Consolidate figure2_canonical_metadata assets with exact source object identities

zero-latency/demon-attack/vla/starvla-qwengr00t-h1/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6/config.full.yaml

@@ -0,0 +1,281 @@
+framework:
+  name: QwenGR00T
+  qwenvl:
+    base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    vl_hidden_dim: 2048
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: DiT-B
+    action_hidden_dim: 1024
+    hidden_size: 1024
+    add_pos_embed: true
+    max_seq_len: 1024
+    action_dim: 7
+    state_dim: 7
+    action_horizon: 1
+    repeated_diffusion_steps: 8
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_timestep_buckets: 1000
+    num_inference_timesteps: 4
+    num_target_vision_tokens: 32
+    diffusion_model_cfg:
+      cross_attention_dim: 2560
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: true
+      norm_type: ada_norm
+      num_layers: 16
+      output_dim: 1024
+      positional_embeddings: null
+      attention_head_dim: 64
+    loss_type: l1
+    future_action_window_size: 0
+    past_action_window_size: 0
+    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/latency-sensitive-bench/playground/Datasets/rl_games
+    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
+    eval_sequential_step_sampling: null
+    num_workers: 8
+    eval_num_workers: 0
+    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: 8
+    load_all_data_for_training: true
+    num_obs_frames: 1
+    image_mode: single
+    prompt_mode: raw
+    stitch_grid:
+    - 2
+    - 2
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: latency-sensitive-bench/demon_attack_200ep
+  config_name: null
+  source_subdir: demon_attack_fix_latency_0_200ep
+  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: true
+  skip_verification: false
+  target_latency_unit: observation_steps
+  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-GR00T-Bridge-RT-1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-GR00T-Bridge-RT-1
+  checkpoint_filename: checkpoints/steps_20000_pytorch_model.pt
+trainer:
+  max_train_steps: 5000
+  num_warmup_steps: 0
+  save_interval: 500
+  eval_interval: 300
+  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: 1.0e-06
+    qwen_vl_interface: 1.0e-06
+    action_model: 1.0e-05
+  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: none
+  is_resume: true
+  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/results/Checkpoints/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6/checkpoints/steps_4500_state
+  resume_step: 4500
+  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: /lustre/fsw/portfolios/nvr/users/zihwang/latency-sensitive-bench
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/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: 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: gr00t
+  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: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/playground/Datasets/rl_games/demon_attack_train__bridge/latency_prompt_map.json
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: true
+      interval_steps: 300
+      latencies:
+      - 0
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: true
+      latencies:
+      - 0
+      - 2
+      - 4
+      - 6
+      - 8
+      num_episodes: 20
+      max_steps_per_episode: 3600
+  task: demon_attack
+  initialization_mode: bridge
+  action_carrier: bridge
+model: gr00t
+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: talha15032/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: false
+  save_training_state: true
+  save_safetensors_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: talha15032/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/results/Checkpoints/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6
+config_yaml: null
+is_debug: false
+version_id: '0.21'

zero-latency/demon-attack/vla/starvla-qwengr00t-h1/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6/config.yaml

@@ -0,0 +1,136 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: false
+  save_safetensors_file: false
+  save_training_state: true
+  sync:
+    enabled: false
+    keep_last_n: 0
+    repo_id: talha15032/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6
+datasets:
+  vla_data:
+    data_mix: demon_attack_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: demon_attack_train__bridge__val
+    include_state: true
+    latency_curriculum:
+      enabled: false
+    obs_image_size: null
+    per_device_batch_size: 8
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 6
+    action_hidden_dim: 1024
+    action_horizon: 1
+    action_model_type: DiT-B
+    add_pos_embed: true
+    diffusion_model_cfg:
+      attention_head_dim: 64
+      cross_attention_dim: 2560
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: true
+      norm_type: ada_norm
+      num_layers: 16
+      output_dim: 1024
+      positional_embeddings: null
+    hidden_size: 1024
+    max_seq_len: 1024
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_inference_timesteps: 4
+    num_target_vision_tokens: 32
+    num_timestep_buckets: 1000
+    repeated_diffusion_steps: 8
+    state_dim: 7
+  name: QwenGR00T
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/results/Checkpoints/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6
+rl_games:
+  env_eval:
+    action_chunk_execution:
+      chunk_size: null
+      enabled: false
+    demon_attack:
+      noop_max: 30
+    distributed_mode: none
+    enabled: true
+    eval_backend: eval_core
+    fixed_episode_seeds: true
+    frameskip: 4
+    ghost_trail:
+      history_frames: 5
+      gamma: 1.3
+      min_alpha: 35
+      scroll_px_per_step: 4.0
+      ground_fraction: 0.22
+    image_size: 224
+    image_transform: raw_rgb
+    latency:
+      prompt_map_path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/playground/Datasets/rl_games/demon_attack_train__bridge/latency_prompt_map.json
+    latency_seed_stride: 0
+    mid_train:
+      enabled: true
+      interval_steps: 300
+      latencies:
+      - 0
+      max_steps_per_episode: 3600
+      num_episodes: 5
+    prompt_mode: raw
+    seed: 42
+    task_description: 'You are playing Demon Attack from a single game image. Choose
+      exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'
+    task_seed_stride: 0
+    vectorized:
+      enabled: false
+  model_alias: gr00t
+  task: demon_attack
+run_id: gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/results/Checkpoints
+seed: 42
+trainer:
+  distributed_backend: none
+  eval_action_classification: false
+  eval_action_classification_interval: null
+  eval_interval: 300
+  eval_num_batches: 100
+  freeze_llm_layers: []
+  freeze_modules: ''
+  freeze_tied_embedding: false
+  freeze_vit: false
+  gradient_accumulation_steps: 16
+  gradient_clipping: 1.0
+  is_resume: true
+  learning_rate:
+    action_model: 1.0e-05
+    base: 1.0e-06
+    qwen_vl_interface: 1.0e-06
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 5000
+  num_warmup_steps: 0
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: true
+    weight_decay: 1.0e-08
+  per_latency_eval_num_batches: null
+  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/results/Checkpoints/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6/checkpoints/steps_4500_state
+  profile_timing:
+    enabled: false
+  resume_step: 4500
+  save_interval: 500
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: zihanwang-ai-northwestern-university
+wandb_project: starVLA_rl_games

zero-latency/demon-attack/vla/starvla-qwengr00t-h1/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.009211494587361813,
+        0.00885747093707323,
+        0.17263218760490417,
+        0.1680275797843933,
+        0.3163609206676483,
+        0.32491034269332886,
+        0.0
+      ],
+      "std": [
+        0.09535734355449677,
+        0.09356050193309784,
+        0.3784366250038147,
+        0.3730704188346863,
+        0.46596139669418335,
+        0.4689920246601105,
+        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": [
+        0.0,
+        0.0,
+        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": 435000,
+    "num_trajectories": 180
+  }
+}

zero-latency/demon-attack/vla/starvla-qwengr00t-h1/gr00t_bridge_demon_attack_single_latency_clean_data_exp3_action_1e-5_backbone_1e-6/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.009733290411531925,
+        0.009130473248660564,
+        0.17506690323352814,
+        0.17256741225719452,
+        0.3143470585346222,
+        0.31915488839149475,
+        0.0
+      ],
+      "std": [
+        0.09819413721561432,
+        0.09510758519172668,
+        0.3800537884235382,
+        0.37778180837631226,
+        0.46411585807800293,
+        0.466117799282074,
+        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": [
+        0.0,
+        0.0,
+        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": 68014,
+    "num_trajectories": 20
+  }
+}

zero-latency/flappy/vla/starvla-qwenpi_v3-h1/pi05_flappy_fix_latency_0_200ep/config.full.yaml

@@ -0,0 +1,233 @@
+framework:
+  name: QwenPI_v3
+  qwenvl:
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    vl_hidden_dim: 2560
+    num_vl_layers: 36
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: LayerwiseFM
+    action_dim: 7
+    state_dim: 7
+    action_horizon: 1
+    repeated_diffusion_steps: 2
+    num_inference_timesteps: 4
+    add_pos_embed: true
+    max_seq_len: 1024
+    num_target_vision_tokens: 32
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_timestep_buckets: 1000
+    diffusion_model_cfg:
+      action_dit_hidden_dim: 1024
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: false
+      norm_type: ada_norm
+      positional_embeddings: null
+      attention_head_dim: 64
+      output_dim: 1024
+      num_layers: 36
+      input_embedding_dim: 1024
+      cross_attention_dim: 1024
+      num_attention_heads: 16
+    loss_type: l1
+    future_action_window_size: 0
+    past_action_window_size: 0
+    action_env_dim: 2
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/flappy_fix_latency_0_200ep
+    data_mix: flappy_train__bridge
+    eval_data_mix: flappy_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    eval_sequential_step_sampling: null
+    num_workers: 4
+    eval_num_workers: null
+    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: 16
+    load_all_data_for_training: true
+    obs_image_size: null
+    video_backend: torchvision_av
+dataset:
+  source_hf: ''
+  config_name: null
+  source_subdir: null
+  converted_name: flappy_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train
+  mixed_converted_name: flappy_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  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-PI_v3-Bridge-RT_1
+  checkpoint_hf_repo_id: StarVLA/Qwen3VL-PI_v3-Bridge-RT_1
+  checkpoint_filename: checkpoints/steps_50000_pytorch_model.pt
+trainer:
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  save_interval: 100
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  eval_action_classification: true
+  eval_action_classification_interval: null
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 5.0e-06
+    qwen_vl_interface: 1.0e-05
+    action_model: 5.0e-06
+  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: 8
+  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: pi-0.5
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: ''
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: /workspace/starVLA/data/flappy_fix_latency_0_200ep/flappy_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: flappy
+  initialization_mode: bridge
+  action_carrier: bridge
+model: pi05
+env: flappy
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    pi0: playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+    pi05: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    gr00t: playground/Pretrained_models/Qwen3-VL-4B-Instruct
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_pt_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: pi05_flappy_fix_latency_0_200ep
+output_dir: /workspace/starVLA/results/Checkpoints/pi05_flappy_fix_latency_0_200ep
+config_yaml: null
+is_debug: false
+version_id: '0.21'

zero-latency/flappy/vla/starvla-qwenpi_v3-h1/pi05_flappy_fix_latency_0_200ep/config.yaml

@@ -0,0 +1,96 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: false
+  save_pt_file: false
+  sync:
+    enabled: false
+    keep_last_n: 0
+    repo_id: null
+datasets:
+  vla_data:
+    data_mix: flappy_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: flappy_train__bridge__val
+    latency_curriculum:
+      enabled: false
+    obs_image_size: null
+    per_device_batch_size: 16
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 2
+    action_horizon: 1
+    add_pos_embed: true
+    diffusion_model_cfg:
+      action_dit_hidden_dim: 1024
+      attention_head_dim: 64
+      cross_attention_dim: 1024
+      dropout: 0.2
+      final_dropout: true
+      input_embedding_dim: 1024
+      interleave_self_attention: false
+      norm_type: ada_norm
+      num_attention_heads: 16
+      num_layers: 36
+      output_dim: 1024
+      positional_embeddings: null
+    max_seq_len: 1024
+    noise_beta_alpha: 1.5
+    noise_beta_beta: 1.0
+    noise_s: 0.999
+    num_inference_timesteps: 4
+    num_target_vision_tokens: 32
+    num_timestep_buckets: 1000
+    state_dim: 7
+  name: QwenPI_v3
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+    num_vl_layers: 36
+    vl_hidden_dim: 2560
+output_dir: /workspace/starVLA/results/Checkpoints/pi05_flappy_fix_latency_0_200ep
+rl_games:
+  env_eval:
+    enabled: true
+    mid_train:
+      enabled: false
+      interval_steps: 100
+  task: flappy
+run_id: pi05_flappy_fix_latency_0_200ep
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  distributed_backend: none
+  eval_action_classification: true
+  eval_action_classification_interval: null
+  eval_interval: 100
+  eval_num_batches: 100
+  freeze_modules: ''
+  gradient_accumulation_steps: 8
+  gradient_clipping: 1.0
+  is_resume: false
+  learning_rate:
+    action_model: 5.0e-06
+    base: 5.0e-06
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: false
+    weight_decay: 1.0e-08
+  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

zero-latency/flappy/vla/starvla-qwenpi_v3-h1/pi05_flappy_fix_latency_0_200ep/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.813460648059845,
+        0.18653932213783264,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.3898015022277832,
+        0.3898015022277832,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 279882,
+    "num_trajectories": 180
+  }
+}

zero-latency/flappy/vla/starvla-qwenpi_v3-h1/pi05_flappy_fix_latency_0_200ep/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.9196249842643738,
+        0.08037500083446503,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.2717601954936981,
+        0.2717601954936981,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 72000,
+    "num_trajectories": 20
+  }
+}