Publish zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenpi_v3-h1/mikasa_intercept_grab_fast_qwenpi_v3_formal_latest_r8gpu8_20260907T082154Z

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenpi_v3-h1/mikasa_intercept_grab_fast_qwenpi_v3_formal_latest_r8gpu8_20260907T082154Z/README.md

@@ -0,0 +1,4 @@
+# qwenpi_v3 — mikasa_intercept_grab_fast
+
+RGB + continuous state, action horizon 1, VLA 20 Hz.
+See task_contract.json for normalization, eval_l0.json for zero-latency results, and provenance.json for exact teacher, backbone, and dataset revisions.

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenpi_v3-h1/mikasa_intercept_grab_fast_qwenpi_v3_formal_latest_r8gpu8_20260907T082154Z/checkpoints/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7467625bb4ce0bc4d7ea7a601191b138750e66216567d00c2ca95a412360f752
+size 10922637469

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenpi_v3-h1/mikasa_intercept_grab_fast_qwenpi_v3_formal_latest_r8gpu8_20260907T082154Z/config.yaml

@@ -0,0 +1,62 @@
+framework:
+  name: QwenPI_v3
+  qwenvl:
+    base_vlm: Qwen/Qwen3-VL-4B-Instruct
+    attn_implementation: flash_attention_2
+    vl_hidden_dim: 2560
+    num_vl_layers: 36
+  action_model:
+    action_model_type: LayerwiseFM
+    action_dim: 7
+    state_dim: 7
+    state_encoding: continuous_projector
+    task_objective: null
+    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
+      num_layers: 36
+      input_embedding_dim: 1024
+      cross_attention_dim: 1024
+      output_dim: 1024
+      num_attention_heads: 16
+    action_env_dim: 7
+    future_action_window_size: 0
+    past_action_window_size: 0
+version_id: '0.21'
+datasets:
+  vla_data:
+    include_state: true
+    obs_image_size:
+    - 224
+    - 224
+    image_mode: single
+    stitch_grid:
+    - 1
+    - 2
+    num_obs_frames: 1
+trainer:
+  pretrained_checkpoint: null
+backbone:
+  repo_id: Qwen/Qwen3-VL-4B-Instruct
+  revision: ebb281ec70b05090aa6165b016eac8ec08e71b17
+  allow_patterns:
+  - '*.json'
+  - '*.safetensors'
+  - '*.txt'
+  - '*.jinja'
+  - '*.model'

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenpi_v3-h1/mikasa_intercept_grab_fast_qwenpi_v3_formal_latest_r8gpu8_20260907T082154Z/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-qwenpi_v3-h1/mikasa_intercept_grab_fast_qwenpi_v3_formal_latest_r8gpu8_20260907T082154Z/provenance.json

@@ -0,0 +1,84 @@
+{
+  "task": "mikasa",
+  "model": "qwenpi_v3",
+  "run_id": "mikasa_intercept_grab_fast_qwenpi_v3_formal_latest_r8gpu8_20260907T082154Z",
+  "smoke": false,
+  "teacher": {
+    "repo_id": "latency-sensitive-bench/benchmark-models",
+    "revision": "46abd9ece5745d808d16aa8013766a26b4f6c371",
+    "path": "zero-latency/mikasa-intercept-grab-fast/small-policy/ppo-v1/checkpoint.pt"
+  },
+  "backbone": {
+    "repo_id": "Qwen/Qwen3-VL-4B-Instruct",
+    "revision": "ebb281ec70b05090aa6165b016eac8ec08e71b17",
+    "allow_patterns": [
+      "*.json",
+      "*.safetensors",
+      "*.txt",
+      "*.jinja",
+      "*.model"
+    ]
+  },
+  "code_revisions": {
+    "benchmark": {
+      "revision": "2f18d4620afd1861059d8cecfc2b1410a4a4777d",
+      "dirty": true
+    },
+    "starvla": {
+      "revision": "73cd1e6175d8c6e5896e742b6a8c8ca3aef52dca",
+      "dirty": false
+    },
+    "haic": {
+      "revision": "b96533e17dfb6d5a1bd3a67032a66acc76e2fd80",
+      "dirty": true
+    },
+    "parkour": {
+      "revision": "26274b2ad83cd6f45d061973a4d102fd492010de",
+      "dirty": false
+    },
+    "mikasa": {
+      "revision": "a5f72ac4faf1f83ea1ccbf7f33ca04173554d3bc",
+      "dirty": true
+    },
+    "gr00t": {
+      "revision": "aa0c592853102010080e4f61609db8e3dc0779d4",
+      "dirty": false
+    }
+  },
+  "datasets": [
+    {
+      "repo_id": "latency-sensitive-bench/benchmark-datasets",
+      "path": "zero-latency/mikasa-intercept-grab-fast/teacher-rollouts-h1-state-v1",
+      "revision": "5421c9d272d7a939bb04045b047af84519c1fba8"
+    }
+  ],
+  "seed": 0,
+  "training": {
+    "sft_updates": 5000,
+    "dagger_updates": 0,
+    "global_batch": 128,
+    "bootstrap_weight": 1.0,
+    "dagger_weight": 0.0,
+    "micro_batch": 16,
+    "num_gpus": 8,
+    "optimizer": {
+      "name": "AdamW",
+      "betas": [
+        0.9,
+        0.95
+      ],
+      "eps": 1e-08,
+      "weight_decay": 1e-08,
+      "fused": true
+    },
+    "learning_rate": {
+      "base": 3e-05,
+      "action_model": 0.0001
+    },
+    "scheduler": "cosine_with_min_lr",
+    "scheduler_kwargs": {
+      "min_lr": 1e-06
+    },
+    "warmup_updates": 100
+  }
+}

zero-latency/mikasa-intercept-grab-fast/vla/starvla-qwenpi_v3-h1/mikasa_intercept_grab_fast_qwenpi_v3_formal_latest_r8gpu8_20260907T082154Z/task_contract.json

@@ -0,0 +1,131 @@
+{
+  "format": "latency_bench_starvla_h1_v1",
+  "task": "mikasa_intercept_grab_fast",
+  "robot_type": "latency_mikasa_h1",
+  "state_dim": 7,
+  "action_dim": 7,
+  "action_horizon": 1,
+  "fps": 20.0,
+  "state_key": "observation.state",
+  "action_key": "action",
+  "video_keys": [
+    "video.top",
+    "video.wrist"
+  ],
+  "auxiliary": {},
+  "normalization": {
+    "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
+      ],
+      "min": [
+        -0.41803836822509766,
+        -0.37668657302856445,
+        0.015313848853111267,
+        -3.1415457725524902,
+        -0.07448819279670715,
+        -0.40120765566825867,
+        0.02547522261738777
+      ],
+      "max": [
+        -0.004026889801025391,
+        0.23487472534179688,
+        0.3442644476890564,
+        3.1415677070617676,
+        0.9648659229278564,
+        0.0931963101029396,
+        0.07999999821186066
+      ],
+      "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
+      ]
+    },
+    "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
+      ],
+      "min": [
+        -1.0,
+        -1.0,
+        -1.0,
+        -0.9346221089363098,
+        -1.0,
+        -0.5651661157608032,
+        -1.0
+      ],
+      "max": [
+        0.3204230070114136,
+        0.9921634197235107,
+        0.6247130036354065,
+        0.8002768158912659,
+        0.6306125521659851,
+        1.0,
+        0.774597704410553
+      ],
+      "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
+      ]
+    }
+  },
+  "prompt": "Intercept the rolling ball and grasp it to stop it."
+}