Upload folder using huggingface_hub

pi05_deadly_corridor_fix_latency_0_1000ep/checkpoints/steps_500_state/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78f85251f875a0763fda512abbefa6dcdc25423f33b943a9a985fd34860afb7b
+size 11412998004

pi05_deadly_corridor_fix_latency_0_1000ep/checkpoints/steps_500_state/optimizer.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1169237449cec9abdbdcba181bab6dac0ff3fbc7f1495c31c8d1f908bd15218
+size 22793150843

pi05_deadly_corridor_fix_latency_0_1000ep/checkpoints/steps_500_state/random_states_0.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4982813c7d8e9ab37cb1f5ae79c64f24ffa676f72ac9a387315a71881ea673fc
+size 14821

pi05_deadly_corridor_fix_latency_0_1000ep/config.full.yaml

@@ -0,0 +1,237 @@
+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: true
+      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: 7
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/deadly_corridor_fix_latency_0_1000ep
+    data_mix: deadly_corridor_train__bridge
+    eval_data_mix: deadly_corridor_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    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: deadly_corridor_train
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: deadly_corridor_train
+  mixed_converted_name: deadly_corridor_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  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: 500
+  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: 4
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: You are playing Deadly Corridor in VizDoom. Choose actions from
+      MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.
+    deadly:
+      action_layout: multibinary_7
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: /workspace/starVLA/data/deadly_corridor_fix_latency_0_1000ep/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: null
+  initialization_mode: bridge
+  action_carrier: bridge
+model: pi05
+env: deadly_corridor
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: 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_deadly_corridor_fix_latency_0_1000ep
+output_dir: /workspace/starVLA/results/Checkpoints/pi05_deadly_corridor_fix_latency_0_1000ep
+config_yaml: null
+is_debug: false
+version_id: '0.21'

pi05_deadly_corridor_fix_latency_0_1000ep/config.yaml

@@ -0,0 +1,99 @@
+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: deadly_corridor_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: deadly_corridor_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: 7
+    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: true
+      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_deadly_corridor_fix_latency_0_1000ep
+rl_games:
+  env_eval:
+    deadly:
+      action_layout: multibinary_7
+    enabled: true
+    mid_train:
+      enabled: false
+      interval_steps: 100
+  task: deadly_corridor
+run_id: pi05_deadly_corridor_fix_latency_0_1000ep
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  cc_f1_tolerance: 1
+  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: 500
+  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

pi05_deadly_corridor_fix_latency_0_1000ep/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.924872100353241,
+        0.0039678397588431835,
+        0.16351675987243652,
+        0.6617416739463806,
+        0.23405033349990845,
+        0.11331836879253387,
+        0.5559674501419067
+      ],
+      "std": [
+        0.26355794072151184,
+        0.06286244094371796,
+        0.36991649866104126,
+        0.47308459877967834,
+        0.42337605357170105,
+        0.3170303404331207,
+        0.49692174792289734
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        0.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 38308,
+    "num_trajectories": 900
+  }
+}

pi05_deadly_corridor_fix_latency_0_1000ep/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.9250290989875793,
+        0.004190919455140829,
+        0.1720605343580246,
+        0.6565774083137512,
+        0.23189756274223328,
+        0.11525028944015503,
+        0.5634458661079407
+      ],
+      "std": [
+        0.26333996653556824,
+        0.06459999084472656,
+        0.37743079662323,
+        0.4748457670211792,
+        0.4220407009124756,
+        0.319315642118454,
+        0.4959460198879242
+      ],
+      "max": [
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        0.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 4295,
+    "num_trajectories": 100
+  }
+}

pi05_deadly_corridor_fix_latency_0_1000ep/hydra/.hydra/config.yaml

@@ -0,0 +1,231 @@
+framework:
+  qwenvl:
+    base_vlm: /workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    enable_gradient_checkpointing: true
+  action_model:
+    state_dim: 7
+    loss_type: l1
+    action_horizon: 1
+    future_action_window_size: 0
+    past_action_window_size: 0
+    action_model_type: LayerwiseFM
+    action_dim: 7
+    action_env_dim: 7
+    repeated_diffusion_steps: 2
+    num_inference_timesteps: 4
+    num_target_vision_tokens: 32
+    add_pos_embed: true
+    max_seq_len: 1024
+    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
+      output_dim: 1024
+      dropout: 0.2
+      final_dropout: true
+      interleave_self_attention: true
+      norm_type: ada_norm
+      positional_embeddings: null
+      attention_head_dim: 64
+  name: QwenPI_v3
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /workspace/starVLA/data/deadly_corridor_fix_latency_0_1000ep
+    data_mix: deadly_corridor_train__bridge
+    eval_data_mix: deadly_corridor_train__bridge__val
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    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: ${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: deadly_corridor_train
+  mixed_converted_name: deadly_corridor_mixed_latency_train
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  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-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: 500
+  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: /workspace/starVLA/playground/Pretrained_models/Qwen3VL-PI_v3-Bridge-RT_1/checkpoints/steps_50000_pytorch_model.pt
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: false
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /workspace/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: dongqianyu99-zhejiang-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: playground/Datasets/rl_games
+  dataset_cache_dir: null
+  base_model_dir: ${bridge_base_model.local_dir.${model}}
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: pi-0.5
+  env_eval:
+    image_size: 224
+    frameskip: 4
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: You are playing Deadly Corridor in VizDoom. Choose actions from
+      MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.
+    deadly:
+      action_layout: multibinary_7
+    enabled: true
+    distributed_mode: none
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: /workspace/starVLA/data/deadly_corridor_fix_latency_0_1000ep/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: null
+  initialization_mode: bridge
+  action_carrier: bridge
+model: pi05
+env: deadly_corridor
+init: bridge
+bridge_base_model:
+  repo_id:
+    openvla: Qwen/Qwen3-VL-4B-Instruct
+    pi0: StarVLA/Qwen2.5-VL-3B-Instruct-Action
+    pi05: Qwen/Qwen3-VL-4B-Instruct
+    gr00t: Qwen/Qwen3-VL-4B-Instruct
+  local_dir:
+    openvla: 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_deadly_corridor_fix_latency_0_1000ep
+output_dir: null
+config_yaml: null
+is_debug: false
+version_id: '0.21'

pi05_deadly_corridor_fix_latency_0_1000ep/hydra/.hydra/hydra.yaml

@@ -0,0 +1,277 @@
+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=pi05
+    - env=deadly_corridor
+    - init=bridge
+    - mode=single
+    - run_id=pi05_deadly_corridor_fix_latency_0_1000ep
+    - run_root_dir=/workspace/starVLA/results/Checkpoints
+    - seed=42
+    - wandb_entity=dongqianyu99-zhejiang-university
+    - wandb_project=starVLA_rl_games
+    - rl_games.env_eval.enabled=true
+    - checkpoint.sync.enabled=false
+    - checkpoint.sync.keep_last_n=0
+    - checkpoint.local.keep_last_n=1
+    - checkpoint.save_best_model=false
+    - checkpoint.save_pt_file=false
+    - trainer.is_resume=false
+    - trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-PI_v3-Bridge-RT_1/checkpoints/steps_50000_pytorch_model.pt
+    - trainer.resume_step=0
+    - trainer.max_train_steps=500
+    - 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=8
+    - trainer.distributed_backend=none
+    - trainer.learning_rate.base=5e-06
+    - trainer.learning_rate.qwen_vl_interface=1e-05
+    - trainer.learning_rate.action_model=5e-06
+    - 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=QwenPI_v3
+    - framework.qwenvl.attn_implementation=flash_attention_2
+    - framework.qwenvl.enable_gradient_checkpointing=true
+    - framework.action_model.action_model_type=LayerwiseFM
+    - framework.action_model.action_dim=7
+    - framework.action_model.action_env_dim=7
+    - framework.action_model.state_dim=7
+    - framework.action_model.loss_type=l1
+    - framework.action_model.action_horizon=1
+    - framework.action_model.future_action_window_size=0
+    - framework.action_model.past_action_window_size=0
+    - framework.action_model.repeated_diffusion_steps=2
+    - framework.action_model.num_inference_timesteps=4
+    - framework.action_model.num_target_vision_tokens=32
+    - framework.action_model.add_pos_embed=true
+    - framework.action_model.max_seq_len=1024
+    - framework.action_model.noise_beta_alpha=1.5
+    - framework.action_model.noise_beta_beta=1.0
+    - framework.action_model.noise_s=0.999
+    - framework.action_model.num_timestep_buckets=1000
+    - framework.action_model.diffusion_model_cfg.action_dit_hidden_dim=1024
+    - framework.action_model.diffusion_model_cfg.dropout=0.2
+    - framework.action_model.diffusion_model_cfg.final_dropout=true
+    - framework.action_model.diffusion_model_cfg.interleave_self_attention=true
+    - framework.action_model.diffusion_model_cfg.norm_type=ada_norm
+    - framework.action_model.diffusion_model_cfg.output_dim=1024
+    - framework.action_model.diffusion_model_cfg.attention_head_dim=64
+    - 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=16
+    - datasets.vla_data.load_all_data_for_training=true
+    - datasets.vla_data.video_backend=torchvision_av
+    - datasets.vla_data.data_root_dir=/workspace/starVLA/data/deadly_corridor_fix_latency_0_1000ep
+    - datasets.vla_data.data_mix=deadly_corridor_train__bridge
+    - datasets.vla_data.eval_data_mix=deadly_corridor_train__bridge__val
+    - framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    - rl_games.task=deadly_corridor
+    - rl_games.model_alias=pi-0.5
+    - rl_games.initialization_mode=bridge
+    - rl_games.action_carrier=bridge
+    - rl_games.env_eval.distributed_mode=none
+    - rl_games.env_eval.latency.mode=single
+    - rl_games.env_eval.frameskip=4
+    - rl_games.env_eval.image_size=224
+    - rl_games.env_eval.seed=42
+    - rl_games.env_eval.fixed_episode_seeds=true
+    - rl_games.env_eval.latency_seed_stride=0
+    - rl_games.env_eval.task_seed_stride=0
+    - rl_games.env_eval.task_description='You are playing Deadly Corridor in VizDoom.
+      Choose actions from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT,
+      TURN_RIGHT, ATTACK.'
+    - rl_games.env_eval.latency.values=[0]
+    - rl_games.env_eval.latency.prompt_map_path=/workspace/starVLA/data/deadly_corridor_fix_latency_0_1000ep/deadly_corridor_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]
+    - rl_games.env_eval.deadly.action_layout=multibinary_7
+  job:
+    name: train_starvla_hydra
+    chdir: false
+    override_dirname: checkpoint.local.keep_last_n=1,checkpoint.save_best_model=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=deadly_corridor_train__bridge,datasets.vla_data.data_root_dir=/workspace/starVLA/data/deadly_corridor_fix_latency_0_1000ep,datasets.vla_data.eval_data_mix=deadly_corridor_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=16,datasets.vla_data.sequential_step_sampling=false,datasets.vla_data.shuffle=true,datasets.vla_data.video_backend=torchvision_av,env=deadly_corridor,framework.action_model.action_dim=7,framework.action_model.action_env_dim=7,framework.action_model.action_horizon=1,framework.action_model.action_model_type=LayerwiseFM,framework.action_model.add_pos_embed=true,framework.action_model.diffusion_model_cfg.action_dit_hidden_dim=1024,framework.action_model.diffusion_model_cfg.attention_head_dim=64,framework.action_model.diffusion_model_cfg.dropout=0.2,framework.action_model.diffusion_model_cfg.final_dropout=true,framework.action_model.diffusion_model_cfg.interleave_self_attention=true,framework.action_model.diffusion_model_cfg.norm_type=ada_norm,framework.action_model.diffusion_model_cfg.output_dim=1024,framework.action_model.future_action_window_size=0,framework.action_model.loss_type=l1,framework.action_model.max_seq_len=1024,framework.action_model.noise_beta_alpha=1.5,framework.action_model.noise_beta_beta=1.0,framework.action_model.noise_s=0.999,framework.action_model.num_inference_timesteps=4,framework.action_model.num_target_vision_tokens=32,framework.action_model.num_timestep_buckets=1000,framework.action_model.past_action_window_size=0,framework.action_model.repeated_diffusion_steps=2,framework.action_model.state_dim=7,framework.name=QwenPI_v3,framework.qwenvl.attn_implementation=flash_attention_2,framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct,framework.qwenvl.enable_gradient_checkpointing=true,init=bridge,mode=single,model=pi05,rl_games.action_carrier=bridge,rl_games.env_eval.deadly.action_layout=multibinary_7,rl_games.env_eval.distributed_mode=none,rl_games.env_eval.enabled=true,rl_games.env_eval.fixed_episode_seeds=true,rl_games.env_eval.frameskip=4,rl_games.env_eval.image_size=224,rl_games.env_eval.latency.mode=single,rl_games.env_eval.latency.prompt_map_path=/workspace/starVLA/data/deadly_corridor_fix_latency_0_1000ep/deadly_corridor_train__bridge/latency_prompt_map.json,rl_games.env_eval.latency.values=[0],rl_games.env_eval.latency_seed_stride=0,rl_games.env_eval.mid_train.enabled=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 Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD,
+      MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.',rl_games.env_eval.task_seed_stride=0,rl_games.initialization_mode=bridge,rl_games.model_alias=pi-0.5,rl_games.task=deadly_corridor,run_id=pi05_deadly_corridor_fix_latency_0_1000ep,run_root_dir=/workspace/starVLA/results/Checkpoints,seed=42,trainer.distributed_backend=none,trainer.eval_interval=100,trainer.eval_num_batches=100,trainer.gradient_accumulation_steps=8,trainer.gradient_clipping=1.0,trainer.is_resume=false,trainer.learning_rate.action_model=5e-06,trainer.learning_rate.base=5e-06,trainer.learning_rate.qwen_vl_interface=1e-05,trainer.logging_frequency=1,trainer.loss_scale.vla=1.0,trainer.loss_scale.vlm=0.1,trainer.lr_scheduler_type=cosine_with_min_lr,trainer.max_grad_norm=1.0,trainer.max_train_steps=500,trainer.num_warmup_steps=100,trainer.optimizer.betas=[0.9,0.95],trainer.optimizer.eps=1e-08,trainer.optimizer.fused=false,trainer.optimizer.name=AdamW,trainer.optimizer.weight_decay=1e-08,trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-PI_v3-Bridge-RT_1/checkpoints/steps_50000_pytorch_model.pt,trainer.resume_step=0,trainer.save_format=pt,trainer.save_interval=100,trainer.scheduler_specific_kwargs.min_lr=1e-06,trainer.weight_decay=0.0,wandb_entity=dongqianyu99-zhejiang-university,wandb_project=starVLA_rl_games
+    id: ???
+    num: ???
+    config_name: train
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.3
+    version_base: '1.1'
+    cwd: /workspace/starVLA
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /workspace/starVLA/examples/rl_games/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /workspace/starVLA/results/Checkpoints/pi05_deadly_corridor_fix_latency_0_1000ep/hydra
+    choices:
+      cross_task_setup: null
+      checkpoint: default
+      mode: single
+      init: bridge
+      env: deadly_corridor
+      model: pi05
+      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

pi05_deadly_corridor_fix_latency_0_1000ep/hydra/.hydra/overrides.yaml

@@ -0,0 +1,115 @@
+- model=pi05
+- env=deadly_corridor
+- init=bridge
+- mode=single
+- run_id=pi05_deadly_corridor_fix_latency_0_1000ep
+- run_root_dir=/workspace/starVLA/results/Checkpoints
+- seed=42
+- wandb_entity=dongqianyu99-zhejiang-university
+- wandb_project=starVLA_rl_games
+- rl_games.env_eval.enabled=true
+- checkpoint.sync.enabled=false
+- checkpoint.sync.keep_last_n=0
+- checkpoint.local.keep_last_n=1
+- checkpoint.save_best_model=false
+- checkpoint.save_pt_file=false
+- trainer.is_resume=false
+- trainer.pretrained_checkpoint=/workspace/starVLA/playground/Pretrained_models/Qwen3VL-PI_v3-Bridge-RT_1/checkpoints/steps_50000_pytorch_model.pt
+- trainer.resume_step=0
+- trainer.max_train_steps=500
+- 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=8
+- trainer.distributed_backend=none
+- trainer.learning_rate.base=5e-06
+- trainer.learning_rate.qwen_vl_interface=1e-05
+- trainer.learning_rate.action_model=5e-06
+- 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=QwenPI_v3
+- framework.qwenvl.attn_implementation=flash_attention_2
+- framework.qwenvl.enable_gradient_checkpointing=true
+- framework.action_model.action_model_type=LayerwiseFM
+- framework.action_model.action_dim=7
+- framework.action_model.action_env_dim=7
+- framework.action_model.state_dim=7
+- framework.action_model.loss_type=l1
+- framework.action_model.action_horizon=1
+- framework.action_model.future_action_window_size=0
+- framework.action_model.past_action_window_size=0
+- framework.action_model.repeated_diffusion_steps=2
+- framework.action_model.num_inference_timesteps=4
+- framework.action_model.num_target_vision_tokens=32
+- framework.action_model.add_pos_embed=true
+- framework.action_model.max_seq_len=1024
+- framework.action_model.noise_beta_alpha=1.5
+- framework.action_model.noise_beta_beta=1.0
+- framework.action_model.noise_s=0.999
+- framework.action_model.num_timestep_buckets=1000
+- framework.action_model.diffusion_model_cfg.action_dit_hidden_dim=1024
+- framework.action_model.diffusion_model_cfg.dropout=0.2
+- framework.action_model.diffusion_model_cfg.final_dropout=true
+- framework.action_model.diffusion_model_cfg.interleave_self_attention=true
+- framework.action_model.diffusion_model_cfg.norm_type=ada_norm
+- framework.action_model.diffusion_model_cfg.output_dim=1024
+- framework.action_model.diffusion_model_cfg.attention_head_dim=64
+- 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=16
+- datasets.vla_data.load_all_data_for_training=true
+- datasets.vla_data.video_backend=torchvision_av
+- datasets.vla_data.data_root_dir=/workspace/starVLA/data/deadly_corridor_fix_latency_0_1000ep
+- datasets.vla_data.data_mix=deadly_corridor_train__bridge
+- datasets.vla_data.eval_data_mix=deadly_corridor_train__bridge__val
+- framework.qwenvl.base_vlm=/workspace/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+- rl_games.task=deadly_corridor
+- rl_games.model_alias=pi-0.5
+- rl_games.initialization_mode=bridge
+- rl_games.action_carrier=bridge
+- rl_games.env_eval.distributed_mode=none
+- rl_games.env_eval.latency.mode=single
+- rl_games.env_eval.frameskip=4
+- rl_games.env_eval.image_size=224
+- rl_games.env_eval.seed=42
+- rl_games.env_eval.fixed_episode_seeds=true
+- rl_games.env_eval.latency_seed_stride=0
+- rl_games.env_eval.task_seed_stride=0
+- rl_games.env_eval.task_description='You are playing Deadly Corridor in VizDoom.
+  Choose actions from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT,
+  TURN_RIGHT, ATTACK.'
+- rl_games.env_eval.latency.values=[0]
+- rl_games.env_eval.latency.prompt_map_path=/workspace/starVLA/data/deadly_corridor_fix_latency_0_1000ep/deadly_corridor_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]
+- rl_games.env_eval.deadly.action_layout=multibinary_7

pi05_deadly_corridor_fix_latency_0_1000ep/summary.jsonl

@@ -0,0 +1,5 @@
+{"steps": 100}
+{"steps": 200}
+{"steps": 300}
+{"steps": 400}
+{"steps": 500}