Publish asterix qwenpi_v3 H8 evaluation artifacts

zero-latency/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-zero-s42-v1/README.md

@@ -1,7 +1,8 @@
-# Asterix qwenpi_v3 zero-latency H8 VLA
+# Asterix qwenpi_v3 zero-trained H8 VLA
 
 Current RGB and eight binary decision-history windows; action horizon one. Initialized from the pinned Qwen3-VL-4B backbone and random action head.
 
 - Training: 5000 steps, seed 42, global batch 128
 - W&B: https://wandb.ai/saberrr-zju/starvla_tasks/runs/ec20df07
-- Formal zero-latency evaluation: pending.
+- Evaluation: 100 episodes, seeds 1000000–1000099, GPU simulator, parallel 32, capacity 1, 60/15 Hz, 7200 raw-frame cap, zero latency
+- Mean raw return: 5758.5

zero-latency/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-zero-s42-v1/eval_l0.json

@@ -0,0 +1,216 @@
+{
+  "checkpoint_path": "/home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/bundles/asterix-qwenpi_v3-hist8-zero-s42-v1/checkpoints/model.pt",
+  "experiment_name": "asterix-qwenpi_v3-hist8-zero-s42-v1-step5000-history01-zero-par32",
+  "latency": 0,
+  "latency_type": "zero",
+  "lengths": [
+    5393,
+    5007,
+    4638,
+    5027,
+    6893,
+    3443,
+    5401,
+    4638,
+    5015,
+    5027,
+    5019,
+    5401,
+    5393,
+    4630,
+    4618,
+    5015,
+    5397,
+    7075,
+    7200,
+    5015,
+    3928,
+    6873,
+    6873,
+    4638,
+    2377,
+    3815,
+    7200,
+    5393,
+    5393,
+    4618,
+    2377,
+    3443,
+    5397,
+    5381,
+    5015,
+    4630,
+    6893,
+    5015,
+    5401,
+    6869,
+    5003,
+    3928,
+    3716,
+    5003,
+    3443,
+    6878,
+    4243,
+    4634,
+    4634,
+    3443,
+    5027,
+    4634,
+    5023,
+    6873,
+    6897,
+    5011,
+    4634,
+    5397,
+    2377,
+    7200,
+    5023,
+    4239,
+    6898,
+    6906,
+    6889,
+    7055,
+    3928,
+    3954,
+    6893,
+    3716,
+    3954,
+    5023,
+    5015,
+    4243,
+    3928,
+    3716,
+    5401,
+    5007,
+    5015,
+    5397,
+    6893,
+    5405,
+    4239,
+    5397,
+    4634,
+    5397,
+    4239,
+    4630,
+    4243,
+    5397,
+    6869,
+    4634,
+    5401,
+    4618,
+    2377,
+    7059,
+    5015,
+    5405,
+    5381,
+    5405
+  ],
+  "mean_length": 5096.1,
+  "mean_return": 5758.5,
+  "returns": [
+    6700.0,
+    5700.0,
+    5700.0,
+    5700.0,
+    9100.0,
+    2250.0,
+    6700.0,
+    5700.0,
+    5700.0,
+    5700.0,
+    5700.0,
+    6700.0,
+    6700.0,
+    5700.0,
+    5700.0,
+    5700.0,
+    6700.0,
+    7900.0,
+    9300.0,
+    5700.0,
+    3000.0,
+    9100.0,
+    9100.0,
+    5700.0,
+    1450.0,
+    2900.0,
+    8300.0,
+    6700.0,
+    6700.0,
+    5700.0,
+    1450.0,
+    2250.0,
+    6700.0,
+    6700.0,
+    5700.0,
+    5700.0,
+    9100.0,
+    5700.0,
+    6700.0,
+    9100.0,
+    5700.0,
+    3000.0,
+    2150.0,
+    5700.0,
+    2250.0,
+    7700.0,
+    3600.0,
+    5700.0,
+    5700.0,
+    2250.0,
+    5700.0,
+    5700.0,
+    5700.0,
+    9100.0,
+    9100.0,
+    5700.0,
+    5700.0,
+    6700.0,
+    1450.0,
+    9300.0,
+    5700.0,
+    3600.0,
+    7700.0,
+    7700.0,
+    9100.0,
+    7900.0,
+    3000.0,
+    2050.0,
+    9100.0,
+    2150.0,
+    2050.0,
+    5700.0,
+    5700.0,
+    3600.0,
+    3000.0,
+    2150.0,
+    6700.0,
+    5700.0,
+    5700.0,
+    6700.0,
+    9100.0,
+    6700.0,
+    3600.0,
+    6700.0,
+    5700.0,
+    6700.0,
+    3600.0,
+    5700.0,
+    3600.0,
+    6700.0,
+    9100.0,
+    5700.0,
+    6700.0,
+    5700.0,
+    1450.0,
+    7900.0,
+    5700.0,
+    6700.0,
+    6700.0,
+    6700.0
+  ],
+  "seed": 1000000,
+  "std_return": 2132.5812411254115,
+  "suite_name": "fixed_0",
+  "timestamp_utc": "2026-09-24T12:19:22.283186+00:00"
+}

zero-latency/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-zero-s42-v1/evaluation/action_audit.json

@@ -0,0 +1,153 @@
+{
+  "eval_protocol_version": "hist8-history01-sf-render-v4",
+  "history_encoding": "binary_0_1",
+  "episodes": 100,
+  "episode_seeds": [
+    1000000,
+    1000001,
+    1000002,
+    1000003,
+    1000004,
+    1000005,
+    1000006,
+    1000007,
+    1000008,
+    1000009,
+    1000010,
+    1000011,
+    1000012,
+    1000013,
+    1000014,
+    1000015,
+    1000016,
+    1000017,
+    1000018,
+    1000019,
+    1000020,
+    1000021,
+    1000022,
+    1000023,
+    1000024,
+    1000025,
+    1000026,
+    1000027,
+    1000028,
+    1000029,
+    1000030,
+    1000031,
+    1000032,
+    1000033,
+    1000034,
+    1000035,
+    1000036,
+    1000037,
+    1000038,
+    1000039,
+    1000040,
+    1000041,
+    1000042,
+    1000043,
+    1000044,
+    1000045,
+    1000046,
+    1000047,
+    1000048,
+    1000049,
+    1000050,
+    1000051,
+    1000052,
+    1000053,
+    1000054,
+    1000055,
+    1000056,
+    1000057,
+    1000058,
+    1000059,
+    1000060,
+    1000061,
+    1000062,
+    1000063,
+    1000064,
+    1000065,
+    1000066,
+    1000067,
+    1000068,
+    1000069,
+    1000070,
+    1000071,
+    1000072,
+    1000073,
+    1000074,
+    1000075,
+    1000076,
+    1000077,
+    1000078,
+    1000079,
+    1000080,
+    1000081,
+    1000082,
+    1000083,
+    1000084,
+    1000085,
+    1000086,
+    1000087,
+    1000088,
+    1000089,
+    1000090,
+    1000091,
+    1000092,
+    1000093,
+    1000094,
+    1000095,
+    1000096,
+    1000097,
+    1000098,
+    1000099
+  ],
+  "issued_actions": 127448,
+  "action_counts": {
+    "6": 41388,
+    "3": 9645,
+    "7": 10390,
+    "4": 2204,
+    "5": 4494,
+    "1": 3244,
+    "8": 30040,
+    "0": 25337,
+    "2": 706
+  },
+  "largest_action_fraction": 0.32474420940305065,
+  "raw_score_min": [
+    -0.08517017960548401,
+    -0.10428443551063538,
+    -0.08317434787750244,
+    -0.09850794076919556,
+    -0.09002876281738281,
+    -0.07311639189720154,
+    -0.26507842540740967,
+    -0.08630828559398651,
+    -0.16656984388828278
+  ],
+  "raw_score_max": [
+    1.112355351448059,
+    1.2103047370910645,
+    1.1933003664016724,
+    1.1528035402297974,
+    1.1515668630599976,
+    1.1958045959472656,
+    1.1107509136199951,
+    1.2056286334991455,
+    1.1282782554626465
+  ],
+  "raw_score_mean": [
+    0.19860854102436143,
+    0.022158051565253748,
+    0.00810792909272446,
+    0.06813755351500361,
+    0.014625414859000431,
+    0.03114063633102587,
+    0.3257789390649875,
+    0.0847539822291615,
+    0.23716105913420343
+  ]
+}

zero-latency/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-zero-s42-v1/evaluation/eval_config.yaml

@@ -0,0 +1,161 @@
+experiment:
+  name: asterix-qwenpi_v3-hist8-zero-s42-v1-step5000-history01-zero-par32
+  seed: 1000000
+backend:
+  type: sample_factory
+  algo: APPO
+  device: cuda
+  train_dir: results/sample_factory
+  restart_behavior: resume
+  run_mode: eval
+  extra_args:
+  - --encoder_conv_architecture
+  - convnet_simple
+executor:
+  mode: simulated
+  simulated_worker_capacity: 1
+  simulated_inference_pool: true
+  inference_devices:
+  - cuda:0
+  - cuda:1
+  inference_batch_size: 16
+env:
+  name: gymnasium
+  task_name: asterix
+  env_id: LatencyBench/Asterix-v0
+  registration_imports:
+  - latency_bench.envs.gymnasium_asterix
+  make_kwargs:
+    base_env_id: ALE/Asterix-v5
+    render_mode: rgb_array
+    screen_size: 84
+    noop_max: 30
+    base_make_kwargs:
+      obs_type: rgb
+      frameskip: 1
+      repeat_action_probability: 0.0
+      full_action_space: false
+      mode: 0
+      difficulty: 0
+      max_num_frames_per_episode: 108000
+  env_fps: 60
+  obs_fps: 15
+  frame_stack: 1
+  simulator: gpu
+  terminal_on_life_loss: false
+  clip_reward: false
+  noop_action: noop
+  action_map:
+    noop: 0
+    up: 1
+    right: 2
+    left: 3
+    down: 4
+    upright: 5
+    upleft: 6
+    downright: 7
+    downleft: 8
+  action_order:
+  - noop
+  - up
+  - right
+  - left
+  - down
+  - upright
+  - upleft
+  - downright
+  - downleft
+  base_prompt: 'Collect useful objects and avoid lyres. Choose exactly one action
+    from: noop, up, right, left, down, upright, upleft, downright, downleft.'
+  obs_resize:
+  - 224
+  - 224
+  gpu_env_device: auto
+  noop_max: 30
+  action_history_decisions: 8
+latency:
+  method: zero
+scheduler:
+  hold_policy: hold
+  ordering_policy: issue_order_fifo
+policy:
+  type: starvla
+  actions:
+  - noop
+  - up
+  - right
+  - left
+  - down
+  - upright
+  - upleft
+  - downright
+  - downleft
+  checkpoint_path: /home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/bundles/asterix-qwenpi_v3-hist8-zero-s42-v1/checkpoints/model.pt
+  model_config_path: /home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/bundles/asterix-qwenpi_v3-hist8-zero-s42-v1/config.yaml
+  task_manifest_path: /home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/bundles/asterix-qwenpi_v3-hist8-zero-s42-v1/manifest.json
+  device: cuda:0
+  unnorm_key: new_embodiment
+  prompt_mode: latency_neutral
+  action_layout: discrete_9
+  backbone_path: /home/ubuntu/.cache/huggingface/hub/models--Qwen--Qwen3-VL-4B-Instruct/snapshots/ebb281ec70b05090aa6165b016eac8ec08e71b17
+  worker_python_executable: /home/ubuntu/miniconda3/envs/starvla_rl_games_openvla/bin/python3.10
+  state_source: transport
+training:
+  train_for_env_steps: 1000000
+  num_workers: 4
+  num_envs_per_worker: 16
+  worker_num_splits: 1
+  num_policies: 1
+  batch_size: 1024
+  rollout: 64
+  recurrence: 32
+  num_epochs: 2
+  num_batches_per_epoch: 8
+  max_policy_lag: 300
+  learning_rate: 0.000303
+  lr_schedule: constant
+  nonlinearity: elu
+  gamma: 0.99
+  gae_lambda: 0.95
+  ppo_clip_ratio: 0.1
+  ppo_clip_value: 0.2
+  exploration_loss: entropy
+  exploration_loss_coeff: 0.003
+  value_loss_coeff: 0.5
+  max_grad_norm: 0.0
+  adam_eps: 1.0e-06
+  obs_scale: 1.0
+  adaptive_stddev: false
+  with_vtrace: false
+  async_rl: true
+  use_rnn: false
+  normalize_input: true
+  normalize_returns: true
+  stats_avg: 100
+  experiment_summaries_interval: 1
+  save_every_sec: 600
+  keep_checkpoints: 5
+evaluation:
+  eval_interval_steps: null
+  eval_episodes: 100
+  eval_parallel_envs: 32
+  eval_max_steps: 7200
+  eval_deterministic: true
+  eval_suites:
+    fixed: []
+    normal: []
+    uniform: []
+  latency_bench_env_backend: asterix_gpu_batched
+  eval_raw_reward: true
+logging:
+  output_dir: /home/ubuntu/codex/atari-hist8-20260924/outputs/tasks/atari_hist8_v1/evals/asterix-qwenpi_v3-hist8-zero-s42-v1/zero_to_zero
+  video:
+    enabled: false
+    num_bins: 1
+  save_step_records: true
+  save_action_records: true
+  save_latency_records: true
+  wandb_project: null
+  wandb_group: null
+  wandb_job_type: null
+  wandb_tags: ''

zero-latency/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-zero-s42-v1/evaluation/provenance.json

@@ -0,0 +1,46 @@
+{
+  "collection": {
+    "episodes": 300,
+    "filter_preset": "none",
+    "history": "decision_v1",
+    "history_decisions": 8,
+    "history_source": "teacher_transport",
+    "max_episode_raw_frames": 7200,
+    "rollout_overrides": {
+      "env_fps": 60,
+      "fixed_latency_ms": 0.0,
+      "latency_type": "fixed",
+      "obs_fps": 15,
+      "seed": 42
+    },
+    "student_image_frames": 1,
+    "teacher_checkpoint": "best_000024112_12345344_reward_5614.500.pth",
+    "teacher_path": "zero-latency/asterix/small-policy/sample-factory-hist8-cpu15m-s0-v1",
+    "teacher_repo": "latency-sensitive-bench/benchmark-models",
+    "teacher_revision": "dfaeecb84cd6b5dfcc3adc0935d0ae86d0b504d2",
+    "teacher_sha256": "0fe54ff4b3a2ccec438aca789f1a9f86fcef7cda44cda5156e3b774262c50beb",
+    "train_labels": 337557,
+    "training_latency_condition": "zero"
+  },
+  "training_run_id": "asterix-qwenpi_v3-hist8-zero-s42-v1",
+  "training_step": 5000,
+  "global_batch": 128,
+  "base_model": "Qwen/Qwen3-VL-4B-Instruct",
+  "wandb_run_id": "ec20df07",
+  "checkpoint_sha256": "381d1ed73a6f3949cb1880d6f1feb18db13068aeb6db39a7974ae9fb1fba0f56",
+  "eval_protocol_version": "hist8-history01-sf-render-v4",
+  "history_encoding": "binary_0_1",
+  "episodes": 100,
+  "env_condition": "zero",
+  "parallel_envs": 32,
+  "max_episode_raw_frames": 7200,
+  "inference_devices": [
+    "cuda:0",
+    "cuda:1"
+  ],
+  "episode_seeds": [
+    1000000,
+    1000099
+  ],
+  "profile_sha256": null
+}

zero-latency/asterix/vla/starvla-qwenpi_v3-h1/asterix-qwenpi_v3-hist8-zero-s42-v1/publication_status.json

@@ -1,4 +1,7 @@
 {
+  "episodes": 100,
+  "mean_return": 5758.5,
   "training_status": "complete",
-  "zero_evaluation_status": "pending"
+  "zero_evaluation_status": "complete",
+  "zero_mean_return": 5758.5
 }