.0,
+    "screen_size": 84,
+    "noop_max": 30,
+    "reset_mode": "fire",
+    "frame_stack": 4,
+    "terminal_on_life_loss": true,
+    "clip_reward": true,
+    "repeat_action_probability": 0.0,
+    "full_action_space": false,
+    "max_episode_steps": 25000,
+    "simulator": "gpu",
+    "gpu_env_device": "auto",
+    "gpu_env_batch_size": 512,
+    "mode": "train",
+    "latency_type": "fixed",
+    "fixed_latency_ms": 100.0,
+    "add_latency_info": false,
+    "hold_policy": "hold",
+    "ordering_policy": "latest_ready",
+    "eval_episodes": 100,
+    "eval_parallel_envs": 50,
+    "eval_max_steps": 7200,
+    "eval_deterministic": true,
+    "eval_raw_reward": true,
+    "episode_metrics_path": "results/checkpoints/memory/rollout_teacher/demon_attack_teacher_fixed_l6_obs15_seed0/episode_metrics.jsonl"
+  },
+  "git_hash": "bedb62f3a97309f6dacdb6a2953e3312d9776765",
+  "git_repo_name": "git@github.com:ZihanWang314/latency-sensitive-bench.git",
+  "eval_env_frameskip": 1,
+  "output_dir": "outputs/memory/rollout_teacher/demon_attack/train/teacher/fixed_l6/obs15/seed_0",
+  "wandb_unique_id": "demon-attack-obs15-fixed_l6-s0"
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/checkpoints/steps_7000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ab6d4d4eb795a9a09da9c63d489771039440689877642e3f9ec48dd9112119a
+size 9138230516

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/config.full.yaml

@@ -0,0 +1,209 @@
+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_200ep
+    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
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+    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: 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
+  skip_verification: 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: 7000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  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: 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_200ep/demon_attack_train__bridge/latency_prompt_map.json
+      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: 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: 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: demon_attack_fix_latency_2_200ep
+output_dir: /workspace/starVLA/results/Checkpoints/demon_attack_fix_latency_2_200ep
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/config.yaml

@@ -0,0 +1,75 @@
+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: demon_attack_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: demon_attack_train__bridge__val
+    latency_curriculum:
+      enabled: false
+    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
+  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/demon_attack_fix_latency_2_200ep
+rl_games:
+  env_eval:
+    enabled: true
+    mid_train:
+      enabled: false
+      interval_steps: 100
+  task: demon_attack
+run_id: demon_attack_fix_latency_2_200ep
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  cc_f1_tolerance: 1
+  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: 7000
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: false
+    weight_decay: 1.0e-08
+  per_latency_eval_num_batches: null
+  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/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.06905864179134369,
+        0.37893569469451904,
+        0.13702447712421417,
+        0.1011248454451561,
+        0.12969808280467987,
+        0.18415828049182892,
+        0.0
+      ],
+      "std": [
+        0.2532638907432556,
+        0.4849342703819275,
+        0.3443855047225952,
+        0.3010394871234894,
+        0.3360588848590851,
+        0.3878043293952942,
+        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": 349558,
+    "num_trajectories": 180
+  }
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.07371096312999725,
+        0.3668539524078369,
+        0.13910865783691406,
+        0.1111111119389534,
+        0.13123494386672974,
+        0.17798034846782684,
+        0.0
+      ],
+      "std": [
+        0.2612755298614502,
+        0.48196175694465637,
+        0.3461320102214813,
+        0.3142721354961395,
+        0.3376607596874237,
+        0.38254332542419434,
+        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": 52326,
+    "num_trajectories": 20
+  }
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/hydra/.hydra/config.yaml

@@ -0,0 +1,207 @@
+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: 6
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_2_200ep
+    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
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+    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: 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
+  skip_verification: 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: 7000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  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: /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 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: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_2_200ep/demon_attack_train__bridge/latency_prompt_map.json
+      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: 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: 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: demon_attack_fix_latency_2_200ep
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/hydra/.hydra/hydra.yaml

@@ -0,0 +1,259 @@
+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=demon_attack_fix_latency_2_200ep
+    - 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=false
+    - 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=7000
+    - 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=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.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.latency_curriculum.enabled=false
+    - datasets.vla_data.latency_curriculum.strategy=exclusive
+    - 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/demon_attack_fix_latency_2_200ep
+    - datasets.vla_data.data_mix=demon_attack_train__bridge
+    - datasets.vla_data.eval_data_mix=demon_attack_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=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=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_2_200ep/demon_attack_train__bridge/latency_prompt_map.json
+    - 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=demon_attack_train__bridge,datasets.vla_data.data_root_dir=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_2_200ep,datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val,datasets.vla_data.include_state=true,datasets.vla_data.latency_curriculum.enabled=false,datasets.vla_data.latency_curriculum.strategy=exclusive,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=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=/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.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/demon_attack_fix_latency_2_200ep/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=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_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=demon_attack_fix_latency_2_200ep,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=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=7000,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=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: /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/demon_attack_fix_latency_2_200ep/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/demon_attack_fix_latency_2_200ep/hydra/.hydra/overrides.yaml

@@ -0,0 +1,97 @@
+- model=openvla
+- env=demon_attack
+- init=bridge
+- mode=single
+- run_id=demon_attack_fix_latency_2_200ep
+- 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=false
+- 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=7000
+- 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=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.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.latency_curriculum.enabled=false
+- datasets.vla_data.latency_curriculum.strategy=exclusive
+- 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/demon_attack_fix_latency_2_200ep
+- datasets.vla_data.data_mix=demon_attack_train__bridge
+- datasets.vla_data.eval_data_mix=demon_attack_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=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=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_2_200ep/demon_attack_train__bridge/latency_prompt_map.json
+- 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/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/summary.jsonl

@@ -0,0 +1,70 @@
+{"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}
+{"steps": 5100}
+{"steps": 5200}
+{"steps": 5300}
+{"steps": 5400}
+{"steps": 5500}
+{"steps": 5600}
+{"steps": 5700}
+{"steps": 5800}
+{"steps": 5900}
+{"steps": 6000}
+{"steps": 6100}
+{"steps": 6200}
+{"steps": 6300}
+{"steps": 6400}
+{"steps": 6500}
+{"steps": 6600}
+{"steps": 6700}
+{"steps": 6800}
+{"steps": 6900}
+{"steps": 7000}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/checkpoints/steps_7000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62e9b2a2b38f2b52cedc248c6e135b2593b212d1965fc3efae42e8130b57b373
+size 9138230516

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/config.full.yaml

@@ -0,0 +1,209 @@
+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_4_200ep
+    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
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+    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: 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
+  skip_verification: 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: 7000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  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: 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_4_200ep/demon_attack_train__bridge/latency_prompt_map.json
+      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: 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: 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: demon_attack_fix_latency_4_200ep
+output_dir: /workspace/starVLA/results/Checkpoints/demon_attack_fix_latency_4_200ep
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/config.yaml

@@ -0,0 +1,75 @@
+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: demon_attack_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: demon_attack_train__bridge__val
+    latency_curriculum:
+      enabled: false
+    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
+  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/demon_attack_fix_latency_2_200ep
+rl_games:
+  env_eval:
+    enabled: true
+    mid_train:
+      enabled: false
+      interval_steps: 100
+  task: demon_attack
+run_id: demon_attack_fix_latency_4_200ep
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  cc_f1_tolerance: 1
+  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: 7000
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: false
+    weight_decay: 1.0e-08
+  per_latency_eval_num_batches: null
+  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/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.02783012203872204,
+        0.6656418442726135,
+        0.008087566122412682,
+        0.061184193938970566,
+        0.1591275930404663,
+        0.07812871038913727,
+        0.0
+      ],
+      "std": [
+        0.16479234397411346,
+        0.47177085280418396,
+        0.0895964726805687,
+        0.23992209136486053,
+        0.36527156829833984,
+        0.26814261078834534,
+        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,
+        0.0,
+        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": 332733,
+    "num_trajectories": 180
+  }
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.028564849868416786,
+        0.675146758556366,
+        0.00799223780632019,
+        0.06372412294149399,
+        0.15450015664100647,
+        0.07007186114788055,
+        0.0
+      ],
+      "std": [
+        0.1666077822446823,
+        0.46833404898643494,
+        0.08902717381715775,
+        0.24433407187461853,
+        0.36147764325141907,
+        0.2552240490913391,
+        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,
+        0.0,
+        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": 60809,
+    "num_trajectories": 20
+  }
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/hydra/.hydra/config.yaml

@@ -0,0 +1,207 @@
+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: 6
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_4_200ep
+    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
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+    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: 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
+  skip_verification: 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: 7000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  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: /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 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: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_4_200ep/demon_attack_train__bridge/latency_prompt_map.json
+      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: 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: 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: demon_attack_fix_latency_4_200ep
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/hydra/.hydra/hydra.yaml

@@ -0,0 +1,259 @@
+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=demon_attack_fix_latency_4_200ep
+    - 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=false
+    - 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=7000
+    - 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=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.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.latency_curriculum.enabled=false
+    - datasets.vla_data.latency_curriculum.strategy=exclusive
+    - 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/demon_attack_fix_latency_4_200ep
+    - datasets.vla_data.data_mix=demon_attack_train__bridge
+    - datasets.vla_data.eval_data_mix=demon_attack_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=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=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_4_200ep/demon_attack_train__bridge/latency_prompt_map.json
+    - 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=demon_attack_train__bridge,datasets.vla_data.data_root_dir=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_4_200ep,datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val,datasets.vla_data.include_state=true,datasets.vla_data.latency_curriculum.enabled=false,datasets.vla_data.latency_curriculum.strategy=exclusive,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=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=/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.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/demon_attack_fix_latency_4_200ep/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=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_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=demon_attack_fix_latency_4_200ep,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=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=7000,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=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: /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/demon_attack_fix_latency_4_200ep/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/demon_attack_fix_latency_4_200ep/hydra/.hydra/overrides.yaml

@@ -0,0 +1,97 @@
+- model=openvla
+- env=demon_attack
+- init=bridge
+- mode=single
+- run_id=demon_attack_fix_latency_4_200ep
+- 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=false
+- 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=7000
+- 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=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.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.latency_curriculum.enabled=false
+- datasets.vla_data.latency_curriculum.strategy=exclusive
+- 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/demon_attack_fix_latency_4_200ep
+- datasets.vla_data.data_mix=demon_attack_train__bridge
+- datasets.vla_data.eval_data_mix=demon_attack_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=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=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_4_200ep/demon_attack_train__bridge/latency_prompt_map.json
+- 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/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/summary.jsonl

@@ -0,0 +1,70 @@
+{"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}
+{"steps": 5100}
+{"steps": 5200}
+{"steps": 5300}
+{"steps": 5400}
+{"steps": 5500}
+{"steps": 5600}
+{"steps": 5700}
+{"steps": 5800}
+{"steps": 5900}
+{"steps": 6000}
+{"steps": 6100}
+{"steps": 6200}
+{"steps": 6300}
+{"steps": 6400}
+{"steps": 6500}
+{"steps": 6600}
+{"steps": 6700}
+{"steps": 6800}
+{"steps": 6900}
+{"steps": 7000}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/checkpoints/steps_7000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:375ff9f24479e219e8197bbc2eee3e71e61e1853e0c80eefe438ca947d5c5b1b
+size 9138230516

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/config.full.yaml

@@ -0,0 +1,209 @@
+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_6_200ep
+    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
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+    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: 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
+  skip_verification: 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: 7000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  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: 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_6_200ep/demon_attack_train__bridge/latency_prompt_map.json
+      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: 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: 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: demon_attack_fix_latency_6_200ep
+output_dir: /workspace/starVLA/results/Checkpoints/demon_attack_fix_latency_6_200ep
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/config.yaml

@@ -0,0 +1,75 @@
+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: demon_attack_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: demon_attack_train__bridge__val
+    latency_curriculum:
+      enabled: false
+    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
+  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/demon_attack_fix_latency_6_200ep
+rl_games:
+  env_eval:
+    enabled: true
+    mid_train:
+      enabled: false
+      interval_steps: 100
+  task: demon_attack
+run_id: demon_attack_fix_latency_6_200ep
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  cc_f1_tolerance: 1
+  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: 7000
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: false
+    weight_decay: 1.0e-08
+  per_latency_eval_num_batches: null
+  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/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.06990445405244827,
+        0.6139481067657471,
+        0.008142709732055664,
+        0.061076752841472626,
+        0.1516113430261612,
+        0.09531665593385696,
+        0.0
+      ],
+      "std": [
+        0.2553316354751587,
+        0.48614779114723206,
+        0.08989176899194717,
+        0.23972871899604797,
+        0.3591064512729645,
+        0.29394957423210144,
+        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,
+        0.0,
+        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": 310953,
+    "num_trajectories": 180
+  }
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.07470002770423889,
+        0.6113184094429016,
+        0.008066286332905293,
+        0.058384548872709274,
+        0.15318627655506134,
+        0.09434445202350616,
+        0.0
+      ],
+      "std": [
+        0.2629486620426178,
+        0.48751503229141235,
+        0.08943457156419754,
+        0.23448707163333893,
+        0.36014819145202637,
+        0.2922261357307434,
+        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,
+        0.0,
+        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": 54672,
+    "num_trajectories": 20
+  }
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/hydra/.hydra/config.yaml

@@ -0,0 +1,207 @@
+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: 6
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_6_200ep
+    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
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+    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: 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
+  skip_verification: 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: 7000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  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: /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 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: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_6_200ep/demon_attack_train__bridge/latency_prompt_map.json
+      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: 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: 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: demon_attack_fix_latency_6_200ep
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/hydra/.hydra/hydra.yaml

@@ -0,0 +1,259 @@
+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=demon_attack_fix_latency_6_200ep
+    - 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=false
+    - 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=7000
+    - 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=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.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.latency_curriculum.enabled=false
+    - datasets.vla_data.latency_curriculum.strategy=exclusive
+    - 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/demon_attack_fix_latency_6_200ep
+    - datasets.vla_data.data_mix=demon_attack_train__bridge
+    - datasets.vla_data.eval_data_mix=demon_attack_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=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=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_6_200ep/demon_attack_train__bridge/latency_prompt_map.json
+    - 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=demon_attack_train__bridge,datasets.vla_data.data_root_dir=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_6_200ep,datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val,datasets.vla_data.include_state=true,datasets.vla_data.latency_curriculum.enabled=false,datasets.vla_data.latency_curriculum.strategy=exclusive,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=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=/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.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/demon_attack_fix_latency_6_200ep/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=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_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=demon_attack_fix_latency_6_200ep,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=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=7000,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=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: /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/demon_attack_fix_latency_6_200ep/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/demon_attack_fix_latency_6_200ep/hydra/.hydra/overrides.yaml

@@ -0,0 +1,97 @@
+- model=openvla
+- env=demon_attack
+- init=bridge
+- mode=single
+- run_id=demon_attack_fix_latency_6_200ep
+- 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=false
+- 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=7000
+- 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=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.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.latency_curriculum.enabled=false
+- datasets.vla_data.latency_curriculum.strategy=exclusive
+- 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/demon_attack_fix_latency_6_200ep
+- datasets.vla_data.data_mix=demon_attack_train__bridge
+- datasets.vla_data.eval_data_mix=demon_attack_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=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=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_6_200ep/demon_attack_train__bridge/latency_prompt_map.json
+- 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/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/summary.jsonl

@@ -0,0 +1,70 @@
+{"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}
+{"steps": 5100}
+{"steps": 5200}
+{"steps": 5300}
+{"steps": 5400}
+{"steps": 5500}
+{"steps": 5600}
+{"steps": 5700}
+{"steps": 5800}
+{"steps": 5900}
+{"steps": 6000}
+{"steps": 6100}
+{"steps": 6200}
+{"steps": 6300}
+{"steps": 6400}
+{"steps": 6500}
+{"steps": 6600}
+{"steps": 6700}
+{"steps": 6800}
+{"steps": 6900}
+{"steps": 7000}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/checkpoints/steps_7000_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d52d791412c41f17f607087c988d99267ef09ac097022b3fc3d97ebbb5b0b00
+size 9138230516

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/config.full.yaml

@@ -0,0 +1,209 @@
+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_8_200ep
+    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
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+    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: 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
+  skip_verification: 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: 7000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  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: 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_8_200ep/demon_attack_train__bridge/latency_prompt_map.json
+      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: 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: 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: demon_attack_fix_latency_8_200ep
+output_dir: /workspace/starVLA/results/Checkpoints/demon_attack_fix_latency_8_200ep
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/config.yaml

@@ -0,0 +1,75 @@
+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: demon_attack_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: demon_attack_train__bridge__val
+    latency_curriculum:
+      enabled: false
+    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
+  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/demon_attack_fix_latency_8_200ep
+rl_games:
+  env_eval:
+    enabled: true
+    mid_train:
+      enabled: false
+      interval_steps: 100
+  task: demon_attack
+run_id: demon_attack_fix_latency_8_200ep
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  cc_f1_tolerance: 1
+  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: 7000
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: false
+    weight_decay: 1.0e-08
+  per_latency_eval_num_batches: null
+  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/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.08639267086982727,
+        0.627116858959198,
+        0.006486095488071442,
+        0.005628371611237526,
+        0.148070827126503,
+        0.12630519270896912,
+        0.0
+      ],
+      "std": [
+        0.28125613927841187,
+        0.48417720198631287,
+        0.08015890419483185,
+        0.07480393350124359,
+        0.35502511262893677,
+        0.33193621039390564,
+        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,
+        0.0,
+        0.0,
+        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": 282142,
+    "num_trajectories": 180
+  }
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.09224364906549454,
+        0.6070592999458313,
+        0.007906598970293999,
+        0.0057144262827932835,
+        0.1609153002500534,
+        0.12616074085235596,
+        0.0
+      ],
+      "std": [
+        0.2894008159637451,
+        0.48834389448165894,
+        0.08855940401554108,
+        0.07536990940570831,
+        0.36744311451911926,
+        0.33201834559440613,
+        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,
+        0.0,
+        0.0,
+        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": 40599,
+    "num_trajectories": 20
+  }
+}

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/hydra/.hydra/config.yaml

@@ -0,0 +1,207 @@
+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: 6
+  name: QwenOFT
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_8_200ep
+    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
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+    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: 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
+  skip_verification: 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: 7000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  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: /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 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: /inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_8_200ep/demon_attack_train__bridge/latency_prompt_map.json
+      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: 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: 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: demon_attack_fix_latency_8_200ep
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: '0.21'

latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/hydra/.hydra/hydra.yaml

@@ -0,0 +1,259 @@
+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=demon_attack_fix_latency_8_200ep
+    - 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=false
+    - 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=7000
+    - 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=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.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.latency_curriculum.enabled=false
+    - datasets.vla_data.latency_curriculum.strategy=exclusive
+    - 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/demon_attack_fix_latency_8_200ep
+    - datasets.vla_data.data_mix=demon_attack_train__bridge
+    - datasets.vla_data.eval_data_mix=demon_attack_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=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=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_8_200ep/demon_attack_train__bridge/latency_prompt_map.json
+    - 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=demon_attack_train__bridge,datasets.vla_data.data_root_dir=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_8_200ep,datasets.vla_data.eval_data_mix=demon_attack_train__bridge__val,datasets.vla_data.include_state=true,datasets.vla_data.latency_curriculum.enabled=false,datasets.vla_data.latency_curriculum.strategy=exclusive,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=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=/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.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/demon_attack_fix_latency_8_200ep/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=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_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=demon_attack_fix_latency_8_200ep,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=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=7000,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=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: /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/demon_attack_fix_latency_8_200ep/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/demon_attack_fix_latency_8_200ep/hydra/.hydra/overrides.yaml

@@ -0,0 +1,97 @@
+- model=openvla
+- env=demon_attack
+- init=bridge
+- mode=single
+- run_id=demon_attack_fix_latency_8_200ep
+- 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=false
+- 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=7000
+- 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=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.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.latency_curriculum.enabled=false
+- datasets.vla_data.latency_curriculum.strategy=exclusive
+- 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/demon_attack_fix_latency_8_200ep
+- datasets.vla_data.data_mix=demon_attack_train__bridge
+- datasets.vla_data.eval_data_mix=demon_attack_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=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=/inspire/hdd/global_user/liumingyu-253208120284/lzj/starvla/data/demon_attack_fix_latency_8_200ep/demon_attack_train__bridge/latency_prompt_map.json
+- 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/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/summary.jsonl

@@ -0,0 +1,70 @@
+{"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}
+{"steps": 5100}
+{"steps": 5200}
+{"steps": 5300}
+{"steps": 5400}
+{"steps": 5500}
+{"steps": 5600}
+{"steps": 5700}
+{"steps": 5800}
+{"steps": 5900}
+{"steps": 6000}
+{"steps": 6100}
+{"steps": 6200}
+{"steps": 6300}
+{"steps": 6400}
+{"steps": 6500}
+{"steps": 6600}
+{"steps": 6700}
+{"steps": 6800}
+{"steps": 6900}
+{"steps": 7000}

latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:791c66ee0313063e3eeec01a3eeb788e642f8e22952e232caa6562998d0754dd
+size 25211531190

latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/config.full.yaml

@@ -0,0 +1,280 @@
+framework:
+  name: WanOFT
+  world_model:
+    base_wm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    extract_layers:
+    - -1
+    num_frames: 5
+  qwenvl:
+    base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    attn_implementation: sdpa
+    flex_backend: triton
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    action_hidden_dim: 3072
+    action_horizon: 8
+    future_action_window_size: 7
+    past_action_window_size: 0
+    loss_type: current_discrete_ce
+    class_weights: null
+    future_loss_weight: null
+    action_query_source: mean
+    action_query_num_heads: 24
+    state_dim: 7
+    action_env_dim: 6
+  kv_memory:
+    enabled: false
+    window: 4
+    rollout_len: 8
+    packed_train: false
+    rebased_sink: true
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/data/demon_attack_fix_latency_6_200ep_context5
+    data_mix: demon_attack_train__bridge
+    eval_data_mix: null
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    eval_sequential_step_sampling: null
+    num_workers: 8
+    eval_num_workers: 8
+    prefetch_factor: 4
+    persistent_workers: true
+    pin_memory: true
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+      phase_distributions: null
+      new_latency_passes: 1.0
+      replay_passes: 0.25
+      target_total_passes: 2.0
+      final_equalization: true
+      step_budget_mode: auto
+      eval_at_phase_end: false
+      save_at_phase_end: false
+      computed_plan: null
+    per_device_batch_size: 1
+    load_all_data_for_training: true
+    num_obs_frames: 1
+    image_mode: single
+    prompt_mode: raw
+    stitch_grid:
+    - 2
+    - 2
+    obs_image_size:
+    - 224
+    - 224
+    video_backend: torchvision_av
+    observation_indices:
+    - -4
+    - -3
+    - -2
+    - -1
+    - 0
+    language_indices:
+    - 0
+    state_indices:
+    - 0
+    action_indices:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    - 5
+    - 6
+    - 7
+    pack_image_sequence: true
+    image_sequence_length: 5
+    context_images_column: observation.context_images
+dataset:
+  source_hf: latency-sensitive-bench/memory-data
+  config_name: demon_attack_fixed_latency_6_200ep_7k2steps
+  source_subdir: null
+  converted_name: demon_attack_train__bridge
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: demon_attack_train__bridge
+  mixed_converted_name: demon_attack_mixed_latency_train__bridge
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  skip_verification: false
+  target_latency_unit: observation_steps
+  verify_rows: 200
+  max_episodes: 200
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Wan-AI/Wan2.2-TI2V-5B-Diffusers
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/WM4A-Wan2d2-OFT-LIBERO-4in1
+  checkpoint_hf_repo_id: StarVLA/WM4A-Wan2d2-OFT-LIBERO-4in1
+  checkpoint_filename: checkpoints/steps_60000_pytorch_model.pt
+trainer:
+  max_train_steps: 5000
+  num_warmup_steps: 0
+  save_interval: 500
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  eval_action_classification: false
+  eval_action_classification_interval: null
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.5e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+    action_query_proj: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  freeze_vit: false
+  freeze_tied_embedding: false
+  freeze_llm_layers: []
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  profile_timing:
+    enabled: false
+    log_interval: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 16
+  distributed_backend: deepspeed
+  is_resume: true
+  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_4000_state
+  resume_step: 4000
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: true
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: zihanwang-ai-northwestern-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: data/demon_attack_fix_latency_6_200ep_context5
+  dataset_cache_dir: null
+  base_model_dir: playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+  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: wan_oft
+  env_eval:
+    image_size: 224
+    frameskip: 4
+    image_transform: raw_rgb
+    prompt_mode: raw
+    ghost_trail:
+      history_frames: 5
+      gamma: 1.3
+      min_alpha: 35
+      scroll_px_per_step: 4.0
+      ground_fraction: 0.22
+    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.'
+    eval_parallel_envs: 5
+    action_chunk_execution:
+      enabled: false
+      chunk_size: null
+    demon_attack:
+      noop_max: 30
+    enabled: true
+    eval_backend: eval_core
+    distributed_mode: rank_sharded
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: data/demon_attack_fix_latency_6_200ep_context5/demon_attack_train__bridge/latency_prompt_map.json
+      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:
+      - 6
+      num_episodes: 20
+      max_steps_per_episode: 3600
+  task: demon_attack
+  initialization_mode: bridge
+  action_carrier: bridge
+model: wan_oft
+env: demon_attack
+init: wan_oft_libero
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: true
+  save_training_state: true
+  save_safetensors_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: wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce
+config_yaml: null
+is_debug: false
+version_id: '0.21'