Publish MIKASA StarVLA zero-latency policy

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenoft-v1/checkpoints/steps_5000_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cabf456e491c6a29d1230c2d11bda7d870c83d049b95b2ef783de3564955db6e
+size 9785060075

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenoft-v1/config.full.yaml

@@ -0,0 +1,81 @@
+run_id: mikasa_grabfast_h1_gb64_mb8_ga8_expandable
+run_root_dir: results/Checkpoints
+seed: 42
+wandb_entity: saberrr-zju
+wandb_project: starvla_mikasa
+is_debug: false
+version_id: '0.21'
+framework:
+  name: QwenOFT
+  qwenvl:
+    base_vlm: ./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: l1
+    action_env_dim: 7
+    state_dim: 7
+    action_horizon: 1
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    lerobot_version: v3.0
+    include_state: false
+    data_root_dir: /home/xinyuan/latency-sensitive-bench-mikasa-h1/outputs/mikasa/grabfast_h1_dataset/data_lerobot
+    data_mix: local/intercept_grab_fast_vla_v0_h1_train
+    eval_data_mix: local/intercept_grab_fast_vla_v0_h1_val
+    custom_mixtures_path: /home/xinyuan/latency-sensitive-bench-mikasa-h1/outputs/mikasa/grabfast_h1_dataset/data_lerobot/mikasa_intercept_grab_fast_h1.json
+    sequential_step_sampling: true
+    per_device_batch_size: 8
+    load_all_data_for_training: true
+    pin_memory: true
+    persistent_workers: true
+    obs_image_size:
+    - 224
+    - 224
+    video_backend: pyav
+trainer:
+  max_train_steps: 5000
+  num_warmup_steps: 100
+  save_interval: 1000
+  eval_interval: 500
+  eval_num_batches: 100
+  learning_rate:
+    base: 3.0e-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: ''
+  freeze_vit: false
+  freeze_tied_embedding: false
+  freeze_llm_layers: []
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  logging_frequency: 5
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 8
+  pretrained_checkpoint: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: true
+  profile_timing:
+    enabled: true
+    log_interval: 10
+checkpoint:
+  save_training_state: true
+  save_pt_file: true
+config_yaml: examples/MIKASA/train_files/starvla_qwenoft_intercept_grab_fast_h1.yaml
+output_dir: results/Checkpoints/mikasa_grabfast_h1_gb64_mb8_ga8_expandable

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenoft-v1/config.yaml

@@ -0,0 +1,58 @@
+checkpoint:
+  save_pt_file: true
+  save_training_state: true
+datasets:
+  vla_data:
+    data_mix: local/intercept_grab_fast_vla_v0_h1_train
+    dataset_py: lerobot_datasets
+    eval_data_mix: local/intercept_grab_fast_vla_v0_h1_val
+    per_device_batch_size: 8
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 7
+    action_hidden_dim: 2560
+    action_horizon: 1
+    action_model_type: MLP
+    loss_type: l1
+  name: QwenOFT
+  qwenvl:
+    attn_implementation: flash_attention_2
+    base_vlm: ./playground/Pretrained_models/Qwen3-VL-4B-Instruct
+    enable_gradient_checkpointing: true
+output_dir: results/Checkpoints/mikasa_grabfast_h1_gb64_mb8_ga8_expandable
+run_id: mikasa_grabfast_h1_gb64_mb8_ga8_expandable
+run_root_dir: results/Checkpoints
+seed: 42
+trainer:
+  eval_interval: 500
+  eval_num_batches: 100
+  freeze_llm_layers: []
+  freeze_modules: ''
+  freeze_tied_embedding: false
+  freeze_vit: false
+  gradient_accumulation_steps: 8
+  learning_rate:
+    action_model: 0.0001
+    base: 3.0e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 5
+  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: true
+    weight_decay: 1.0e-08
+  pretrained_checkpoint: null
+  profile_timing:
+    enabled: true
+    log_interval: 10
+  save_interval: 1000
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: saberrr-zju
+wandb_project: starvla_mikasa

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenoft-v1/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        -0.11131570488214493,
+        0.040192198008298874,
+        -0.24009454250335693,
+        -0.23081380128860474,
+        -0.41221386194229126,
+        0.18129923939704895,
+        0.013331156224012375
+      ],
+      "std": [
+        0.22855518758296967,
+        0.5893623232841492,
+        0.26777687668800354,
+        0.27581456303596497,
+        0.3006018102169037,
+        0.3146352171897888,
+        0.40770089626312256
+      ],
+      "max": [
+        0.3204230070114136,
+        0.9921634197235107,
+        0.6247130036354065,
+        0.8002768158912659,
+        0.6306125521659851,
+        1.0,
+        0.774597704410553
+      ],
+      "min": [
+        -1.0,
+        -1.0,
+        -1.0,
+        -0.9346221089363098,
+        -1.0,
+        -0.5651661157608032,
+        -1.0
+      ],
+      "q01": [
+        -0.8506493151187897,
+        -1.0,
+        -0.9802482903003692,
+        -0.6838930904865265,
+        -1.0,
+        -0.4035292208194733,
+        -0.831850928068161
+      ],
+      "q99": [
+        0.2628934979438781,
+        0.8585556507110594,
+        0.18524706959724418,
+        0.5309111773967736,
+        0.22950317859649652,
+        0.8685460448265068,
+        0.6467343747615811
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        -0.18996892869472504,
+        -0.09119448065757751,
+        0.08502458781003952,
+        -2.616269588470459,
+        0.33758142590522766,
+        -0.17528389394283295,
+        0.04933461919426918
+      ],
+      "std": [
+        0.08925069123506546,
+        0.1439206600189209,
+        0.08522012829780579,
+        1.4712677001953125,
+        0.18586333096027374,
+        0.10226143896579742,
+        0.01495052408427
+      ],
+      "max": [
+        -0.004026889801025391,
+        0.23487472534179688,
+        0.3442644476890564,
+        3.1415677070617676,
+        0.9648659229278564,
+        0.0931963101029396,
+        0.07999999821186066
+      ],
+      "min": [
+        -0.41803836822509766,
+        -0.37668657302856445,
+        0.015313848853111267,
+        -3.1415457725524902,
+        -0.07448819279670715,
+        -0.40120765566825867,
+        0.02547522261738777
+      ],
+      "q01": [
+        -0.3719064712524414,
+        -0.3454733848571777,
+        0.018063589930534363,
+        -3.1388832092285157,
+        -0.029389445297420023,
+        -0.3605163902044296,
+        0.0291226364672184
+      ],
+      "q99": [
+        -0.03631005287170415,
+        0.17424454689025873,
+        0.3186076879501343,
+        3.1387048721313477,
+        0.7408432960510251,
+        0.053878990560769925,
+        0.07999999821186066
+      ]
+    },
+    "num_transitions": 8911,
+    "num_trajectories": 250
+  }
+}

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenoft-v1/dataset_statistics_eval.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        -0.10304215550422668,
+        0.033736441284418106,
+        -0.23379924893379211,
+        -0.22262747585773468,
+        -0.39024659991264343,
+        0.17639224231243134,
+        0.00261433282867074
+      ],
+      "std": [
+        0.21771568059921265,
+        0.5799816250801086,
+        0.2730005979537964,
+        0.2694971263408661,
+        0.30399200320243835,
+        0.32376307249069214,
+        0.4108830988407135
+      ],
+      "max": [
+        0.3265194892883301,
+        0.9437140822410583,
+        0.4582591652870178,
+        0.7764082551002502,
+        0.48147255182266235,
+        1.0,
+        0.710970401763916
+      ],
+      "min": [
+        -1.0,
+        -1.0,
+        -1.0,
+        -0.7210672497749329,
+        -1.0,
+        -0.5640764236450195,
+        -1.0
+      ],
+      "q01": [
+        -0.8553266936540603,
+        -1.0,
+        -1.0,
+        -0.6864310455322266,
+        -0.9788408172130585,
+        -0.45269877403974534,
+        -0.7478000450134278
+      ],
+      "q99": [
+        0.26953339189291,
+        0.8566095483303069,
+        0.184562908411026,
+        0.4908112260699267,
+        0.2535213321447372,
+        0.928561645746231,
+        0.6371558833122251
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        -0.17818187177181244,
+        -0.0853741466999054,
+        0.08197368681430817,
+        -2.7943994998931885,
+        0.33197715878486633,
+        -0.19013282656669617,
+        0.04959884285926819
+      ],
+      "std": [
+        0.09514310210943222,
+        0.14849503338336945,
+        0.0847383365035057,
+        1.0725171566009521,
+        0.1959533393383026,
+        0.11053295433521271,
+        0.015129948034882545
+      ],
+      "max": [
+        -0.017928123474121094,
+        0.23338603973388672,
+        0.3442644476890564,
+        3.1403441429138184,
+        1.1020687818527222,
+        0.09329112619161606,
+        0.07999999821186066
+      ],
+      "min": [
+        -0.3875160217285156,
+        -0.3706817626953125,
+        0.016500279307365417,
+        -3.141533374786377,
+        -0.06919845193624496,
+        -0.4373460114002228,
+        0.02718203142285347
+      ],
+      "q01": [
+        -0.37409615993499756,
+        -0.359880256652832,
+        0.017494037225842477,
+        -3.1352380514144897,
+        -0.031069158706814048,
+        -0.3954590344429016,
+        0.02878103971481323
+      ],
+      "q99": [
+        -0.03042309761047366,
+        0.2084942722320556,
+        0.3209171247482299,
+        3.1326345920562746,
+        0.91342860341072,
+        0.062003581337630675,
+        0.07999999821186066
+      ]
+    },
+    "num_transitions": 948,
+    "num_trajectories": 25
+  }
+}

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenoft-v1/summary.jsonl

@@ -0,0 +1 @@
+{"steps": 5000}