Consolidate paper experiment assets by task and run (901-1050)

.gitattributes

@@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+evaluations/deadly-corridor/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/train/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/episode_metrics.jsonl filter=lfs diff=lfs merge=lfs -text
+evaluations/flappy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/train/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/episode_metrics.jsonl filter=lfs diff=lfs merge=lfs -text

evaluations/deadly-corridor/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/train/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/episode_metrics.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47e9f369b1c80159e15bd6ed5db7be2b4f9b5377fbdbe329701de93ad7352a08
+size 192030635

evaluations/deadly-corridor/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/train/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/git.diff

@@ -0,0 +1,1741 @@
+diff --git a/configs/experiments/fixed_latency_curve.yaml b/configs/experiments/fixed_latency_curve.yaml
+deleted file mode 100644
+index 3ab184f..0000000
+--- a/configs/experiments/fixed_latency_curve.yaml
++++ /dev/null
+@@ -1,17 +0,0 @@
+-__base__: ../base/dummy.yaml
+-
+-experiment:
+-  name: fixed_latency_curve
+-
+-executor:
+-  mode: simulated
+-
+-latency:
+-  method: fixed
+-  fixed_latency_ms: 100
+-
+-evaluation:
+-  eval_episodes: 20
+-
+-logging:
+-  output_dir: outputs/fixed_latency_curve
+diff --git a/configs/experiments/flappy/frame_stack/base.yaml b/configs/experiments/flappy/frame_stack/base.yaml
+deleted file mode 100644
+index f1557d5..0000000
+--- a/configs/experiments/flappy/frame_stack/base.yaml
++++ /dev/null
+@@ -1,123 +0,0 @@
+-experiment:
+-  name: flappy_base
+-  seed: 0
+-
+-backend:
+-  type: sample_factory
+-  algo: APPO
+-  device: cuda
+-  train_dir: results/checkpoints
+-  restart_behavior: resume
+-  run_mode: train
+-
+-executor:
+-  mode: simulated
+-
+-env:
+-  name: flappy
+-  # note that env_fps must be an integer multiple of obs_fps!
+-  env_fps: 30
+-  obs_fps: 30
+-  gym_id: FlappyBird-v0
+-  render_mode: rgb_array
+-  observation_mode: image
+-  use_lidar: false
+-  normalize_obs: true
+-  audio_on: false
+-  frame_stack: 4
+-  obs_resize:
+-    - 84
+-    - 84
+-  # for training it could be true, but for realtime mode, THIS MUST BE FALSE!
+-  use_gpu_render: true
+-  simulator: gpu
+-  gpu_render_device: auto
+-  gpu_render_profile: false
+-  gpu_render_profile_interval: 200
+-  noop_action: noop
+-  action_map:
+-    noop: 0
+-    flap: 1
+-  oneshot_actions:
+-    - flap
+-
+-latency:
+-  method: normal
+-  mean_latency_ms: 66.7
+-  std_latency_ms: 33.3
+-  min_latency_ms: 0
+-  max_latency_ms: 500.0
+-
+-
+-scheduler:
+-  # [one_frame_then_noop, hold]
+-  # if hold, default hold duration is: env_fps / obs_fps
+-  hold_policy: one_frame_then_noop
+-  # can be either `issue_order_fifo` or `latest ready`
+-  ordering_policy: latest_ready
+-
+-policy:
+-  type: random
+-  actions:
+-    - noop
+-    - flap
+-
+-training:
+-  train_for_env_steps: 25000000
+-  num_workers: 2
+-  num_envs_per_worker: 128
+-  worker_num_splits: 1
+-  num_policies: 1
+-  batch_size: 4096
+-  rollout: 128
+-  recurrence: 1
+-  num_epochs: 8
+-  num_batches_per_epoch: 8
+-  max_policy_lag: 400
+-  learning_rate: 0.00025
+-  gamma: 0.99
+-  gae_lambda: 0.95
+-  ppo_clip_ratio: 0.1
+-  ppo_clip_value: 0.2
+-  value_loss_coeff: 0.5
+-  max_grad_norm: 0.5
+-  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: 1000000
+-  eval_episodes: 100
+-  eval_parallel_envs: 100
+-  eval_max_steps: 3600
+-  eval_deterministic: true
+-  eval_raw_reward: false
+-  eval_suites:
+-    fixed:
+-      - [0, 8]    # [start, end]
+-    normal:
+-      - [2.0, 0.5, 0, 15]   # low latency, low jitter
+-      - [2.0, 1.0, 0, 15]   # low latency, medium jitter
+-      - [4.0, 0.5, 0, 15]   # high latency, low jitter
+-      - [4.0, 1.0, 0, 15]   # high latency, medium jitter
+-      - [4.0, 2.0, 0, 15]   # high latency, high jitter
+-    uniform:
+-      - [1, 3]  # low latency, low jitter
+-      - [0, 4]  # low latency, high jitter
+-      - [3, 5]  # high latency, low jitter
+-      - [2, 6]  # high latency, high jitter
+-
+-logging:
+-  output_dir: outputs/flappy_base
+-  save_step_records: true
+-  save_action_records: true
+-  save_latency_records: true
+-  wandb_project: latency-sensitive-bench
+-  wandb_group: flappy-frame_stack
+-  wandb_job_type: sample_factory
+-  wandb_tags: ""
+diff --git a/latency_bench/data/exporters/starvla_lerobot.py b/latency_bench/data/exporters/starvla_lerobot.py
+index c693460..4ad276a 100644
+--- a/latency_bench/data/exporters/starvla_lerobot.py
++++ b/latency_bench/data/exporters/starvla_lerobot.py
+@@ -18,12 +18,17 @@ from latency_bench.utils.io import write_jsonl
+ from latency_bench.data.action_layouts import action_spec_for_env
+ from latency_bench.data.filters import density_weights_clean_v1
+ 
+-FPS_BY_ENV = {
+-    "flappy": 30,
+-    "demon_attack": 60,
+-    "deadly_corridor": 35,
+-}
+ STATE_DIM = 1
++BRIDGE_DIM = 7
++DEADLY_CORRIDOR_BRIDGE_ACTION_LABELS = (
++    "MOVE_FORWARD",
++    "MOVE_BACKWARD",
++    "MOVE_LEFT",
++    "MOVE_RIGHT",
++    "TURN_LEFT",
++    "TURN_RIGHT",
++    "ATTACK",
++)
+ 
+ 
+ def _png_bytes(image: Any) -> bytes:
+@@ -64,30 +69,38 @@ def _load_rows(raw_export_dir: Path, split: str) -> list[dict[str, Any]]:
+ 
+ 
+ def _episode_groups(rows: list[dict[str, Any]]) -> list[list[dict[str, Any]]]:
+-    """Group canonical rows by episode and order by decision step.
++    """Group canonical rows by trajectory and order by decision step.
+ 
+     Args:
+         rows: Canonical rollout rows from one split.
+ 
+     Returns:
+-        Episode row groups sorted by episode index.
++        Trajectory row groups sorted by episode and trajectory index.
+     """
+-    grouped: dict[int, list[dict[str, Any]]] = {}
++    grouped: dict[tuple[int, int], list[dict[str, Any]]] = {}
+     for row in rows:
+-        grouped.setdefault(int(row["episode_idx"]), []).append(row)
++        key = (int(row["episode_idx"]), int(row["trajectory_idx"]))
++        grouped.setdefault(key, []).append(row)
+     return [
+         sorted(episode_rows, key=lambda row: int(row["decision_step"]))
+         for _, episode_rows in sorted(grouped.items())
+     ]
+ 
+ 
+-def _write_episode(path: Path, rows: list[dict[str, Any]], *, action_dim: int) -> None:
++def _write_episode(
++    path: Path,
++    rows: list[dict[str, Any]],
++    *,
++    action_dim: int,
++    state_dim: int,
++) -> None:
+     """Write one LeRobot episode parquet file.
+ 
+     Args:
+         path: Destination episode parquet path.
+         rows: Converted LeRobot frame rows.
+         action_dim: Expected action vector dimension.
++        state_dim: Expected state carrier dimension.
+     """
+     path.parent.mkdir(parents=True, exist_ok=True)
+     table = pa.table(
+@@ -97,8 +110,8 @@ def _write_episode(path: Path, rows: list[dict[str, Any]], *, action_dim: int) -
+                 type=pa.struct([("bytes", pa.binary()), ("path", pa.string())]),
+             ),
+             "observation.state": pa.array(
+-                [[0.0] * STATE_DIM for _ in rows],
+-                type=pa.list_(pa.float32(), STATE_DIM),
++                [[0.0] * state_dim for _ in rows],
++                type=pa.list_(pa.float32(), state_dim),
+             ),
+             "action": pa.array(
+                 [row["action"] for row in rows],
+@@ -131,11 +144,47 @@ def _row_action_vector(row: dict[str, Any], *, action_dim: int) -> list[float]:
+     return [1.0 if idx == action_id else 0.0 for idx in range(action_dim)]
+ 
+ 
++def _bridge_action_vector(
++    row: dict[str, Any],
++    *,
++    env_name: str,
++    active_action_dim: int,
++) -> list[float]:
++    if env_name == "deadly_corridor":
++        action_text = str(row["action_text"]).upper()
++        return [
++            1.0 if label in action_text else 0.0
++            for label in DEADLY_CORRIDOR_BRIDGE_ACTION_LABELS
++        ]
++    action = _row_action_vector(row, action_dim=active_action_dim)
++    return [*action, *([0.0] * (BRIDGE_DIM - active_action_dim))]
++
++
++def _target_action_labels(
++    *,
++    env_name: str,
++    source_action_labels: list[str],
++    action_carrier: str,
++) -> list[str]:
++    if action_carrier == "native":
++        return source_action_labels
++    if env_name == "deadly_corridor":
++        return list(DEADLY_CORRIDOR_BRIDGE_ACTION_LABELS)
++    return [
++        *source_action_labels,
++        *[
++            f"BRIDGE_PAD_{idx}"
++            for idx in range(len(source_action_labels), BRIDGE_DIM)
++        ],
++    ]
++
++
+ def _write_metadata(
+     dataset_dir: Path,
+     *,
+-    fps: int,
++    fps: float,
+     action_labels: list[str],
++    state_dim: int,
+     episode_lengths: list[int],
+     task_prompts: list[str],
+ ) -> None:
+@@ -144,7 +193,8 @@ def _write_metadata(
+     Args:
+         dataset_dir: LeRobot dataset split directory.
+         fps: Frame rate stored in metadata.
+-        action_labels: Canonical action labels in vector order.
++        action_labels: Target action labels in vector order.
++        state_dim: State carrier dimension.
+         episode_lengths: Number of frames per converted episode.
+         task_prompts: Unique prompts used as LeRobot tasks.
+     """
+@@ -155,10 +205,10 @@ def _write_metadata(
+         json.dumps(
+             {
+                 "codebase_version": "v2.0",
+-                "fps": int(fps),
++                "fps": fps,
+                 "features": {
+                     "observation.image": {"dtype": "image"},
+-                    "observation.state": {"dtype": "float32", "shape": [STATE_DIM]},
++                    "observation.state": {"dtype": "float32", "shape": [state_dim]},
+                     "action": {"dtype": "float32", "shape": [action_dim], "names": action_labels},
+                     "timestamp": {"dtype": "float64", "shape": [1]},
+                     "episode_index": {"dtype": "int64", "shape": [1]},
+@@ -183,9 +233,11 @@ def _convert_split(
+     *,
+     env_name: str,
+     action_labels: list[str],
+-    fps: int,
++    fps: float,
+     reward_field: str,
+-    obs_stride: int = 1,
++    action_carrier: str,
++    active_action_dim: int,
++    state_dim: int,
+     density_weighting: bool = False,
+ ) -> dict[str, Any]:
+     """Convert one canonical rollout split into a LeRobot dataset directory.
+@@ -195,13 +247,11 @@ def _convert_split(
+         output_dir: Destination LeRobot split directory.
+         env_name: Canonical benchmark environment name.
+         action_labels: Canonical action labels in vector order.
+-        fps: Frame rate used for timestamps.
++        fps: Observation frame rate used for timestamps.
+         reward_field: Source canonical reward column.
+-        obs_stride: Raw frames per observation decision (``round(env_fps/obs_fps)``).
+-            Canonical rows are recorded once per raw frame; keeping only the obs
+-            decisions (``decision_step % obs_stride == 0``) makes adjacent LeRobot
+-            rows adjacent observations, so KV-memory rollouts unroll at the eval
+-            observation cadence.
++        action_carrier: Target action/state carrier.
++        active_action_dim: Number of active environment action dimensions.
++        state_dim: Target state carrier dimension.
+         density_weighting: When True, write per-row ``density_weight`` from clean_v1's
+             event balancing (train split). When False, weights are uniform 1.0 (val
+             split keeps an unweighted held-out loss).
+@@ -214,7 +264,6 @@ def _convert_split(
+     task_prompts: list[str] = []
+     episode_lengths: list[int] = []
+     for new_episode_idx, episode_rows in enumerate(_episode_groups(rows)):
+-        episode_rows = [row for row in episode_rows if int(row["decision_step"]) % obs_stride == 0]
+         density_weights = (
+             density_weights_clean_v1(env_name, episode_rows)
+             if density_weighting
+@@ -231,8 +280,14 @@ def _convert_split(
+             out_rows.append(
+                 {
+                     "image_bytes": _png_bytes(row["image"]),
+-                    "action": _row_action_vector(row, action_dim=len(action_labels)),
+-                    "timestamp": float(frame_idx) / int(fps),
++                    "action": _bridge_action_vector(
++                        row,
++                        env_name=env_name,
++                        active_action_dim=active_action_dim,
++                    )
++                    if action_carrier == "bridge"
++                    else _row_action_vector(row, action_dim=len(action_labels)),
++                    "timestamp": float(frame_idx) / fps,
+                     "episode_index": new_episode_idx,
+                     "frame_index": frame_idx,
+                     "task_index": prompt_to_task_index[prompt],
+@@ -248,11 +303,13 @@ def _convert_split(
+             output_dir / f"data/chunk-{episode_chunk:03d}/episode_{new_episode_idx:06d}.parquet",
+             out_rows,
+             action_dim=len(action_labels),
++            state_dim=state_dim,
+         )
+     _write_metadata(
+         output_dir,
+         fps=fps,
+         action_labels=action_labels,
++        state_dim=state_dim,
+         episode_lengths=episode_lengths,
+         task_prompts=task_prompts,
+     )
+@@ -269,6 +326,7 @@ def convert_canonical_rollout_to_lerobot(
+     raw_export_dir: Path | str,
+     output_dir: Path | str,
+     *,
++    action_carrier: str = "native",
+     force: bool = False,
+ ) -> dict[str, Any]:
+     """Convert a canonical rollout export into train/val LeRobot datasets.
+@@ -276,6 +334,7 @@ def convert_canonical_rollout_to_lerobot(
+     Args:
+         raw_export_dir: Directory with canonical rollout parquet files and metadata.
+         output_dir: Destination LeRobot train dataset directory.
++        action_carrier: Target action/state carrier, ``native`` or ``bridge``.
+         force: Remove existing train/val output directories before writing.
+ 
+     Returns:
+@@ -286,11 +345,20 @@ def convert_canonical_rollout_to_lerobot(
+     metadata = json.loads((raw_export_dir / "metadata.json").read_text(encoding="utf-8"))
+     env_name = str(metadata["env_name"])
+     action_spec = action_spec_for_env(env_name)
+-    action_labels = list(metadata["action_labels"])
++    source_action_labels = list(metadata["action_labels"])
++    action_labels = _target_action_labels(
++        env_name=env_name,
++        source_action_labels=source_action_labels,
++        action_carrier=action_carrier,
++    )
++    active_action_dim = {
++        "flappy": 2,
++        "demon_attack": 6,
++        "deadly_corridor": 7,
++    }[env_name]
++    state_dim = {"native": STATE_DIM, "bridge": BRIDGE_DIM}[action_carrier]
+     reward_field = str(metadata["reward_field"])
+-    obs_stride = int(metadata.get("obs_stride_raw_frames", 1))
+-    # Rows survive at the observation cadence, so timestamps advance at obs_fps.
+-    fps = max(1, int(FPS_BY_ENV[env_name]) // obs_stride)
++    fps = metadata["obs_fps"]
+     val_output_dir = output_dir.with_name(f"{output_dir.name}__val")
+     if output_dir.exists() and force:
+         shutil.rmtree(output_dir)
+@@ -303,7 +371,9 @@ def convert_canonical_rollout_to_lerobot(
+         action_labels=action_labels,
+         fps=fps,
+         reward_field=reward_field,
+-        obs_stride=obs_stride,
++        action_carrier=action_carrier,
++        active_action_dim=active_action_dim,
++        state_dim=state_dim,
+         density_weighting=True,
+     )
+     val_manifest = _convert_split(
+@@ -313,7 +383,9 @@ def convert_canonical_rollout_to_lerobot(
+         action_labels=action_labels,
+         fps=fps,
+         reward_field=reward_field,
+-        obs_stride=obs_stride,
++        action_carrier=action_carrier,
++        active_action_dim=active_action_dim,
++        state_dim=state_dim,
+         density_weighting=False,
+     )
+     manifest = {
+@@ -321,9 +393,13 @@ def convert_canonical_rollout_to_lerobot(
+         "format": "starvla_lerobot_v2_image_parquet",
+         "source": str(raw_export_dir),
+         "action_layout": str(metadata["action_layout"]),
++        "action_carrier": action_carrier,
+         "action_labels": action_labels,
+         "action_dim": len(action_labels),
+-        "active_action_dim": len(action_spec.labels),
++        "active_action_dim": active_action_dim
++        if action_carrier == "bridge"
++        else len(action_spec.labels),
++        "state_dim": state_dim,
+         "validation_dataset_name": val_output_dir.name,
+         "validation_episodes": val_manifest["episodes"],
+         "validation_frames": val_manifest["frames"],
+diff --git a/latency_bench/data/hf_parquet.py b/latency_bench/data/hf_parquet.py
+index 7a26426..6163c0d 100644
+--- a/latency_bench/data/hf_parquet.py
++++ b/latency_bench/data/hf_parquet.py
+@@ -205,7 +205,9 @@ def completed_episode_indices_from_shards(output_dir: Path) -> set[int]:
+             with shard_path.open("r", encoding="utf-8") as handle:
+                 for line in handle:
+                     if line.strip():
+-                        completed_episode_indices.add(int(json.loads(line)["episode_idx"]))
++                        row = json.loads(line)
++                        row["trajectory_idx"]
++                        completed_episode_indices.add(int(row["episode_idx"]))
+                         break
+     return completed_episode_indices
+ 
+diff --git a/latency_bench/data/rollout_schema.py b/latency_bench/data/rollout_schema.py
+index 11426b9..21d2bf5 100644
+--- a/latency_bench/data/rollout_schema.py
++++ b/latency_bench/data/rollout_schema.py
+@@ -7,11 +7,12 @@ from datasets import Features, Image as HFImage, Sequence as HFSequence, Value
+ from latency_bench.data.action_layouts import ActionSpec
+ 
+ 
+-SCHEMA_VERSION = 2
++SCHEMA_VERSION = 3
+ CANONICAL_ROW_FIELDS = (
+     "image",
+     "env_name",
+     "episode_idx",
++    "trajectory_idx",
+     "decision_step",
+     "split",
+     "seed",
+@@ -26,6 +27,7 @@ def _rollout_feature_map(action_spec: ActionSpec) -> dict[str, Any]:
+         "image": HFImage(),
+         "env_name": Value("string"),
+         "episode_idx": Value("int64"),
++        "trajectory_idx": Value("int64"),
+         "decision_step": Value("int64"),
+         "split": Value("string"),
+         "seed": Value("int64"),
+diff --git a/latency_bench/data/sf_teacher_rollout.py b/latency_bench/data/sf_teacher_rollout.py
+index af2b64f..bbbd157 100644
+--- a/latency_bench/data/sf_teacher_rollout.py
++++ b/latency_bench/data/sf_teacher_rollout.py
+@@ -355,6 +355,7 @@ def _write_filtered_episode_images(
+     image_writer: RolloutImageWriter,
+     rows: list[dict[str, Any]],
+     frames_by_decision_step: Mapping[int, np.ndarray],
++    trajectory_by_decision_step: Mapping[int, int],
+     *,
+     context_window: int,
+ ) -> list[dict[str, Any]]:
+@@ -364,6 +365,7 @@ def _write_filtered_episode_images(
+         image_writer: Image writer shared by the rollout run.
+         rows: Final rows selected for materialization.
+         frames_by_decision_step: Captured frames keyed by decision step.
++        trajectory_by_decision_step: Trajectory index keyed by decision step.
+         context_window: Total image window length including the current ``image`` field.
+ 
+     Returns:
+@@ -371,10 +373,15 @@ def _write_filtered_episode_images(
+     """
+     if not rows:
+         return []
+-    available_decision_steps = sorted(frames_by_decision_step)
+     required_steps: set[int] = set()
+     for row in rows:
+         decision_step = int(row["decision_step"])
++        trajectory_idx = int(row["trajectory_idx"])
++        available_decision_steps = sorted(
++            step
++            for step, step_trajectory_idx in trajectory_by_decision_step.items()
++            if step_trajectory_idx == trajectory_idx
++        )
+         required_steps.add(decision_step)
+         if int(context_window) > 1:
+             required_steps.update(
+@@ -396,6 +403,12 @@ def _write_filtered_episode_images(
+     for row in rows:
+         written = dict(row)
+         decision_step = int(written["decision_step"])
++        trajectory_idx = int(written["trajectory_idx"])
++        available_decision_steps = sorted(
++            step
++            for step, step_trajectory_idx in trajectory_by_decision_step.items()
++            if step_trajectory_idx == trajectory_idx
++        )
+         written["image"] = written_paths_by_step[decision_step]
+         if int(context_window) > 1:
+             written["context_images"] = [
+@@ -804,6 +817,7 @@ def _replay_accepted_specs(
+     active_episode_rows: list[list[dict[str, Any]]] = [[] for _ in range(driver.slot_count)]
+     active_episode_frames: list[list[np.ndarray]] = [[] for _ in range(driver.slot_count)]
+     active_decision_steps = [0 for _ in range(driver.slot_count)]
++    active_trajectory_indices = [0 for _ in range(driver.slot_count)]
+     active_ghost_streams: list[Any | None] = [None for _ in range(driver.slot_count)]
+     rnn_states = torch.zeros([driver.slot_count, rnn_size], dtype=torch.float32, device=device)
+     image_writer = RolloutImageWriter(
+@@ -844,6 +858,7 @@ def _replay_accepted_specs(
+                 active_episode_rows[slot] = []
+                 active_episode_frames[slot] = []
+                 active_decision_steps[slot] = 0
++                active_trajectory_indices[slot] = 0
+                 rnn_states[slot].zero_()
+ 
+         with tqdm(total=len(specs_to_replay), initial=0, unit="episode") as progress:
+@@ -881,21 +896,21 @@ def _replay_accepted_specs(
+                         next_decision_step=next_decision_step,
+                         max_steps_per_episode=int(max_steps_per_episode),
+                     )
+-                    if not life_loss_boundary:
+-                        active_episode_frames[slot].append(np.array(image_frame, copy=True))
+-                        row = teacher_dataset_row(
+-                            env_name=env_name,
+-                            spec=spec,
+-                            decision_step=decision_step,
+-                            image=str(image_path),
+-                            raw_action=raw_actions[batch_idx],
+-                            info=step_result.info,
+-                            reward=step_result.reward,
+-                            env_fps=env_fps,
+-                            obs_fps=obs_fps,
+-                            image_transform=image_transform_config.normalized_transform(),
+-                        )
+-                        active_episode_rows[slot].append(row)
++                    active_episode_frames[slot].append(np.array(image_frame, copy=True))
++                    row = teacher_dataset_row(
++                        env_name=env_name,
++                        spec=spec,
++                        trajectory_idx=active_trajectory_indices[slot],
++                        decision_step=decision_step,
++                        image=str(image_path),
++                        raw_action=raw_actions[batch_idx],
++                        info=step_result.info,
++                        reward=step_result.reward,
++                        env_fps=env_fps,
++                        obs_fps=obs_fps,
++                        image_transform=image_transform_config.normalized_transform(),
++                    )
++                    active_episode_rows[slot].append(row)
+                     rnn_states[slot] = new_rnn_states[batch_idx]
+                     if episode_done:
+                         final_rows = _finalize_attempt_rows(
+@@ -931,6 +946,10 @@ def _replay_accepted_specs(
+                             image_writer,
+                             shard_rows,
+                             frames_by_decision_step,
++                            {
++                                int(row["decision_step"]): int(row["trajectory_idx"])
++                                for row in final_rows
++                            },
+                             context_window=int(context_window),
+                         )
+                         write_episode_row_shard(output_dir, shard_rows)
+@@ -943,12 +962,14 @@ def _replay_accepted_specs(
+                         active_episode_rows[slot] = []
+                         active_episode_frames[slot] = []
+                         active_decision_steps[slot] = 0
++                        active_trajectory_indices[slot] = 0
+                         if _is_demon_ghost_trail:
+                             active_ghost_streams[slot] = None
+                         rnn_states[slot].zero_()
+                     else:
+                         if life_loss_boundary:
+                             active_obs[slot], reset_info = driver.continue_after_life_loss(slot)
++                            active_trajectory_indices[slot] += 1
+                             if _is_demon_ghost_trail:
+                                 active_ghost_streams[slot].clear()
+                                 active_image_frames[slot] = frame_from_env_raw_rgb_info(reset_info).copy()
+@@ -981,6 +1002,7 @@ def _replay_accepted_specs(
+                         active_episode_rows[slot] = []
+                         active_episode_frames[slot] = []
+                         active_decision_steps[slot] = 0
++                        active_trajectory_indices[slot] = 0
+                         rnn_states[slot].zero_()
+                 else:
+                     if isinstance(driver, BatchedTeacherRolloutDriver):
+diff --git a/latency_bench/data/sf_teacher_rows.py b/latency_bench/data/sf_teacher_rows.py
+index efa6f3e..4e7ce9a 100644
+--- a/latency_bench/data/sf_teacher_rows.py
++++ b/latency_bench/data/sf_teacher_rows.py
+@@ -14,6 +14,7 @@ def teacher_dataset_row(
+     *,
+     env_name: str,
+     spec: Mapping[str, Any],
++    trajectory_idx: int,
+     decision_step: int,
+     image: str,
+     raw_action: Any,
+@@ -28,6 +29,7 @@ def teacher_dataset_row(
+     Args:
+         env_name: Canonical benchmark environment name.
+         spec: Episode spec containing index, split, and seed.
++        trajectory_idx: Zero-based trajectory index within the full-game episode.
+         decision_step: Zero-based policy decision step within the episode.
+         image: Path to the saved pre-step image.
+         raw_action: Raw teacher action for this decision.
+@@ -48,6 +50,7 @@ def teacher_dataset_row(
+         "image": image,
+         "env_name": str(env_name),
+         "episode_idx": int(spec["episode_idx"]),
++        "trajectory_idx": int(trajectory_idx),
+         "decision_step": int(decision_step),
+         "split": str(spec["split"]),
+         "seed": int(spec["seed"]),
+diff --git a/scripts/experiment_runner/run_manifest_jobs.py b/scripts/experiment_runner/run_manifest_jobs.py
+index e2d1b5e..f425a27 100644
+--- a/scripts/experiment_runner/run_manifest_jobs.py
++++ b/scripts/experiment_runner/run_manifest_jobs.py
+@@ -73,11 +73,6 @@ def _event_writer(event_log: Path) -> Callable[[dict[str, Any]], None]:
+     return write
+ 
+ 
+-def _register_process(process: subprocess.Popen) -> None:
+-    with _ACTIVE_PROCESSES_LOCK:
+-        _ACTIVE_PROCESSES.add(process)
+-
+-
+ def _forget_process(process: subprocess.Popen) -> None:
+     with _ACTIVE_PROCESSES_LOCK:
+         _ACTIVE_PROCESSES.discard(process)
+@@ -89,11 +84,10 @@ def _active_processes() -> list[subprocess.Popen]:
+ 
+ 
+ def _signal_process_group(process: subprocess.Popen, sig: int) -> None:
+-    if process.poll() is None:
+-        try:
+-            os.killpg(process.pid, sig)
+-        except ProcessLookupError:
+-            pass
++    try:
++        os.killpg(process.pid, sig)
++    except ProcessLookupError:
++        pass
+ 
+ 
+ def _terminate_active_processes(sig: int = signal.SIGTERM) -> None:
+@@ -118,15 +112,24 @@ def _terminate_active_processes(sig: int = signal.SIGTERM) -> None:
+             pass
+ 
+ 
+-def _run_command(command: str, *, gpu: str) -> None:
++def _run_command(
++    command: str,
++    *,
++    gpu: str,
++    cancel_event: threading.Event | None = None,
++) -> bool:
+     env = os.environ.copy()
+     env["CUDA_VISIBLE_DEVICES"] = str(gpu)
+-    process = subprocess.Popen(command, shell=True, env=env, start_new_session=True)
+-    _register_process(process)
++    with _ACTIVE_PROCESSES_LOCK:
++        if cancel_event is not None and cancel_event.is_set():
++            return False
++        process = subprocess.Popen(command, shell=True, env=env, start_new_session=True)
++        _ACTIVE_PROCESSES.add(process)
+     try:
+         returncode = process.wait()
+         if returncode != 0:
+             raise subprocess.CalledProcessError(returncode, command)
++        return True
+     finally:
+         _forget_process(process)
+ 
+@@ -223,16 +226,6 @@ def run_jobs(
+     dry_run: bool = False,
+     uploader: Callable[[dict[str, str]], None] | None = None,
+ ) -> int:
+-    previous_sigint = signal.getsignal(signal.SIGINT)
+-    previous_sigterm = signal.getsignal(signal.SIGTERM)
+-
+-    def stop_active_processes(signum, frame):
+-        del frame
+-        _terminate_active_processes(signal.SIGTERM)
+-        raise KeyboardInterrupt(signum)
+-
+-    signal.signal(signal.SIGINT, stop_active_processes)
+-    signal.signal(signal.SIGTERM, stop_active_processes)
+     manifest_file = Path(manifest_path).expanduser().resolve()
+     event_log = event_log_path_for_manifest(manifest_file)
+     write_event = _event_writer(event_log)
+@@ -259,6 +252,9 @@ def run_jobs(
+     upload_queue: queue.Queue[UploadTask | None] = queue.Queue()
+     failures: list[tuple[str, str]] = []
+     failure_lock = threading.Lock()
++    cancel_event = threading.Event()
++    signal_event = threading.Event()
++    cleanup_started = False
+     uploader_fn = _default_uploader if uploader is None else uploader
+ 
+     if stage != "upload":
+@@ -280,28 +276,49 @@ def run_jobs(
+             if job is None:
+                 job_queue.task_done()
+                 break
++            if cancel_event.is_set():
++                job_queue.task_done()
++                continue
+             next_stage = _resume_stage(job, states) if stage == "all" else stage
+             job_id = str(job["job_id"])
+             try:
+                 if next_stage == "train":
++                    if cancel_event.is_set():
++                        continue
+                     if dry_run:
+                         print(f"[dry-run] CUDA_VISIBLE_DEVICES={gpu} {job['train_command']}", flush=True)
+                     else:
+                         write_event({"job_id": job_id, "stage": "train_started", "gpu": gpu, "ts": _now()})
+-                        _run_command(str(job["train_command"]), gpu=gpu)
++                        started = _run_command(
++                            str(job["train_command"]),
++                            gpu=gpu,
++                            cancel_event=cancel_event,
++                        )
++                        if not started or cancel_event.is_set():
++                            continue
+                         write_event({"job_id": job_id, "stage": "train_succeeded", "gpu": gpu, "ts": _now()})
+                     next_stage = "eval" if stage == "all" else "complete"
+                 if next_stage == "eval":
++                    if cancel_event.is_set():
++                        continue
+                     if dry_run:
+                         print(f"[dry-run] CUDA_VISIBLE_DEVICES={gpu} {job['eval_command']}", flush=True)
+                     else:
+                         write_event({"job_id": job_id, "stage": "eval_started", "gpu": gpu, "ts": _now()})
+-                        _run_command(str(job["eval_command"]), gpu=gpu)
++                        started = _run_command(
++                            str(job["eval_command"]),
++                            gpu=gpu,
++                            cancel_event=cancel_event,
++                        )
++                        if not started or cancel_event.is_set():
++                            continue
+                         write_event({"job_id": job_id, "stage": "eval_succeeded", "gpu": gpu, "ts": _now()})
+                     next_stage = "upload" if stage == "all" else "complete"
+-                if next_stage == "upload":
++                if next_stage == "upload" and not cancel_event.is_set():
+                     upload_queue.put(UploadTask(job))
+             except subprocess.CalledProcessError:
++                if cancel_event.is_set():
++                    continue
+                 failed_stage = "train_failed" if next_stage == "train" else "eval_failed"
+                 write_event({"job_id": job_id, "stage": failed_stage, "gpu": gpu, "ts": _now()})
+                 record_failure(job_id, failed_stage)
+@@ -311,48 +328,52 @@ def run_jobs(
+     def upload_worker() -> None:
+         while True:
+             task = upload_queue.get()
+-            if task is None:
+-                upload_queue.task_done()
+-                break
+-            job = task.job
+-            job_id = str(job["job_id"])
+-            train_repo_id = str(job["train_upload_repo_id"])
+-            eval_repo_id = str(job["eval_upload_repo_id"])
+-            if dry_run:
+-                targets = [repo_id for repo_id in (eval_repo_id, train_repo_id) if repo_id]
+-                print(f"[dry-run] upload {job_id}: {', '.join(targets) or 'no upload target'}", flush=True)
+-                upload_queue.task_done()
+-                continue
+-            if not train_repo_id and not eval_repo_id:
+-                write_event({"job_id": job_id, "stage": "upload_skipped", "ts": _now()})
+-                write_event({"job_id": job_id, "stage": "job_complete", "ts": _now()})
+-                upload_queue.task_done()
+-                continue
+-            attempt = 0
+-            success = False
+-            while attempt < max_upload_attempts and not success:
+-                attempt += 1
+-                write_event({"job_id": job_id, "stage": "upload_started", "attempt": attempt, "ts": _now()})
+-                try:
+-                    if not dry_run:
+-                        uploader_fn(job)
+-                    write_event({"job_id": job_id, "stage": "upload_succeeded", "attempt": attempt, "ts": _now()})
++            try:
++                if task is None:
++                    break
++                if cancel_event.is_set():
++                    continue
++                job = task.job
++                job_id = str(job["job_id"])
++                train_repo_id = str(job["train_upload_repo_id"])
++                eval_repo_id = str(job["eval_upload_repo_id"])
++                if dry_run:
++                    targets = [repo_id for repo_id in (eval_repo_id, train_repo_id) if repo_id]
++                    print(f"[dry-run] upload {job_id}: {', '.join(targets) or 'no upload target'}", flush=True)
++                    continue
++                if not train_repo_id and not eval_repo_id:
++                    write_event({"job_id": job_id, "stage": "upload_skipped", "ts": _now()})
+                     write_event({"job_id": job_id, "stage": "job_complete", "ts": _now()})
+-                    success = True
+-                except Exception as error:
+-                    write_event(
+-                        {
+-                            "job_id": job_id,
+-                            "stage": "upload_failed",
+-                            "attempt": attempt,
+-                            "error": repr(error),
+-                            "ts": _now(),
+-                        }
+-                    )
+-                    print(f"upload failed for {job_id} (attempt {attempt}): {error!r}", file=sys.stderr)
+-            if not success:
+-                record_failure(job_id, "upload_failed")
+-            upload_queue.task_done()
++                    continue
++                attempt = 0
++                success = False
++                while attempt < max_upload_attempts and not success and not cancel_event.is_set():
++                    attempt += 1
++                    write_event({"job_id": job_id, "stage": "upload_started", "attempt": attempt, "ts": _now()})
++                    try:
++                        uploader_fn(job)
++                        if cancel_event.is_set():
++                            continue
++                        write_event({"job_id": job_id, "stage": "upload_succeeded", "attempt": attempt, "ts": _now()})
++                        write_event({"job_id": job_id, "stage": "job_complete", "ts": _now()})
++                        success = True
++                    except Exception as error:
++                        if cancel_event.is_set():
++                            continue
++                        write_event(
++                            {
++                                "job_id": job_id,
++                                "stage": "upload_failed",
++                                "attempt": attempt,
++                                "error": repr(error),
++                                "ts": _now(),
++                            }
++                        )
++                        print(f"upload failed for {job_id} (attempt {attempt}): {error!r}", file=sys.stderr)
++                if not success and not cancel_event.is_set():
++                    record_failure(job_id, "upload_failed")
++            finally:
++                upload_queue.task_done()
+ 
+     upload_threads = [
+         threading.Thread(target=upload_worker, name=f"upload-{idx}", daemon=True)
+@@ -362,17 +383,44 @@ def run_jobs(
+         threading.Thread(target=train_worker, args=(gpu,), name=f"train-{gpu}", daemon=True)
+         for gpu in gpus if stage != "upload"
+     ]
++    previous_sigint = signal.getsignal(signal.SIGINT)
++    previous_sigterm = signal.getsignal(signal.SIGTERM)
++
++    def stop_active_processes(signum, frame):
++        nonlocal cleanup_started
++        del frame
++        if signal_event.is_set():
++            return
++        signal_event.set()
++        cancel_event.set()
++        if cleanup_started:
++            return
++        raise KeyboardInterrupt(signum)
++
++    signal.signal(signal.SIGINT, stop_active_processes)
++    signal.signal(signal.SIGTERM, stop_active_processes)
+     try:
+         for thread in upload_threads + train_threads:
+             thread.start()
+         job_queue.join()
+-        for _ in upload_threads:
+-            upload_queue.put(None)
+         upload_queue.join()
+-        for thread in train_threads + upload_threads:
+-            thread.join()
+         return 1 if failures else 0
+     finally:
++        cleanup_started = True
++        cancel_event.set()
++        _terminate_active_processes(signal.SIGTERM)
++        for thread in train_threads:
++            thread.join()
++        _terminate_active_processes(signal.SIGTERM)
++        if signal_event.is_set():
++            for _ in upload_threads:
++                upload_queue.put(None)
++        else:
++            upload_queue.join()
++            for _ in upload_threads:
++                upload_queue.put(None)
++            for thread in upload_threads:
++                thread.join()
+         _terminate_active_processes(signal.SIGTERM)
+         signal.signal(signal.SIGINT, previous_sigint)
+         signal.signal(signal.SIGTERM, previous_sigterm)
+diff --git a/scripts/rollout_data/prepare_starvla_lerobot_dataset.py b/scripts/rollout_data/prepare_starvla_lerobot_dataset.py
+index e202657..2521899 100644
+--- a/scripts/rollout_data/prepare_starvla_lerobot_dataset.py
++++ b/scripts/rollout_data/prepare_starvla_lerobot_dataset.py
+@@ -26,6 +26,7 @@ def convert_raw_export_to_lerobot(
+     raw_export_dir: Path,
+     lerobot_root: Path,
+     data_mix: str,
++    action_carrier: str = "native",
+ ) -> dict:
+     """Convert a raw rollout export into a named LeRobot data mix.
+ 
+@@ -33,6 +34,7 @@ def convert_raw_export_to_lerobot(
+         raw_export_dir: Canonical rollout export directory.
+         lerobot_root: Directory that will contain the LeRobot datasets.
+         data_mix: Train dataset directory name under ``lerobot_root``.
++        action_carrier: Target action/state carrier.
+ 
+     Returns:
+         Conversion manifest returned by the LeRobot adapter.
+@@ -41,6 +43,7 @@ def convert_raw_export_to_lerobot(
+     return convert_canonical_rollout_to_lerobot(
+         raw_export_dir,
+         output_dir,
++        action_carrier=action_carrier,
+         force=True,
+     )
+ 
+@@ -92,6 +95,7 @@ def build_parser() -> argparse.ArgumentParser:
+     parser.add_argument("raw_export_dir")
+     parser.add_argument("lerobot_root")
+     parser.add_argument("data_mix")
++    parser.add_argument("--action-carrier", choices=("native", "bridge"), default="native")
+     parser.add_argument("--push-to-hub", default="")
+     parser.add_argument("--private", action="store_true")
+     return parser
+@@ -115,6 +119,7 @@ def main(argv: Sequence[str] | None = None) -> int:
+         raw_export_dir=raw_export_dir,
+         lerobot_root=lerobot_root,
+         data_mix=data_mix,
++        action_carrier=args.action_carrier,
+     )
+     if args.push_to_hub:
+         push_lerobot_dataset(
+diff --git a/tests/data/test_rollout_dataset_schema.py b/tests/data/test_rollout_dataset_schema.py
+index 06a6091..55ccebf 100644
+--- a/tests/data/test_rollout_dataset_schema.py
++++ b/tests/data/test_rollout_dataset_schema.py
+@@ -8,6 +8,7 @@ def test_canonical_rollout_fields_exclude_legacy_aliases():
+         "image",
+         "env_name",
+         "episode_idx",
++        "trajectory_idx",
+         "decision_step",
+         "split",
+         "seed",
+@@ -36,6 +37,7 @@ def test_rollout_features_include_env_specific_action_fields():
+     assert "action_tuple" not in deadly_features
+     assert "action" not in deadly_features
+     for features in (flappy_features, deadly_features):
++        assert "trajectory_idx" in features
+         assert "decision_step" in features
+         assert "latency_raw_frames" in features
+         assert "raw_reward" in features
+diff --git a/tests/data/test_rollout_hf_parquet.py b/tests/data/test_rollout_hf_parquet.py
+index 9fa4c48..4e56de6 100644
+--- a/tests/data/test_rollout_hf_parquet.py
++++ b/tests/data/test_rollout_hf_parquet.py
+@@ -15,6 +15,7 @@ def test_hf_parquet_writes_canonical_columns_only(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -56,6 +57,7 @@ def test_hf_parquet_shards_sort_by_decision_step(tmp_path):
+             "image": str(image_path),
+             "env_name": "flappy",
+             "episode_idx": 0,
++            "trajectory_idx": 0,
+             "decision_step": decision_step,
+             "split": "train",
+             "seed": 11,
+@@ -92,6 +94,7 @@ def test_hf_parquet_embeds_context_images_when_rows_provide_them(tmp_path):
+         "context_images": [str(path) for path in context_paths],
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -116,3 +119,18 @@ def test_hf_parquet_embeds_context_images_when_rows_provide_them(tmp_path):
+     assert "context_images" in table.column_names
+     assert len(exported["context_images"]) == 3
+     assert exported["context_images"][0]["bytes"].startswith(b"\x89PNG\r\n\x1a\n")
++
++
++def test_completed_episode_shards_require_trajectory_identity(tmp_path):
++    import json
++
++    import pytest
++
++    from latency_bench.data.hf_parquet import completed_episode_indices_from_shards
++
++    shard_path = tmp_path / "episode_rows" / "train" / "episode_000000.jsonl"
++    shard_path.parent.mkdir(parents=True)
++    shard_path.write_text(json.dumps({"episode_idx": 0}) + "\n", encoding="utf-8")
++
++    with pytest.raises(KeyError, match="trajectory_idx"):
++        completed_episode_indices_from_shards(tmp_path)
+diff --git a/tests/data/test_sf_teacher_dataset_export.py b/tests/data/test_sf_teacher_dataset_export.py
+index 62b30a6..ba4644b 100644
+--- a/tests/data/test_sf_teacher_dataset_export.py
++++ b/tests/data/test_sf_teacher_dataset_export.py
+@@ -1299,7 +1299,7 @@ def test_write_filtered_episode_images_preserves_empty_filtered_episode(tmp_path
+ 
+     image_writer = RolloutImageWriter(max_workers=0, row_buffer_size=1, export_image_size=None)
+ 
+-    rows = _write_filtered_episode_images(image_writer, [], {}, context_window=5)
++    rows = _write_filtered_episode_images(image_writer, [], {}, {}, context_window=5)
+ 
+     assert rows == []
+ 
+@@ -1314,6 +1314,7 @@ def test_write_filtered_episode_images_resets_context_after_decision_step_gap(tm
+         {
+             "image": str(tmp_path / "images" / "train" / "episode_000000" / f"decision_step_{step:06d}.png"),
+             "decision_step": step,
++            "trajectory_idx": 0,
+         }
+         for step in [0, 1, 4, 5]
+     ]
+@@ -1326,6 +1327,7 @@ def test_write_filtered_episode_images_resets_context_after_decision_step_gap(tm
+         image_writer,
+         rows,
+         frames_by_decision_step,
++        {step: 0 for step in frames_by_decision_step},
+         context_window=4,
+     )
+ 
+@@ -1733,13 +1735,24 @@ def test_collect_teacher_dataset_demon_attack_exports_full_game_across_life_loss
+         image_writer_workers=0,
+         row_buffer_size=1,
+         filter_preset="none",
++        context_window=3,
+     )
+     train_rows = _read_shard_rows(output_dir, "train")
+     state = json.loads((output_dir / "probe_state.json").read_text(encoding="utf-8"))
+ 
+     assert ScriptedDemonEnv.reset_seeds == [100, None, None, 100, None, None]
+-    assert [row["decision_step"] for row in train_rows] == [2]
+-    assert train_rows[0]["raw_reward"] == 30.0
++    assert [row["decision_step"] for row in train_rows] == [0, 1, 2]
++    assert [row["trajectory_idx"] for row in train_rows] == [0, 1, 2]
++    assert [row["action_id"] for row in train_rows] == [1, 1, 1]
++    assert [row["raw_reward"] for row in train_rows] == [10.0, 20.0, 30.0]
++    from PIL import Image
++
++    context_values = [
++        [np.asarray(Image.open(path))[0, 0, 0].item() for path in row["context_images"]]
++        for row in train_rows
++    ]
++    current_values = [np.asarray(Image.open(row["image"]))[0, 0, 0].item() for row in train_rows]
++    assert context_values == [[value, value] for value in current_values]
+     assert state["accepted_specs"] == [
+         {
+             "attempt_idx": 0,
+@@ -1900,6 +1913,7 @@ def test_collect_teacher_dataset_skips_runtime_when_probe_and_shards_complete(mo
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 100,
+@@ -2264,6 +2278,7 @@ def test_write_parquet_splits_handles_empty_validation_rows(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -2300,6 +2315,7 @@ def test_write_parquet_splits_embeds_image_bytes(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -2336,6 +2352,7 @@ def test_write_parquet_splits_preserves_rows_across_large_exports(tmp_path):
+             "image": str(image_path),
+             "env_name": "flappy",
+             "episode_idx": episode_idx,
++            "trajectory_idx": 0,
+             "decision_step": 0,
+             "split": "train",
+             "seed": 11 + episode_idx,
+@@ -2390,6 +2407,7 @@ def  test_write_parquet_splits_from_shards_writes_train_and_val_rows(tmp_path):
+                 "image": str(image_path),
+                 "env_name": "flappy",
+                 "episode_idx": episode_idx,
++                "trajectory_idx": 0,
+                 "decision_step": decision_step,
+                 "split": split,
+                 "seed": 11 + episode_idx,
+@@ -2432,6 +2450,7 @@ def test_write_parquet_splits_from_shards_replaces_stale_partial_file(tmp_path):
+         "image": str(stale_image_path),
+         "env_name": "flappy",
+         "episode_idx": 99,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 99,
+@@ -2452,6 +2471,7 @@ def test_write_parquet_splits_from_shards_replaces_stale_partial_file(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -2536,6 +2556,8 @@ def test_prepare_starvla_lerobot_dataset_converts_from_raw_export(monkeypatch, t
+         str(raw_export),
+         str(tmp_path / "lerobot"),
+         "demon_attack_train",
++        "--action-carrier",
++        "bridge",
+     ])
+ 
+     assert result == 0
+@@ -2543,7 +2565,7 @@ def test_prepare_starvla_lerobot_dataset_converts_from_raw_export(monkeypatch, t
+         (
+             raw_export,
+             tmp_path / "lerobot" / "demon_attack_train",
+-            {"force": True},
++            {"action_carrier": "bridge", "force": True},
+         )
+     ]
+ 
+diff --git a/tests/data/test_starvla_lerobot_exporter.py b/tests/data/test_starvla_lerobot_exporter.py
+index 64b5f47..58391c7 100644
+--- a/tests/data/test_starvla_lerobot_exporter.py
++++ b/tests/data/test_starvla_lerobot_exporter.py
+@@ -3,6 +3,7 @@ from __future__ import annotations
+ import json
+ 
+ import pyarrow.parquet as pq
++import pytest
+ 
+ from helpers import save_test_image
+ 
+@@ -20,6 +21,7 @@ def test_starvla_lerobot_exporter_reads_canonical_rollout_schema(tmp_path):
+         "image": str(image_path),
+         "env_name": "deadly_corridor",
+         "episode_idx": 3,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 14,
+@@ -38,6 +40,7 @@ def test_starvla_lerobot_exporter_reads_canonical_rollout_schema(tmp_path):
+                 "action_layout": "deadly_corridor_joint_54",
+                 "reward_field": "raw_reward",
+                 "action_labels": list(action_spec.labels),
++                "obs_fps": 8.75,
+             }
+         ),
+         encoding="utf-8",
+@@ -54,10 +57,102 @@ def test_starvla_lerobot_exporter_reads_canonical_rollout_schema(tmp_path):
+     assert table.to_pylist()[0]["done"] is True
+ 
+ 
+-def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+-    """Adjacent LeRobot rows must be adjacent obs decisions (every obs_stride raw
+-    frames), with contiguous frame_index, so KV-memory rollouts unroll at the eval
+-    observation cadence rather than at every raw frame."""
++@pytest.mark.parametrize(
++    ("env_name", "obs_fps", "latency", "action_values", "expected_action", "active_action_dim"),
++    [
++        ("flappy", 30, 3, (1, [0.0, 1.0], "FLAP"), [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0], 2),
++        (
++            "demon_attack",
++            15,
++            6,
++            (5, [0.0, 0.0, 0.0, 0.0, 0.0, 1.0], "LEFTFIRE"),
++            [0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0],
++            6,
++        ),
++        (
++            "deadly_corridor",
++            8.75,
++            6,
++            (11, None, "MOVE_FORWARD + MOVE_RIGHT + ATTACK"),
++            [1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 1.0],
++            7,
++        ),
++    ],
++)
++def test_starvla_lerobot_exporter_writes_bridge_carrier_with_raw_latency_prompt(
++    tmp_path,
++    env_name,
++    obs_fps,
++    latency,
++    action_values,
++    expected_action,
++    active_action_dim,
++):
++    from latency_bench.data.action_layouts import action_spec_for_env
++    from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
++    from latency_bench.data.hf_parquet import write_parquet_splits
++
++    image_path = tmp_path / "image.png"
++    save_test_image(image_path, 9)
++    raw_export = tmp_path / "raw"
++    action_spec = action_spec_for_env(env_name)
++    action_id, action, action_text = action_values
++    row = {
++        "image": str(image_path),
++        "env_name": env_name,
++        "episode_idx": 0,
++        "trajectory_idx": 0,
++        "decision_step": 0,
++        "split": "train",
++        "seed": 0,
++        "prompt": f"The action latency is fixed at {latency} raw environment frames.",
++        "latency_raw_frames": latency,
++        "latency_ms": latency * 1000 / {"flappy": 30, "demon_attack": 60, "deadly_corridor": 35}[env_name],
++        "raw_reward": 1.0,
++        "action_id": action_id,
++        "action_text": action_text,
++    }
++    if action is not None:
++        row["action"] = action
++    write_parquet_splits(raw_export, [row], [], action_spec=action_spec)
++    (raw_export / "metadata.json").write_text(
++        json.dumps(
++            {
++                "env_name": env_name,
++                "action_layout": action_spec.layout,
++                "reward_field": action_spec.reward_field,
++                "action_labels": list(action_spec.labels),
++                "obs_fps": obs_fps,
++            }
++        ),
++        encoding="utf-8",
++    )
++
++    manifest = convert_canonical_rollout_to_lerobot(
++        raw_export,
++        tmp_path / "lerobot",
++        action_carrier="bridge",
++        force=True,
++    )
++
++    table = pq.read_table(tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000000.parquet")
++    info = json.loads((tmp_path / "lerobot" / "meta" / "info.json").read_text(encoding="utf-8"))
++    task_lines = (tmp_path / "lerobot" / "meta" / "tasks.jsonl").read_text(
++        encoding="utf-8"
++    ).splitlines()
++    tasks = [json.loads(line) for line in task_lines]
++    converted_row = table.to_pylist()[0]
++    assert converted_row["action"] == expected_action
++    assert converted_row["observation.state"] == [0.0] * 7
++    assert info["features"]["action"]["shape"] == [7]
++    assert info["features"]["observation.state"]["shape"] == [7]
++    assert tasks == [{"task_index": 0, "task": row["prompt"]}]
++    assert manifest["action_carrier"] == "bridge"
++    assert manifest["active_action_dim"] == active_action_dim
++
++
++def test_starvla_lerobot_exporter_keeps_every_observation_decision(tmp_path):
++    """Canonical rows already occur at the policy observation cadence."""
+     from latency_bench.data.action_layouts import action_spec_for_env
+     from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
+     from latency_bench.data.hf_parquet import write_parquet_splits
+@@ -66,12 +161,13 @@ def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+     save_test_image(image_path, 9)
+     raw_export = tmp_path / "raw"
+     action_spec = action_spec_for_env("deadly_corridor")
+-    # One episode, 8 contiguous per-raw-frame decision steps; obs_stride = 4.
++    # One trajectory with eight contiguous policy-decision observations.
+     train_rows = [
+         {
+             "image": str(image_path),
+             "env_name": "deadly_corridor",
+             "episode_idx": 0,
++            "trajectory_idx": 0,
+             "decision_step": step,
+             "split": "train",
+             "seed": 7,
+@@ -92,6 +188,7 @@ def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+                 "action_layout": "deadly_corridor_joint_54",
+                 "reward_field": "raw_reward",
+                 "action_labels": list(action_spec.labels),
++                "obs_fps": 8.75,
+                 "obs_stride_raw_frames": 4,
+             }
+         ),
+@@ -102,12 +199,73 @@ def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+ 
+     table = pq.read_table(tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000000.parquet")
+     rows = table.to_pylist()
+-    # Only decision_steps 0 and 4 survive; frame_index is contiguous obs index.
+-    assert [r["frame_index"] for r in rows] == [0, 1]
+-    assert [r["episode_index"] for r in rows] == [0, 0]
++    assert [r["frame_index"] for r in rows] == list(range(8))
++    assert [r["episode_index"] for r in rows] == [0] * 8
+     assert rows[-1]["done"] is True
+ 
+ 
++def test_starvla_lerobot_exporter_splits_life_trajectories(tmp_path):
++    from latency_bench.data.action_layouts import action_spec_for_env
++    from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
++    from latency_bench.data.hf_parquet import write_parquet_splits
++
++    image_path = tmp_path / "image.png"
++    save_test_image(image_path, 9)
++    raw_export = tmp_path / "raw"
++    action_spec = action_spec_for_env("demon_attack")
++    train_rows = [
++        {
++            "image": str(image_path),
++            "env_name": "demon_attack",
++            "episode_idx": 0,
++            "trajectory_idx": trajectory_idx,
++            "decision_step": decision_step,
++            "split": "train",
++            "seed": 7,
++            "prompt": "pick action",
++            "latency_raw_frames": 6,
++            "latency_ms": 100.0,
++            "raw_reward": 1.0,
++            "action_id": 0,
++            "action": [1.0, 0.0, 0.0, 0.0, 0.0, 0.0],
++            "action_text": "NOOP",
++        }
++        for trajectory_idx, decision_step in ((0, 0), (0, 1), (1, 2), (1, 3))
++    ]
++    write_parquet_splits(raw_export, train_rows, [], action_spec=action_spec)
++    (raw_export / "metadata.json").write_text(
++        json.dumps(
++            {
++                "env_name": "demon_attack",
++                "action_layout": action_spec.layout,
++                "reward_field": action_spec.reward_field,
++                "action_labels": list(action_spec.labels),
++                "obs_fps": 15,
++            }
++        ),
++        encoding="utf-8",
++    )
++
++    manifest = convert_canonical_rollout_to_lerobot(
++        raw_export,
++        tmp_path / "lerobot",
++        action_carrier="bridge",
++        force=True,
++    )
++
++    first = pq.read_table(
++        tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000000.parquet"
++    ).to_pylist()
++    second = pq.read_table(
++        tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000001.parquet"
++    ).to_pylist()
++    assert manifest["episodes"] == 2
++    assert [row["frame_index"] for row in first] == [0, 1]
++    assert [row["frame_index"] for row in second] == [0, 1]
++    assert [row["done"] for row in first] == [False, True]
++    assert [row["done"] for row in second] == [False, True]
++
++
+ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+     from datasets import Dataset, Features, Image as HFImage, Sequence as HFSequence, Value
+     from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
+@@ -121,6 +279,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+             "image": HFImage(),
+             "env_name": Value("string"),
+             "episode_idx": Value("int64"),
++            "trajectory_idx": Value("int64"),
+             "decision_step": Value("int64"),
+             "split": Value("string"),
+             "seed": Value("int64"),
+@@ -138,6 +297,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+             "image": [str(image_path)],
+             "env_name": ["flappy"],
+             "episode_idx": [0],
++            "trajectory_idx": [0],
+             "decision_step": [0],
+             "split": ["train"],
+             "seed": [11],
+@@ -156,6 +316,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+             "image": [str(image_path)],
+             "env_name": ["flappy"],
+             "episode_idx": [1],
++            "trajectory_idx": [0],
+             "decision_step": [0],
+             "split": ["val"],
+             "seed": [12],
+@@ -176,6 +337,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+                 "action_layout": "flappy_discrete",
+                 "reward_field": "raw_reward",
+                 "action_labels": ["NOOP", "FLAP"],
++                "obs_fps": 30,
+             }
+         ),
+         encoding="utf-8",
+diff --git a/tests/scripts/test_generate_sweep_configs.py b/tests/scripts/test_generate_sweep_configs.py
+index 23d79cb..1fe6cef 100644
+--- a/tests/scripts/test_generate_sweep_configs.py
++++ b/tests/scripts/test_generate_sweep_configs.py
+@@ -23,6 +23,73 @@ def _read_manifest(path: Path) -> list[dict[str, str]]:
+     return [dict(zip(header, line.split("\t"))) for line in lines[1:]]
+ 
+ 
++@pytest.mark.parametrize(
++    ("env_name", "latency_label", "latency_frames", "obs_fps", "obs_fps_label"),
++    [
++        ("flappy", "fixed_l3", 3, 30, "30"),
++        ("demon_attack", "fixed_l6", 6, 15, "15"),
++        ("deadly_corridor", "fixed_l6", 6, 8.75, "8p75"),
++    ],
++)
++def test_memory_rollout_teacher_spec_emits_single_uploadable_job(
++    tmp_path,
++    env_name,
++    latency_label,
++    latency_frames,
++    obs_fps,
++    obs_fps_label,
++):
++    generated_root = tmp_path / "configs"
++    output_root = tmp_path / "outputs"
++    train_dir = tmp_path / "checkpoints"
++
++    generate_sweep_configs(
++        [
++            "--spec",
++            f"configs/experiments/memory/rollout_teacher/{env_name}.yaml",
++            "--generated-root",
++            str(generated_root),
++            "--output-root",
++            str(output_root),
++            "--train-dir",
++            str(train_dir),
++        ]
++    )
++
++    rows = _read_manifest(generated_root / "manifest.tsv")
++    assert len(rows) == 1
++    row = rows[0]
++    assert row["env"] == env_name
++    assert row["experiment_family"] == "memory_rollout_teacher"
++    assert row["sweep_name"] == "teacher"
++    assert row["seed"] == "0"
++    assert row["frame_stack"] == "4"
++    assert row["obs_fps"] == obs_fps_label
++    assert row["latency_label"] == latency_label
++    assert row["latency_kind"] == "fixed"
++    assert row["latency_fixed_frames"] == str(latency_frames)
++    assert row["train_upload_repo_id"] == "latency-sensitive-bench/paper-experiment-models"
++    assert row["train_upload_repo_type"] == "model"
++    assert row["eval_upload_repo_id"] == ""
++    assert row["train_artifact_repo_path"].startswith(
++        f"memory/rollout_teachers/{env_name}/train/"
++    )
++
++    train_cfg = yaml.safe_load(Path(row["train_config"]).read_text(encoding="utf-8"))
++    assert train_cfg["experiment"]["seed"] == 0
++    assert train_cfg["env"]["frame_stack"] == 4
++    assert train_cfg["env"]["obs_fps"] == pytest.approx(obs_fps)
++    assert train_cfg["latency"]["method"] == "fixed"
++    assert train_cfg["latency"]["fixed_latency_ms"] == pytest.approx(
++        latency_frames * 1000 / train_cfg["env"]["env_fps"]
++    )
++    assert train_cfg["evaluation"]["eval_suites"] == {
++        "fixed": [[latency_frames]],
++        "normal": [],
++        "uniform": [],
++    }
++
++
+ def test_latency_bench_eval_config_disables_full_trace_records(tmp_path):
+     base_config = {
+         "backend": {
+diff --git a/tests/scripts/test_manifest_jobs.py b/tests/scripts/test_manifest_jobs.py
+index bd7bf03..4546ac6 100644
+--- a/tests/scripts/test_manifest_jobs.py
++++ b/tests/scripts/test_manifest_jobs.py
+@@ -1,8 +1,10 @@
+ from __future__ import annotations
+ 
++import json
+ import os
+-import subprocess
++import shlex
+ import signal
++import subprocess
+ import sys
+ import threading
+ import time
+@@ -10,6 +12,7 @@ import types
+ from pathlib import Path
+ 
+ import pytest
++from helpers import REPO_ROOT
+ 
+ import scripts.experiment_runner.run_manifest_jobs as run_manifest_jobs
+ from scripts.experiment_runner.job_manifest import (
+@@ -310,6 +313,186 @@ def test_run_command_cleanup_terminates_child_process_group(tmp_path):
+         os.kill(child_pid, 0)
+ 
+ 
++def test_sigint_cancels_queue_and_kills_all_active_process_groups(tmp_path):
++    helper_path = tmp_path / "ignore_term_job.py"
++    helper_path.write_text(
++        """from __future__ import annotations
++
++import json
++import os
++from pathlib import Path
++import signal
++import subprocess
++import sys
++import time
++
++signal.signal(signal.SIGTERM, signal.SIG_IGN)
++child = subprocess.Popen(
++    [
++        sys.executable,
++        "-c",
++        "import signal, time; signal.signal(signal.SIGTERM, signal.SIG_IGN); time.sleep(60)",
++    ]
++)
++Path(sys.argv[1]).write_text(
++    json.dumps(
++        {
++            "leader_pid": os.getpid(),
++            "leader_pgid": os.getpgid(0),
++            "child_pid": child.pid,
++            "child_pgid": os.getpgid(child.pid),
++        }
++    ),
++    encoding="utf-8",
++)
++time.sleep(60)
++""",
++        encoding="utf-8",
++    )
++    manifest_path = tmp_path / "manifest.tsv"
++    markers = [tmp_path / f"job_{idx}.json" for idx in range(4)]
++    rows = []
++    for idx, marker in enumerate(markers):
++        row = {column: "" for column in MANIFEST_V2_COLUMNS}
++        row.update(
++            {
++                "schema_version": "v2",
++                "job_id": f"job{idx}",
++                "train_command": f"exec {shlex.join([sys.executable, str(helper_path), str(marker)])}",
++            }
++        )
++        rows.append(row)
++    write_manifest(manifest_path, rows)
++
++    runner = subprocess.Popen(
++        [
++            sys.executable,
++            "-m",
++            "scripts.experiment_runner.run_manifest_jobs",
++            "--manifest",
++            str(manifest_path),
++            "--stage",
++            "train",
++            "--gpus",
++            "0,1",
++        ],
++        cwd=REPO_ROOT,
++        stdout=subprocess.PIPE,
++        stderr=subprocess.PIPE,
++        text=True,
++    )
++    process_groups: set[int] = set()
++    try:
++        deadline = time.time() + 10.0
++        while sum(marker.is_file() for marker in markers) < 2 and time.time() < deadline:
++            time.sleep(0.05)
++        assert [marker.is_file() for marker in markers] == [True, True, False, False]
++        for marker in markers[:2]:
++            process_info = json.loads(marker.read_text(encoding="utf-8"))
++            assert process_info["leader_pgid"] == process_info["child_pgid"]
++            process_groups.add(int(process_info["leader_pgid"]))
++
++        runner.send_signal(signal.SIGINT)
++        time.sleep(0.1)
++        runner.send_signal(signal.SIGINT)
++        stdout, stderr = runner.communicate(timeout=10.0)
++        assert runner.returncode != 0, (stdout, stderr)
++        assert [marker.is_file() for marker in markers] == [True, True, False, False]
++
++        deadline = time.time() + 5.0
++        while process_groups and time.time() < deadline:
++            for process_group in list(process_groups):
++                try:
++                    os.killpg(process_group, 0)
++                except ProcessLookupError:
++                    process_groups.remove(process_group)
++            time.sleep(0.05)
++        assert not process_groups
++
++        events = load_events(event_log_path_for_manifest(manifest_path))
++        assert [event["stage"] for event in events] == ["train_started", "train_started"]
++    finally:
++        if runner.poll() is None:
++            runner.kill()
++            runner.wait()
++        for process_group in process_groups:
++            try:
++                os.killpg(process_group, signal.SIGKILL)
++            except ProcessLookupError:
++                pass
++
++
++def test_sigint_does_not_wait_for_in_flight_upload(tmp_path):
++    manifest_path = tmp_path / "manifest.tsv"
++    row = {column: "" for column in MANIFEST_V2_COLUMNS}
++    row.update(
++        {
++            "schema_version": "v2",
++            "job_id": "job0",
++            "train_upload_repo_id": "repo-model",
++            "train_upload_repo_type": "model",
++        }
++    )
++    write_manifest(manifest_path, [row])
++    append_event(
++        event_log_path_for_manifest(manifest_path),
++        {"job_id": "job0", "stage": "eval_succeeded"},
++    )
++    upload_started = tmp_path / "upload_started"
++    helper_path = tmp_path / "blocking_upload_runner.py"
++    helper_path.write_text(
++        """from pathlib import Path
++import sys
++import time
++
++from scripts.experiment_runner.run_manifest_jobs import run_jobs
++
++manifest_path = Path(sys.argv[1])
++upload_started = Path(sys.argv[2])
++
++def blocking_uploader(job):
++    del job
++    upload_started.write_text("started", encoding="utf-8")
++    time.sleep(60)
++
++raise SystemExit(
++    run_jobs(
++        manifest_path=manifest_path,
++        gpus=[],
++        upload_workers=1,
++        max_upload_attempts=1,
++        stage="upload",
++        uploader=blocking_uploader,
++    )
++)
++""",
++        encoding="utf-8",
++    )
++    runner = subprocess.Popen(
++        [sys.executable, str(helper_path), str(manifest_path), str(upload_started)],
++        cwd=REPO_ROOT,
++        env={**os.environ, "PYTHONPATH": str(REPO_ROOT)},
++        stdout=subprocess.PIPE,
++        stderr=subprocess.PIPE,
++        text=True,
++    )
++    try:
++        deadline = time.time() + 10.0
++        while not upload_started.is_file() and time.time() < deadline:
++            time.sleep(0.05)
++        assert upload_started.is_file()
++
++        started_at = time.monotonic()
++        runner.send_signal(signal.SIGINT)
++        stdout, stderr = runner.communicate(timeout=5.0)
++        assert runner.returncode != 0, (stdout, stderr)
++        assert time.monotonic() - started_at < 3.0
++    finally:
++        if runner.poll() is None:
++            runner.kill()
++            runner.wait()
++
++
+ def test_default_uploader_prefers_env_token_over_dotenv(monkeypatch, tmp_path):
+     monkeypatch.chdir(tmp_path)
+     (tmp_path / ".env").write_text("HF_TOKEN=dotenv-token\n", encoding="utf-8")

evaluations/deadly-corridor/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/eval/fixed_6/eval_latency_6/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 145.67367553710938, "episode_return_env": 145.67367553710938, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 42, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 13, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 50}
+{"episode_id": 1, "episode_return": 46.77955627441406, "episode_return_env": 46.77955627441406, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 43, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 12, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 46}
+{"episode_id": 2, "episode_return": 382.1748809814453, "episode_return_env": 382.1748809814453, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 44, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 23, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 92}
+{"episode_id": 3, "episode_return": 112.09428405761719, "episode_return_env": 112.09428405761719, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 45, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 13, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 50}
+{"episode_id": 4, "episode_return": 433.02099609375, "episode_return_env": 433.02099609375, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 46, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 23, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 92}
+{"episode_id": 5, "episode_return": 318.6367950439453, "episode_return_env": 318.6367950439453, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 47, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 87}
+{"episode_id": 6, "episode_return": 49.76713562011719, "episode_return_env": 49.76713562011719, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 48, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 12, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 46}
+{"episode_id": 7, "episode_return": 465.71104431152344, "episode_return_env": 465.71104431152344, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 49, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 24, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 96}
+{"episode_id": 8, "episode_return": 97.55632019042969, "episode_return_env": 97.55632019042969, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 50, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 12, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 47}
+{"episode_id": 9, "episode_return": 107.55245971679688, "episode_return_env": 107.55245971679688, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 51, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 13, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 52}
+{"episode_id": 10, "episode_return": 73.88134765625, "episode_return_env": 73.88134765625, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 52, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 11, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 44}
+{"episode_id": 11, "episode_return": 321.43589782714844, "episode_return_env": 321.43589782714844, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 53, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 21, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 83}
+{"episode_id": 12, "episode_return": 340.1656494140625, "episode_return_env": 340.1656494140625, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 54, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 21, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 84}
+{"episode_id": 13, "episode_return": 358.41978454589844, "episode_return_env": 358.41978454589844, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 55, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 88}
+{"episode_id": 14, "episode_return": 433.22064208984375, "episode_return_env": 433.22064208984375, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 56, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 25, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.4285714285714, "return_raw": null, "survival_steps": 97}
+{"episode_id": 15, "episode_return": 420.3064727783203, "episode_return_env": 420.3064727783203, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 57, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 21, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 82}
+{"episode_id": 16, "episode_return": 409.05506896972656, "episode_return_env": 409.05506896972656, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 58, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 24, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 95}
+{"episode_id": 17, "episode_return": 93.52970886230469, "episode_return_env": 93.52970886230469, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 59, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 13, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 51}
+{"episode_id": 18, "episode_return": 102.85501098632812, "episode_return_env": 102.85501098632812, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 60, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 13, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 49}
+{"episode_id": 19, "episode_return": 67.03433227539062, "episode_return_env": 67.03433227539062, "game_score": null, "mean_latency_ms": 171.42857142857142, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "configured_latency_ms": 171.42857142857142, "env_fps": 35.0, "env_id": "doom_deadly_corridor", "episode_seed": 61, "eval_latency_steps": 6, "frame_ms": 28.571428571428573, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 8.75, "output_dir": "/workspace/evals/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/latency_bench_eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 11, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 171.42857142857142, "p99_latency_ms": 171.42857142857142, "return_raw": null, "survival_steps": 44}

evaluations/deadly-corridor/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/eval/fixed_6/queue_eval_latency_6.json

@@ -0,0 +1,59 @@
+{
+  "checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt",
+  "configured_latency_ms": 171.42857142857142,
+  "eval_latency_steps": 6,
+  "experiment_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400",
+  "latency": 6,
+  "latency_type": "fixed_sweep",
+  "lengths": [
+    50,
+    46,
+    92,
+    50,
+    92,
+    87,
+    46,
+    96,
+    47,
+    52,
+    44,
+    83,
+    84,
+    88,
+    97,
+    82,
+    95,
+    51,
+    49,
+    44
+  ],
+  "mean_length": 68.75,
+  "mean_return": 238.94355316162108,
+  "returns": [
+    145.67367553710938,
+    46.77955627441406,
+    382.1748809814453,
+    112.09428405761719,
+    433.02099609375,
+    318.6367950439453,
+    49.76713562011719,
+    465.71104431152344,
+    97.55632019042969,
+    107.55245971679688,
+    73.88134765625,
+    321.43589782714844,
+    340.1656494140625,
+    358.41978454589844,
+    433.22064208984375,
+    420.3064727783203,
+    409.05506896972656,
+    93.52970886230469,
+    102.85501098632812,
+    67.03433227539062
+  ],
+  "seed": 42,
+  "std_return": 154.6087889571668,
+  "suite_name": "fixed_6",
+  "timestamp_utc": "2026-07-29T08:50:52.080822+00:00",
+  "train_latency_steps": 6
+}

evaluations/deadly-corridor/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/eval/fixed_6/queue_eval_results.jsonl

@@ -0,0 +1 @@
+{"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "configured_latency_ms": 171.42857142857142, "eval_latency_steps": 6, "experiment_name": "starvla_wan_oft_deadly_corridor_raw6_maxsteps14400", "latency": 6, "latency_type": "fixed_sweep", "lengths": [50, 46, 92, 50, 92, 87, 46, 96, 47, 52, 44, 83, 84, 88, 97, 82, 95, 51, 49, 44], "mean_length": 68.75, "mean_return": 238.94355316162108, "returns": [145.67367553710938, 46.77955627441406, 382.1748809814453, 112.09428405761719, 433.02099609375, 318.6367950439453, 49.76713562011719, 465.71104431152344, 97.55632019042969, 107.55245971679688, 73.88134765625, 321.43589782714844, 340.1656494140625, 358.41978454589844, 433.22064208984375, 420.3064727783203, 409.05506896972656, 93.52970886230469, 102.85501098632812, 67.03433227539062], "seed": 42, "std_return": 154.6087889571668, "suite_name": "fixed_6", "timestamp_utc": "2026-07-29T08:50:52.080822+00:00", "train_latency_steps": 6}

evaluations/demon-attack/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/train/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/git.diff

@@ -0,0 +1,1829 @@
+diff --git a/configs/experiments/fixed_latency_curve.yaml b/configs/experiments/fixed_latency_curve.yaml
+deleted file mode 100644
+index 3ab184f..0000000
+--- a/configs/experiments/fixed_latency_curve.yaml
++++ /dev/null
+@@ -1,17 +0,0 @@
+-__base__: ../base/dummy.yaml
+-
+-experiment:
+-  name: fixed_latency_curve
+-
+-executor:
+-  mode: simulated
+-
+-latency:
+-  method: fixed
+-  fixed_latency_ms: 100
+-
+-evaluation:
+-  eval_episodes: 20
+-
+-logging:
+-  output_dir: outputs/fixed_latency_curve
+diff --git a/configs/experiments/flappy/frame_stack/base.yaml b/configs/experiments/flappy/frame_stack/base.yaml
+deleted file mode 100644
+index f1557d5..0000000
+--- a/configs/experiments/flappy/frame_stack/base.yaml
++++ /dev/null
+@@ -1,123 +0,0 @@
+-experiment:
+-  name: flappy_base
+-  seed: 0
+-
+-backend:
+-  type: sample_factory
+-  algo: APPO
+-  device: cuda
+-  train_dir: results/checkpoints
+-  restart_behavior: resume
+-  run_mode: train
+-
+-executor:
+-  mode: simulated
+-
+-env:
+-  name: flappy
+-  # note that env_fps must be an integer multiple of obs_fps!
+-  env_fps: 30
+-  obs_fps: 30
+-  gym_id: FlappyBird-v0
+-  render_mode: rgb_array
+-  observation_mode: image
+-  use_lidar: false
+-  normalize_obs: true
+-  audio_on: false
+-  frame_stack: 4
+-  obs_resize:
+-    - 84
+-    - 84
+-  # for training it could be true, but for realtime mode, THIS MUST BE FALSE!
+-  use_gpu_render: true
+-  simulator: gpu
+-  gpu_render_device: auto
+-  gpu_render_profile: false
+-  gpu_render_profile_interval: 200
+-  noop_action: noop
+-  action_map:
+-    noop: 0
+-    flap: 1
+-  oneshot_actions:
+-    - flap
+-
+-latency:
+-  method: normal
+-  mean_latency_ms: 66.7
+-  std_latency_ms: 33.3
+-  min_latency_ms: 0
+-  max_latency_ms: 500.0
+-
+-
+-scheduler:
+-  # [one_frame_then_noop, hold]
+-  # if hold, default hold duration is: env_fps / obs_fps
+-  hold_policy: one_frame_then_noop
+-  # can be either `issue_order_fifo` or `latest ready`
+-  ordering_policy: latest_ready
+-
+-policy:
+-  type: random
+-  actions:
+-    - noop
+-    - flap
+-
+-training:
+-  train_for_env_steps: 25000000
+-  num_workers: 2
+-  num_envs_per_worker: 128
+-  worker_num_splits: 1
+-  num_policies: 1
+-  batch_size: 4096
+-  rollout: 128
+-  recurrence: 1
+-  num_epochs: 8
+-  num_batches_per_epoch: 8
+-  max_policy_lag: 400
+-  learning_rate: 0.00025
+-  gamma: 0.99
+-  gae_lambda: 0.95
+-  ppo_clip_ratio: 0.1
+-  ppo_clip_value: 0.2
+-  value_loss_coeff: 0.5
+-  max_grad_norm: 0.5
+-  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: 1000000
+-  eval_episodes: 100
+-  eval_parallel_envs: 100
+-  eval_max_steps: 3600
+-  eval_deterministic: true
+-  eval_raw_reward: false
+-  eval_suites:
+-    fixed:
+-      - [0, 8]    # [start, end]
+-    normal:
+-      - [2.0, 0.5, 0, 15]   # low latency, low jitter
+-      - [2.0, 1.0, 0, 15]   # low latency, medium jitter
+-      - [4.0, 0.5, 0, 15]   # high latency, low jitter
+-      - [4.0, 1.0, 0, 15]   # high latency, medium jitter
+-      - [4.0, 2.0, 0, 15]   # high latency, high jitter
+-    uniform:
+-      - [1, 3]  # low latency, low jitter
+-      - [0, 4]  # low latency, high jitter
+-      - [3, 5]  # high latency, low jitter
+-      - [2, 6]  # high latency, high jitter
+-
+-logging:
+-  output_dir: outputs/flappy_base
+-  save_step_records: true
+-  save_action_records: true
+-  save_latency_records: true
+-  wandb_project: latency-sensitive-bench
+-  wandb_group: flappy-frame_stack
+-  wandb_job_type: sample_factory
+-  wandb_tags: ""
+diff --git a/latency_bench/data/exporters/starvla_lerobot.py b/latency_bench/data/exporters/starvla_lerobot.py
+index c693460..4ad276a 100644
+--- a/latency_bench/data/exporters/starvla_lerobot.py
++++ b/latency_bench/data/exporters/starvla_lerobot.py
+@@ -18,12 +18,17 @@ from latency_bench.utils.io import write_jsonl
+ from latency_bench.data.action_layouts import action_spec_for_env
+ from latency_bench.data.filters import density_weights_clean_v1
+ 
+-FPS_BY_ENV = {
+-    "flappy": 30,
+-    "demon_attack": 60,
+-    "deadly_corridor": 35,
+-}
+ STATE_DIM = 1
++BRIDGE_DIM = 7
++DEADLY_CORRIDOR_BRIDGE_ACTION_LABELS = (
++    "MOVE_FORWARD",
++    "MOVE_BACKWARD",
++    "MOVE_LEFT",
++    "MOVE_RIGHT",
++    "TURN_LEFT",
++    "TURN_RIGHT",
++    "ATTACK",
++)
+ 
+ 
+ def _png_bytes(image: Any) -> bytes:
+@@ -64,30 +69,38 @@ def _load_rows(raw_export_dir: Path, split: str) -> list[dict[str, Any]]:
+ 
+ 
+ def _episode_groups(rows: list[dict[str, Any]]) -> list[list[dict[str, Any]]]:
+-    """Group canonical rows by episode and order by decision step.
++    """Group canonical rows by trajectory and order by decision step.
+ 
+     Args:
+         rows: Canonical rollout rows from one split.
+ 
+     Returns:
+-        Episode row groups sorted by episode index.
++        Trajectory row groups sorted by episode and trajectory index.
+     """
+-    grouped: dict[int, list[dict[str, Any]]] = {}
++    grouped: dict[tuple[int, int], list[dict[str, Any]]] = {}
+     for row in rows:
+-        grouped.setdefault(int(row["episode_idx"]), []).append(row)
++        key = (int(row["episode_idx"]), int(row["trajectory_idx"]))
++        grouped.setdefault(key, []).append(row)
+     return [
+         sorted(episode_rows, key=lambda row: int(row["decision_step"]))
+         for _, episode_rows in sorted(grouped.items())
+     ]
+ 
+ 
+-def _write_episode(path: Path, rows: list[dict[str, Any]], *, action_dim: int) -> None:
++def _write_episode(
++    path: Path,
++    rows: list[dict[str, Any]],
++    *,
++    action_dim: int,
++    state_dim: int,
++) -> None:
+     """Write one LeRobot episode parquet file.
+ 
+     Args:
+         path: Destination episode parquet path.
+         rows: Converted LeRobot frame rows.
+         action_dim: Expected action vector dimension.
++        state_dim: Expected state carrier dimension.
+     """
+     path.parent.mkdir(parents=True, exist_ok=True)
+     table = pa.table(
+@@ -97,8 +110,8 @@ def _write_episode(path: Path, rows: list[dict[str, Any]], *, action_dim: int) -
+                 type=pa.struct([("bytes", pa.binary()), ("path", pa.string())]),
+             ),
+             "observation.state": pa.array(
+-                [[0.0] * STATE_DIM for _ in rows],
+-                type=pa.list_(pa.float32(), STATE_DIM),
++                [[0.0] * state_dim for _ in rows],
++                type=pa.list_(pa.float32(), state_dim),
+             ),
+             "action": pa.array(
+                 [row["action"] for row in rows],
+@@ -131,11 +144,47 @@ def _row_action_vector(row: dict[str, Any], *, action_dim: int) -> list[float]:
+     return [1.0 if idx == action_id else 0.0 for idx in range(action_dim)]
+ 
+ 
++def _bridge_action_vector(
++    row: dict[str, Any],
++    *,
++    env_name: str,
++    active_action_dim: int,
++) -> list[float]:
++    if env_name == "deadly_corridor":
++        action_text = str(row["action_text"]).upper()
++        return [
++            1.0 if label in action_text else 0.0
++            for label in DEADLY_CORRIDOR_BRIDGE_ACTION_LABELS
++        ]
++    action = _row_action_vector(row, action_dim=active_action_dim)
++    return [*action, *([0.0] * (BRIDGE_DIM - active_action_dim))]
++
++
++def _target_action_labels(
++    *,
++    env_name: str,
++    source_action_labels: list[str],
++    action_carrier: str,
++) -> list[str]:
++    if action_carrier == "native":
++        return source_action_labels
++    if env_name == "deadly_corridor":
++        return list(DEADLY_CORRIDOR_BRIDGE_ACTION_LABELS)
++    return [
++        *source_action_labels,
++        *[
++            f"BRIDGE_PAD_{idx}"
++            for idx in range(len(source_action_labels), BRIDGE_DIM)
++        ],
++    ]
++
++
+ def _write_metadata(
+     dataset_dir: Path,
+     *,
+-    fps: int,
++    fps: float,
+     action_labels: list[str],
++    state_dim: int,
+     episode_lengths: list[int],
+     task_prompts: list[str],
+ ) -> None:
+@@ -144,7 +193,8 @@ def _write_metadata(
+     Args:
+         dataset_dir: LeRobot dataset split directory.
+         fps: Frame rate stored in metadata.
+-        action_labels: Canonical action labels in vector order.
++        action_labels: Target action labels in vector order.
++        state_dim: State carrier dimension.
+         episode_lengths: Number of frames per converted episode.
+         task_prompts: Unique prompts used as LeRobot tasks.
+     """
+@@ -155,10 +205,10 @@ def _write_metadata(
+         json.dumps(
+             {
+                 "codebase_version": "v2.0",
+-                "fps": int(fps),
++                "fps": fps,
+                 "features": {
+                     "observation.image": {"dtype": "image"},
+-                    "observation.state": {"dtype": "float32", "shape": [STATE_DIM]},
++                    "observation.state": {"dtype": "float32", "shape": [state_dim]},
+                     "action": {"dtype": "float32", "shape": [action_dim], "names": action_labels},
+                     "timestamp": {"dtype": "float64", "shape": [1]},
+                     "episode_index": {"dtype": "int64", "shape": [1]},
+@@ -183,9 +233,11 @@ def _convert_split(
+     *,
+     env_name: str,
+     action_labels: list[str],
+-    fps: int,
++    fps: float,
+     reward_field: str,
+-    obs_stride: int = 1,
++    action_carrier: str,
++    active_action_dim: int,
++    state_dim: int,
+     density_weighting: bool = False,
+ ) -> dict[str, Any]:
+     """Convert one canonical rollout split into a LeRobot dataset directory.
+@@ -195,13 +247,11 @@ def _convert_split(
+         output_dir: Destination LeRobot split directory.
+         env_name: Canonical benchmark environment name.
+         action_labels: Canonical action labels in vector order.
+-        fps: Frame rate used for timestamps.
++        fps: Observation frame rate used for timestamps.
+         reward_field: Source canonical reward column.
+-        obs_stride: Raw frames per observation decision (``round(env_fps/obs_fps)``).
+-            Canonical rows are recorded once per raw frame; keeping only the obs
+-            decisions (``decision_step % obs_stride == 0``) makes adjacent LeRobot
+-            rows adjacent observations, so KV-memory rollouts unroll at the eval
+-            observation cadence.
++        action_carrier: Target action/state carrier.
++        active_action_dim: Number of active environment action dimensions.
++        state_dim: Target state carrier dimension.
+         density_weighting: When True, write per-row ``density_weight`` from clean_v1's
+             event balancing (train split). When False, weights are uniform 1.0 (val
+             split keeps an unweighted held-out loss).
+@@ -214,7 +264,6 @@ def _convert_split(
+     task_prompts: list[str] = []
+     episode_lengths: list[int] = []
+     for new_episode_idx, episode_rows in enumerate(_episode_groups(rows)):
+-        episode_rows = [row for row in episode_rows if int(row["decision_step"]) % obs_stride == 0]
+         density_weights = (
+             density_weights_clean_v1(env_name, episode_rows)
+             if density_weighting
+@@ -231,8 +280,14 @@ def _convert_split(
+             out_rows.append(
+                 {
+                     "image_bytes": _png_bytes(row["image"]),
+-                    "action": _row_action_vector(row, action_dim=len(action_labels)),
+-                    "timestamp": float(frame_idx) / int(fps),
++                    "action": _bridge_action_vector(
++                        row,
++                        env_name=env_name,
++                        active_action_dim=active_action_dim,
++                    )
++                    if action_carrier == "bridge"
++                    else _row_action_vector(row, action_dim=len(action_labels)),
++                    "timestamp": float(frame_idx) / fps,
+                     "episode_index": new_episode_idx,
+                     "frame_index": frame_idx,
+                     "task_index": prompt_to_task_index[prompt],
+@@ -248,11 +303,13 @@ def _convert_split(
+             output_dir / f"data/chunk-{episode_chunk:03d}/episode_{new_episode_idx:06d}.parquet",
+             out_rows,
+             action_dim=len(action_labels),
++            state_dim=state_dim,
+         )
+     _write_metadata(
+         output_dir,
+         fps=fps,
+         action_labels=action_labels,
++        state_dim=state_dim,
+         episode_lengths=episode_lengths,
+         task_prompts=task_prompts,
+     )
+@@ -269,6 +326,7 @@ def convert_canonical_rollout_to_lerobot(
+     raw_export_dir: Path | str,
+     output_dir: Path | str,
+     *,
++    action_carrier: str = "native",
+     force: bool = False,
+ ) -> dict[str, Any]:
+     """Convert a canonical rollout export into train/val LeRobot datasets.
+@@ -276,6 +334,7 @@ def convert_canonical_rollout_to_lerobot(
+     Args:
+         raw_export_dir: Directory with canonical rollout parquet files and metadata.
+         output_dir: Destination LeRobot train dataset directory.
++        action_carrier: Target action/state carrier, ``native`` or ``bridge``.
+         force: Remove existing train/val output directories before writing.
+ 
+     Returns:
+@@ -286,11 +345,20 @@ def convert_canonical_rollout_to_lerobot(
+     metadata = json.loads((raw_export_dir / "metadata.json").read_text(encoding="utf-8"))
+     env_name = str(metadata["env_name"])
+     action_spec = action_spec_for_env(env_name)
+-    action_labels = list(metadata["action_labels"])
++    source_action_labels = list(metadata["action_labels"])
++    action_labels = _target_action_labels(
++        env_name=env_name,
++        source_action_labels=source_action_labels,
++        action_carrier=action_carrier,
++    )
++    active_action_dim = {
++        "flappy": 2,
++        "demon_attack": 6,
++        "deadly_corridor": 7,
++    }[env_name]
++    state_dim = {"native": STATE_DIM, "bridge": BRIDGE_DIM}[action_carrier]
+     reward_field = str(metadata["reward_field"])
+-    obs_stride = int(metadata.get("obs_stride_raw_frames", 1))
+-    # Rows survive at the observation cadence, so timestamps advance at obs_fps.
+-    fps = max(1, int(FPS_BY_ENV[env_name]) // obs_stride)
++    fps = metadata["obs_fps"]
+     val_output_dir = output_dir.with_name(f"{output_dir.name}__val")
+     if output_dir.exists() and force:
+         shutil.rmtree(output_dir)
+@@ -303,7 +371,9 @@ def convert_canonical_rollout_to_lerobot(
+         action_labels=action_labels,
+         fps=fps,
+         reward_field=reward_field,
+-        obs_stride=obs_stride,
++        action_carrier=action_carrier,
++        active_action_dim=active_action_dim,
++        state_dim=state_dim,
+         density_weighting=True,
+     )
+     val_manifest = _convert_split(
+@@ -313,7 +383,9 @@ def convert_canonical_rollout_to_lerobot(
+         action_labels=action_labels,
+         fps=fps,
+         reward_field=reward_field,
+-        obs_stride=obs_stride,
++        action_carrier=action_carrier,
++        active_action_dim=active_action_dim,
++        state_dim=state_dim,
+         density_weighting=False,
+     )
+     manifest = {
+@@ -321,9 +393,13 @@ def convert_canonical_rollout_to_lerobot(
+         "format": "starvla_lerobot_v2_image_parquet",
+         "source": str(raw_export_dir),
+         "action_layout": str(metadata["action_layout"]),
++        "action_carrier": action_carrier,
+         "action_labels": action_labels,
+         "action_dim": len(action_labels),
+-        "active_action_dim": len(action_spec.labels),
++        "active_action_dim": active_action_dim
++        if action_carrier == "bridge"
++        else len(action_spec.labels),
++        "state_dim": state_dim,
+         "validation_dataset_name": val_output_dir.name,
+         "validation_episodes": val_manifest["episodes"],
+         "validation_frames": val_manifest["frames"],
+diff --git a/latency_bench/data/hf_parquet.py b/latency_bench/data/hf_parquet.py
+index 7a26426..6163c0d 100644
+--- a/latency_bench/data/hf_parquet.py
++++ b/latency_bench/data/hf_parquet.py
+@@ -205,7 +205,9 @@ def completed_episode_indices_from_shards(output_dir: Path) -> set[int]:
+             with shard_path.open("r", encoding="utf-8") as handle:
+                 for line in handle:
+                     if line.strip():
+-                        completed_episode_indices.add(int(json.loads(line)["episode_idx"]))
++                        row = json.loads(line)
++                        row["trajectory_idx"]
++                        completed_episode_indices.add(int(row["episode_idx"]))
+                         break
+     return completed_episode_indices
+ 
+diff --git a/latency_bench/data/rollout_schema.py b/latency_bench/data/rollout_schema.py
+index 11426b9..21d2bf5 100644
+--- a/latency_bench/data/rollout_schema.py
++++ b/latency_bench/data/rollout_schema.py
+@@ -7,11 +7,12 @@ from datasets import Features, Image as HFImage, Sequence as HFSequence, Value
+ from latency_bench.data.action_layouts import ActionSpec
+ 
+ 
+-SCHEMA_VERSION = 2
++SCHEMA_VERSION = 3
+ CANONICAL_ROW_FIELDS = (
+     "image",
+     "env_name",
+     "episode_idx",
++    "trajectory_idx",
+     "decision_step",
+     "split",
+     "seed",
+@@ -26,6 +27,7 @@ def _rollout_feature_map(action_spec: ActionSpec) -> dict[str, Any]:
+         "image": HFImage(),
+         "env_name": Value("string"),
+         "episode_idx": Value("int64"),
++        "trajectory_idx": Value("int64"),
+         "decision_step": Value("int64"),
+         "split": Value("string"),
+         "seed": Value("int64"),
+diff --git a/latency_bench/data/sf_teacher_rollout.py b/latency_bench/data/sf_teacher_rollout.py
+index af2b64f..bbbd157 100644
+--- a/latency_bench/data/sf_teacher_rollout.py
++++ b/latency_bench/data/sf_teacher_rollout.py
+@@ -355,6 +355,7 @@ def _write_filtered_episode_images(
+     image_writer: RolloutImageWriter,
+     rows: list[dict[str, Any]],
+     frames_by_decision_step: Mapping[int, np.ndarray],
++    trajectory_by_decision_step: Mapping[int, int],
+     *,
+     context_window: int,
+ ) -> list[dict[str, Any]]:
+@@ -364,6 +365,7 @@ def _write_filtered_episode_images(
+         image_writer: Image writer shared by the rollout run.
+         rows: Final rows selected for materialization.
+         frames_by_decision_step: Captured frames keyed by decision step.
++        trajectory_by_decision_step: Trajectory index keyed by decision step.
+         context_window: Total image window length including the current ``image`` field.
+ 
+     Returns:
+@@ -371,10 +373,15 @@ def _write_filtered_episode_images(
+     """
+     if not rows:
+         return []
+-    available_decision_steps = sorted(frames_by_decision_step)
+     required_steps: set[int] = set()
+     for row in rows:
+         decision_step = int(row["decision_step"])
++        trajectory_idx = int(row["trajectory_idx"])
++        available_decision_steps = sorted(
++            step
++            for step, step_trajectory_idx in trajectory_by_decision_step.items()
++            if step_trajectory_idx == trajectory_idx
++        )
+         required_steps.add(decision_step)
+         if int(context_window) > 1:
+             required_steps.update(
+@@ -396,6 +403,12 @@ def _write_filtered_episode_images(
+     for row in rows:
+         written = dict(row)
+         decision_step = int(written["decision_step"])
++        trajectory_idx = int(written["trajectory_idx"])
++        available_decision_steps = sorted(
++            step
++            for step, step_trajectory_idx in trajectory_by_decision_step.items()
++            if step_trajectory_idx == trajectory_idx
++        )
+         written["image"] = written_paths_by_step[decision_step]
+         if int(context_window) > 1:
+             written["context_images"] = [
+@@ -804,6 +817,7 @@ def _replay_accepted_specs(
+     active_episode_rows: list[list[dict[str, Any]]] = [[] for _ in range(driver.slot_count)]
+     active_episode_frames: list[list[np.ndarray]] = [[] for _ in range(driver.slot_count)]
+     active_decision_steps = [0 for _ in range(driver.slot_count)]
++    active_trajectory_indices = [0 for _ in range(driver.slot_count)]
+     active_ghost_streams: list[Any | None] = [None for _ in range(driver.slot_count)]
+     rnn_states = torch.zeros([driver.slot_count, rnn_size], dtype=torch.float32, device=device)
+     image_writer = RolloutImageWriter(
+@@ -844,6 +858,7 @@ def _replay_accepted_specs(
+                 active_episode_rows[slot] = []
+                 active_episode_frames[slot] = []
+                 active_decision_steps[slot] = 0
++                active_trajectory_indices[slot] = 0
+                 rnn_states[slot].zero_()
+ 
+         with tqdm(total=len(specs_to_replay), initial=0, unit="episode") as progress:
+@@ -881,21 +896,21 @@ def _replay_accepted_specs(
+                         next_decision_step=next_decision_step,
+                         max_steps_per_episode=int(max_steps_per_episode),
+                     )
+-                    if not life_loss_boundary:
+-                        active_episode_frames[slot].append(np.array(image_frame, copy=True))
+-                        row = teacher_dataset_row(
+-                            env_name=env_name,
+-                            spec=spec,
+-                            decision_step=decision_step,
+-                            image=str(image_path),
+-                            raw_action=raw_actions[batch_idx],
+-                            info=step_result.info,
+-                            reward=step_result.reward,
+-                            env_fps=env_fps,
+-                            obs_fps=obs_fps,
+-                            image_transform=image_transform_config.normalized_transform(),
+-                        )
+-                        active_episode_rows[slot].append(row)
++                    active_episode_frames[slot].append(np.array(image_frame, copy=True))
++                    row = teacher_dataset_row(
++                        env_name=env_name,
++                        spec=spec,
++                        trajectory_idx=active_trajectory_indices[slot],
++                        decision_step=decision_step,
++                        image=str(image_path),
++                        raw_action=raw_actions[batch_idx],
++                        info=step_result.info,
++                        reward=step_result.reward,
++                        env_fps=env_fps,
++                        obs_fps=obs_fps,
++                        image_transform=image_transform_config.normalized_transform(),
++                    )
++                    active_episode_rows[slot].append(row)
+                     rnn_states[slot] = new_rnn_states[batch_idx]
+                     if episode_done:
+                         final_rows = _finalize_attempt_rows(
+@@ -931,6 +946,10 @@ def _replay_accepted_specs(
+                             image_writer,
+                             shard_rows,
+                             frames_by_decision_step,
++                            {
++                                int(row["decision_step"]): int(row["trajectory_idx"])
++                                for row in final_rows
++                            },
+                             context_window=int(context_window),
+                         )
+                         write_episode_row_shard(output_dir, shard_rows)
+@@ -943,12 +962,14 @@ def _replay_accepted_specs(
+                         active_episode_rows[slot] = []
+                         active_episode_frames[slot] = []
+                         active_decision_steps[slot] = 0
++                        active_trajectory_indices[slot] = 0
+                         if _is_demon_ghost_trail:
+                             active_ghost_streams[slot] = None
+                         rnn_states[slot].zero_()
+                     else:
+                         if life_loss_boundary:
+                             active_obs[slot], reset_info = driver.continue_after_life_loss(slot)
++                            active_trajectory_indices[slot] += 1
+                             if _is_demon_ghost_trail:
+                                 active_ghost_streams[slot].clear()
+                                 active_image_frames[slot] = frame_from_env_raw_rgb_info(reset_info).copy()
+@@ -981,6 +1002,7 @@ def _replay_accepted_specs(
+                         active_episode_rows[slot] = []
+                         active_episode_frames[slot] = []
+                         active_decision_steps[slot] = 0
++                        active_trajectory_indices[slot] = 0
+                         rnn_states[slot].zero_()
+                 else:
+                     if isinstance(driver, BatchedTeacherRolloutDriver):
+diff --git a/latency_bench/data/sf_teacher_rows.py b/latency_bench/data/sf_teacher_rows.py
+index efa6f3e..4e7ce9a 100644
+--- a/latency_bench/data/sf_teacher_rows.py
++++ b/latency_bench/data/sf_teacher_rows.py
+@@ -14,6 +14,7 @@ def teacher_dataset_row(
+     *,
+     env_name: str,
+     spec: Mapping[str, Any],
++    trajectory_idx: int,
+     decision_step: int,
+     image: str,
+     raw_action: Any,
+@@ -28,6 +29,7 @@ def teacher_dataset_row(
+     Args:
+         env_name: Canonical benchmark environment name.
+         spec: Episode spec containing index, split, and seed.
++        trajectory_idx: Zero-based trajectory index within the full-game episode.
+         decision_step: Zero-based policy decision step within the episode.
+         image: Path to the saved pre-step image.
+         raw_action: Raw teacher action for this decision.
+@@ -48,6 +50,7 @@ def teacher_dataset_row(
+         "image": image,
+         "env_name": str(env_name),
+         "episode_idx": int(spec["episode_idx"]),
++        "trajectory_idx": int(trajectory_idx),
+         "decision_step": int(decision_step),
+         "split": str(spec["split"]),
+         "seed": int(spec["seed"]),
+diff --git a/scripts/experiment_runner/run_manifest_jobs.py b/scripts/experiment_runner/run_manifest_jobs.py
+index e2d1b5e..3ba106a 100644
+--- a/scripts/experiment_runner/run_manifest_jobs.py
++++ b/scripts/experiment_runner/run_manifest_jobs.py
+@@ -73,11 +73,6 @@ def _event_writer(event_log: Path) -> Callable[[dict[str, Any]], None]:
+     return write
+ 
+ 
+-def _register_process(process: subprocess.Popen) -> None:
+-    with _ACTIVE_PROCESSES_LOCK:
+-        _ACTIVE_PROCESSES.add(process)
+-
+-
+ def _forget_process(process: subprocess.Popen) -> None:
+     with _ACTIVE_PROCESSES_LOCK:
+         _ACTIVE_PROCESSES.discard(process)
+@@ -89,11 +84,10 @@ def _active_processes() -> list[subprocess.Popen]:
+ 
+ 
+ def _signal_process_group(process: subprocess.Popen, sig: int) -> None:
+-    if process.poll() is None:
+-        try:
+-            os.killpg(process.pid, sig)
+-        except ProcessLookupError:
+-            pass
++    try:
++        os.killpg(process.pid, sig)
++    except ProcessLookupError:
++        pass
+ 
+ 
+ def _terminate_active_processes(sig: int = signal.SIGTERM) -> None:
+@@ -118,15 +112,24 @@ def _terminate_active_processes(sig: int = signal.SIGTERM) -> None:
+             pass
+ 
+ 
+-def _run_command(command: str, *, gpu: str) -> None:
++def _run_command(
++    command: str,
++    *,
++    gpu: str,
++    cancel_event: threading.Event | None = None,
++) -> bool:
+     env = os.environ.copy()
+     env["CUDA_VISIBLE_DEVICES"] = str(gpu)
+-    process = subprocess.Popen(command, shell=True, env=env, start_new_session=True)
+-    _register_process(process)
++    with _ACTIVE_PROCESSES_LOCK:
++        if cancel_event is not None and cancel_event.is_set():
++            return False
++        process = subprocess.Popen(command, shell=True, env=env, start_new_session=True)
++        _ACTIVE_PROCESSES.add(process)
+     try:
+         returncode = process.wait()
+         if returncode != 0:
+             raise subprocess.CalledProcessError(returncode, command)
++        return True
+     finally:
+         _forget_process(process)
+ 
+@@ -201,9 +204,9 @@ def _resume_stage(job: dict[str, str], states: dict[str, str]) -> str:
+     stage = states.get(str(job["job_id"]), "")
+     if stage in {"job_complete", "upload_skipped", "upload_succeeded"}:
+         return "complete"
+-    if stage in {"eval_succeeded", "upload_failed"}:
++    if stage in {"eval_succeeded", "upload_started", "upload_failed"}:
+         return "upload"
+-    if stage == "train_succeeded":
++    if stage in {"train_succeeded", "eval_started"}:
+         return "eval"
+     if stage == "eval_failed":
+         # Training already succeeded before eval ran; re-eval from the checkpoint
+@@ -223,16 +226,6 @@ def run_jobs(
+     dry_run: bool = False,
+     uploader: Callable[[dict[str, str]], None] | None = None,
+ ) -> int:
+-    previous_sigint = signal.getsignal(signal.SIGINT)
+-    previous_sigterm = signal.getsignal(signal.SIGTERM)
+-
+-    def stop_active_processes(signum, frame):
+-        del frame
+-        _terminate_active_processes(signal.SIGTERM)
+-        raise KeyboardInterrupt(signum)
+-
+-    signal.signal(signal.SIGINT, stop_active_processes)
+-    signal.signal(signal.SIGTERM, stop_active_processes)
+     manifest_file = Path(manifest_path).expanduser().resolve()
+     event_log = event_log_path_for_manifest(manifest_file)
+     write_event = _event_writer(event_log)
+@@ -253,12 +246,16 @@ def run_jobs(
+         jobs = [
+             row
+             for row in rows
+-            if states.get(row["job_id"], "") in {"eval_succeeded", "upload_failed"}
++            if states.get(row["job_id"], "")
++            in {"eval_succeeded", "upload_started", "upload_failed"}
+         ]
+     job_queue: queue.Queue[dict[str, str] | None] = queue.Queue()
+     upload_queue: queue.Queue[UploadTask | None] = queue.Queue()
+     failures: list[tuple[str, str]] = []
+     failure_lock = threading.Lock()
++    cancel_event = threading.Event()
++    signal_event = threading.Event()
++    cleanup_started = False
+     uploader_fn = _default_uploader if uploader is None else uploader
+ 
+     if stage != "upload":
+@@ -280,28 +277,49 @@ def run_jobs(
+             if job is None:
+                 job_queue.task_done()
+                 break
++            if cancel_event.is_set():
++                job_queue.task_done()
++                continue
+             next_stage = _resume_stage(job, states) if stage == "all" else stage
+             job_id = str(job["job_id"])
+             try:
+                 if next_stage == "train":
++                    if cancel_event.is_set():
++                        continue
+                     if dry_run:
+                         print(f"[dry-run] CUDA_VISIBLE_DEVICES={gpu} {job['train_command']}", flush=True)
+                     else:
+                         write_event({"job_id": job_id, "stage": "train_started", "gpu": gpu, "ts": _now()})
+-                        _run_command(str(job["train_command"]), gpu=gpu)
++                        started = _run_command(
++                            str(job["train_command"]),
++                            gpu=gpu,
++                            cancel_event=cancel_event,
++                        )
++                        if not started or cancel_event.is_set():
++                            continue
+                         write_event({"job_id": job_id, "stage": "train_succeeded", "gpu": gpu, "ts": _now()})
+                     next_stage = "eval" if stage == "all" else "complete"
+                 if next_stage == "eval":
++                    if cancel_event.is_set():
++                        continue
+                     if dry_run:
+                         print(f"[dry-run] CUDA_VISIBLE_DEVICES={gpu} {job['eval_command']}", flush=True)
+                     else:
+                         write_event({"job_id": job_id, "stage": "eval_started", "gpu": gpu, "ts": _now()})
+-                        _run_command(str(job["eval_command"]), gpu=gpu)
++                        started = _run_command(
++                            str(job["eval_command"]),
++                            gpu=gpu,
++                            cancel_event=cancel_event,
++                        )
++                        if not started or cancel_event.is_set():
++                            continue
+                         write_event({"job_id": job_id, "stage": "eval_succeeded", "gpu": gpu, "ts": _now()})
+                     next_stage = "upload" if stage == "all" else "complete"
+-                if next_stage == "upload":
++                if next_stage == "upload" and not cancel_event.is_set():
+                     upload_queue.put(UploadTask(job))
+             except subprocess.CalledProcessError:
++                if cancel_event.is_set():
++                    continue
+                 failed_stage = "train_failed" if next_stage == "train" else "eval_failed"
+                 write_event({"job_id": job_id, "stage": failed_stage, "gpu": gpu, "ts": _now()})
+                 record_failure(job_id, failed_stage)
+@@ -311,48 +329,52 @@ def run_jobs(
+     def upload_worker() -> None:
+         while True:
+             task = upload_queue.get()
+-            if task is None:
+-                upload_queue.task_done()
+-                break
+-            job = task.job
+-            job_id = str(job["job_id"])
+-            train_repo_id = str(job["train_upload_repo_id"])
+-            eval_repo_id = str(job["eval_upload_repo_id"])
+-            if dry_run:
+-                targets = [repo_id for repo_id in (eval_repo_id, train_repo_id) if repo_id]
+-                print(f"[dry-run] upload {job_id}: {', '.join(targets) or 'no upload target'}", flush=True)
+-                upload_queue.task_done()
+-                continue
+-            if not train_repo_id and not eval_repo_id:
+-                write_event({"job_id": job_id, "stage": "upload_skipped", "ts": _now()})
+-                write_event({"job_id": job_id, "stage": "job_complete", "ts": _now()})
+-                upload_queue.task_done()
+-                continue
+-            attempt = 0
+-            success = False
+-            while attempt < max_upload_attempts and not success:
+-                attempt += 1
+-                write_event({"job_id": job_id, "stage": "upload_started", "attempt": attempt, "ts": _now()})
+-                try:
+-                    if not dry_run:
+-                        uploader_fn(job)
+-                    write_event({"job_id": job_id, "stage": "upload_succeeded", "attempt": attempt, "ts": _now()})
++            try:
++                if task is None:
++                    break
++                if cancel_event.is_set():
++                    continue
++                job = task.job
++                job_id = str(job["job_id"])
++                train_repo_id = str(job["train_upload_repo_id"])
++                eval_repo_id = str(job["eval_upload_repo_id"])
++                if dry_run:
++                    targets = [repo_id for repo_id in (eval_repo_id, train_repo_id) if repo_id]
++                    print(f"[dry-run] upload {job_id}: {', '.join(targets) or 'no upload target'}", flush=True)
++                    continue
++                if not train_repo_id and not eval_repo_id:
++                    write_event({"job_id": job_id, "stage": "upload_skipped", "ts": _now()})
+                     write_event({"job_id": job_id, "stage": "job_complete", "ts": _now()})
+-                    success = True
+-                except Exception as error:
+-                    write_event(
+-                        {
+-                            "job_id": job_id,
+-                            "stage": "upload_failed",
+-                            "attempt": attempt,
+-                            "error": repr(error),
+-                            "ts": _now(),
+-                        }
+-                    )
+-                    print(f"upload failed for {job_id} (attempt {attempt}): {error!r}", file=sys.stderr)
+-            if not success:
+-                record_failure(job_id, "upload_failed")
+-            upload_queue.task_done()
++                    continue
++                attempt = 0
++                success = False
++                while attempt < max_upload_attempts and not success and not cancel_event.is_set():
++                    attempt += 1
++                    write_event({"job_id": job_id, "stage": "upload_started", "attempt": attempt, "ts": _now()})
++                    try:
++                        uploader_fn(job)
++                        if cancel_event.is_set():
++                            continue
++                        write_event({"job_id": job_id, "stage": "upload_succeeded", "attempt": attempt, "ts": _now()})
++                        write_event({"job_id": job_id, "stage": "job_complete", "ts": _now()})
++                        success = True
++                    except Exception as error:
++                        if cancel_event.is_set():
++                            continue
++                        write_event(
++                            {
++                                "job_id": job_id,
++                                "stage": "upload_failed",
++                                "attempt": attempt,
++                                "error": repr(error),
++                                "ts": _now(),
++                            }
++                        )
++                        print(f"upload failed for {job_id} (attempt {attempt}): {error!r}", file=sys.stderr)
++                if not success and not cancel_event.is_set():
++                    record_failure(job_id, "upload_failed")
++            finally:
++                upload_queue.task_done()
+ 
+     upload_threads = [
+         threading.Thread(target=upload_worker, name=f"upload-{idx}", daemon=True)
+@@ -362,17 +384,44 @@ def run_jobs(
+         threading.Thread(target=train_worker, args=(gpu,), name=f"train-{gpu}", daemon=True)
+         for gpu in gpus if stage != "upload"
+     ]
++    previous_sigint = signal.getsignal(signal.SIGINT)
++    previous_sigterm = signal.getsignal(signal.SIGTERM)
++
++    def stop_active_processes(signum, frame):
++        nonlocal cleanup_started
++        del frame
++        if signal_event.is_set():
++            return
++        signal_event.set()
++        cancel_event.set()
++        if cleanup_started:
++            return
++        raise KeyboardInterrupt(signum)
++
++    signal.signal(signal.SIGINT, stop_active_processes)
++    signal.signal(signal.SIGTERM, stop_active_processes)
+     try:
+         for thread in upload_threads + train_threads:
+             thread.start()
+         job_queue.join()
+-        for _ in upload_threads:
+-            upload_queue.put(None)
+         upload_queue.join()
+-        for thread in train_threads + upload_threads:
+-            thread.join()
+         return 1 if failures else 0
+     finally:
++        cleanup_started = True
++        cancel_event.set()
++        _terminate_active_processes(signal.SIGTERM)
++        for thread in train_threads:
++            thread.join()
++        _terminate_active_processes(signal.SIGTERM)
++        if signal_event.is_set():
++            for _ in upload_threads:
++                upload_queue.put(None)
++        else:
++            upload_queue.join()
++            for _ in upload_threads:
++                upload_queue.put(None)
++            for thread in upload_threads:
++                thread.join()
+         _terminate_active_processes(signal.SIGTERM)
+         signal.signal(signal.SIGINT, previous_sigint)
+         signal.signal(signal.SIGTERM, previous_sigterm)
+diff --git a/scripts/rollout_data/prepare_starvla_lerobot_dataset.py b/scripts/rollout_data/prepare_starvla_lerobot_dataset.py
+index e202657..2521899 100644
+--- a/scripts/rollout_data/prepare_starvla_lerobot_dataset.py
++++ b/scripts/rollout_data/prepare_starvla_lerobot_dataset.py
+@@ -26,6 +26,7 @@ def convert_raw_export_to_lerobot(
+     raw_export_dir: Path,
+     lerobot_root: Path,
+     data_mix: str,
++    action_carrier: str = "native",
+ ) -> dict:
+     """Convert a raw rollout export into a named LeRobot data mix.
+ 
+@@ -33,6 +34,7 @@ def convert_raw_export_to_lerobot(
+         raw_export_dir: Canonical rollout export directory.
+         lerobot_root: Directory that will contain the LeRobot datasets.
+         data_mix: Train dataset directory name under ``lerobot_root``.
++        action_carrier: Target action/state carrier.
+ 
+     Returns:
+         Conversion manifest returned by the LeRobot adapter.
+@@ -41,6 +43,7 @@ def convert_raw_export_to_lerobot(
+     return convert_canonical_rollout_to_lerobot(
+         raw_export_dir,
+         output_dir,
++        action_carrier=action_carrier,
+         force=True,
+     )
+ 
+@@ -92,6 +95,7 @@ def build_parser() -> argparse.ArgumentParser:
+     parser.add_argument("raw_export_dir")
+     parser.add_argument("lerobot_root")
+     parser.add_argument("data_mix")
++    parser.add_argument("--action-carrier", choices=("native", "bridge"), default="native")
+     parser.add_argument("--push-to-hub", default="")
+     parser.add_argument("--private", action="store_true")
+     return parser
+@@ -115,6 +119,7 @@ def main(argv: Sequence[str] | None = None) -> int:
+         raw_export_dir=raw_export_dir,
+         lerobot_root=lerobot_root,
+         data_mix=data_mix,
++        action_carrier=args.action_carrier,
+     )
+     if args.push_to_hub:
+         push_lerobot_dataset(
+diff --git a/scripts/sweeps/generate_sweep_configs.py b/scripts/sweeps/generate_sweep_configs.py
+index 8074c79..bd5fafe 100644
+--- a/scripts/sweeps/generate_sweep_configs.py
++++ b/scripts/sweeps/generate_sweep_configs.py
+@@ -116,6 +116,7 @@ ENV_KEYS_BY_NAME = {
+         "obs_fps",
+         "screen_size",
+         "noop_max",
++        "reset_mode",
+         "frame_stack",
+         "terminal_on_life_loss",
+         "clip_reward",
+diff --git a/tests/data/test_rollout_dataset_schema.py b/tests/data/test_rollout_dataset_schema.py
+index 06a6091..55ccebf 100644
+--- a/tests/data/test_rollout_dataset_schema.py
++++ b/tests/data/test_rollout_dataset_schema.py
+@@ -8,6 +8,7 @@ def test_canonical_rollout_fields_exclude_legacy_aliases():
+         "image",
+         "env_name",
+         "episode_idx",
++        "trajectory_idx",
+         "decision_step",
+         "split",
+         "seed",
+@@ -36,6 +37,7 @@ def test_rollout_features_include_env_specific_action_fields():
+     assert "action_tuple" not in deadly_features
+     assert "action" not in deadly_features
+     for features in (flappy_features, deadly_features):
++        assert "trajectory_idx" in features
+         assert "decision_step" in features
+         assert "latency_raw_frames" in features
+         assert "raw_reward" in features
+diff --git a/tests/data/test_rollout_hf_parquet.py b/tests/data/test_rollout_hf_parquet.py
+index 9fa4c48..4e56de6 100644
+--- a/tests/data/test_rollout_hf_parquet.py
++++ b/tests/data/test_rollout_hf_parquet.py
+@@ -15,6 +15,7 @@ def test_hf_parquet_writes_canonical_columns_only(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -56,6 +57,7 @@ def test_hf_parquet_shards_sort_by_decision_step(tmp_path):
+             "image": str(image_path),
+             "env_name": "flappy",
+             "episode_idx": 0,
++            "trajectory_idx": 0,
+             "decision_step": decision_step,
+             "split": "train",
+             "seed": 11,
+@@ -92,6 +94,7 @@ def test_hf_parquet_embeds_context_images_when_rows_provide_them(tmp_path):
+         "context_images": [str(path) for path in context_paths],
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -116,3 +119,18 @@ def test_hf_parquet_embeds_context_images_when_rows_provide_them(tmp_path):
+     assert "context_images" in table.column_names
+     assert len(exported["context_images"]) == 3
+     assert exported["context_images"][0]["bytes"].startswith(b"\x89PNG\r\n\x1a\n")
++
++
++def test_completed_episode_shards_require_trajectory_identity(tmp_path):
++    import json
++
++    import pytest
++
++    from latency_bench.data.hf_parquet import completed_episode_indices_from_shards
++
++    shard_path = tmp_path / "episode_rows" / "train" / "episode_000000.jsonl"
++    shard_path.parent.mkdir(parents=True)
++    shard_path.write_text(json.dumps({"episode_idx": 0}) + "\n", encoding="utf-8")
++
++    with pytest.raises(KeyError, match="trajectory_idx"):
++        completed_episode_indices_from_shards(tmp_path)
+diff --git a/tests/data/test_sf_teacher_dataset_export.py b/tests/data/test_sf_teacher_dataset_export.py
+index 62b30a6..ba4644b 100644
+--- a/tests/data/test_sf_teacher_dataset_export.py
++++ b/tests/data/test_sf_teacher_dataset_export.py
+@@ -1299,7 +1299,7 @@ def test_write_filtered_episode_images_preserves_empty_filtered_episode(tmp_path
+ 
+     image_writer = RolloutImageWriter(max_workers=0, row_buffer_size=1, export_image_size=None)
+ 
+-    rows = _write_filtered_episode_images(image_writer, [], {}, context_window=5)
++    rows = _write_filtered_episode_images(image_writer, [], {}, {}, context_window=5)
+ 
+     assert rows == []
+ 
+@@ -1314,6 +1314,7 @@ def test_write_filtered_episode_images_resets_context_after_decision_step_gap(tm
+         {
+             "image": str(tmp_path / "images" / "train" / "episode_000000" / f"decision_step_{step:06d}.png"),
+             "decision_step": step,
++            "trajectory_idx": 0,
+         }
+         for step in [0, 1, 4, 5]
+     ]
+@@ -1326,6 +1327,7 @@ def test_write_filtered_episode_images_resets_context_after_decision_step_gap(tm
+         image_writer,
+         rows,
+         frames_by_decision_step,
++        {step: 0 for step in frames_by_decision_step},
+         context_window=4,
+     )
+ 
+@@ -1733,13 +1735,24 @@ def test_collect_teacher_dataset_demon_attack_exports_full_game_across_life_loss
+         image_writer_workers=0,
+         row_buffer_size=1,
+         filter_preset="none",
++        context_window=3,
+     )
+     train_rows = _read_shard_rows(output_dir, "train")
+     state = json.loads((output_dir / "probe_state.json").read_text(encoding="utf-8"))
+ 
+     assert ScriptedDemonEnv.reset_seeds == [100, None, None, 100, None, None]
+-    assert [row["decision_step"] for row in train_rows] == [2]
+-    assert train_rows[0]["raw_reward"] == 30.0
++    assert [row["decision_step"] for row in train_rows] == [0, 1, 2]
++    assert [row["trajectory_idx"] for row in train_rows] == [0, 1, 2]
++    assert [row["action_id"] for row in train_rows] == [1, 1, 1]
++    assert [row["raw_reward"] for row in train_rows] == [10.0, 20.0, 30.0]
++    from PIL import Image
++
++    context_values = [
++        [np.asarray(Image.open(path))[0, 0, 0].item() for path in row["context_images"]]
++        for row in train_rows
++    ]
++    current_values = [np.asarray(Image.open(row["image"]))[0, 0, 0].item() for row in train_rows]
++    assert context_values == [[value, value] for value in current_values]
+     assert state["accepted_specs"] == [
+         {
+             "attempt_idx": 0,
+@@ -1900,6 +1913,7 @@ def test_collect_teacher_dataset_skips_runtime_when_probe_and_shards_complete(mo
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 100,
+@@ -2264,6 +2278,7 @@ def test_write_parquet_splits_handles_empty_validation_rows(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -2300,6 +2315,7 @@ def test_write_parquet_splits_embeds_image_bytes(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -2336,6 +2352,7 @@ def test_write_parquet_splits_preserves_rows_across_large_exports(tmp_path):
+             "image": str(image_path),
+             "env_name": "flappy",
+             "episode_idx": episode_idx,
++            "trajectory_idx": 0,
+             "decision_step": 0,
+             "split": "train",
+             "seed": 11 + episode_idx,
+@@ -2390,6 +2407,7 @@ def  test_write_parquet_splits_from_shards_writes_train_and_val_rows(tmp_path):
+                 "image": str(image_path),
+                 "env_name": "flappy",
+                 "episode_idx": episode_idx,
++                "trajectory_idx": 0,
+                 "decision_step": decision_step,
+                 "split": split,
+                 "seed": 11 + episode_idx,
+@@ -2432,6 +2450,7 @@ def test_write_parquet_splits_from_shards_replaces_stale_partial_file(tmp_path):
+         "image": str(stale_image_path),
+         "env_name": "flappy",
+         "episode_idx": 99,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 99,
+@@ -2452,6 +2471,7 @@ def test_write_parquet_splits_from_shards_replaces_stale_partial_file(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -2536,6 +2556,8 @@ def test_prepare_starvla_lerobot_dataset_converts_from_raw_export(monkeypatch, t
+         str(raw_export),
+         str(tmp_path / "lerobot"),
+         "demon_attack_train",
++        "--action-carrier",
++        "bridge",
+     ])
+ 
+     assert result == 0
+@@ -2543,7 +2565,7 @@ def test_prepare_starvla_lerobot_dataset_converts_from_raw_export(monkeypatch, t
+         (
+             raw_export,
+             tmp_path / "lerobot" / "demon_attack_train",
+-            {"force": True},
++            {"action_carrier": "bridge", "force": True},
+         )
+     ]
+ 
+diff --git a/tests/data/test_starvla_lerobot_exporter.py b/tests/data/test_starvla_lerobot_exporter.py
+index 64b5f47..58391c7 100644
+--- a/tests/data/test_starvla_lerobot_exporter.py
++++ b/tests/data/test_starvla_lerobot_exporter.py
+@@ -3,6 +3,7 @@ from __future__ import annotations
+ import json
+ 
+ import pyarrow.parquet as pq
++import pytest
+ 
+ from helpers import save_test_image
+ 
+@@ -20,6 +21,7 @@ def test_starvla_lerobot_exporter_reads_canonical_rollout_schema(tmp_path):
+         "image": str(image_path),
+         "env_name": "deadly_corridor",
+         "episode_idx": 3,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 14,
+@@ -38,6 +40,7 @@ def test_starvla_lerobot_exporter_reads_canonical_rollout_schema(tmp_path):
+                 "action_layout": "deadly_corridor_joint_54",
+                 "reward_field": "raw_reward",
+                 "action_labels": list(action_spec.labels),
++                "obs_fps": 8.75,
+             }
+         ),
+         encoding="utf-8",
+@@ -54,10 +57,102 @@ def test_starvla_lerobot_exporter_reads_canonical_rollout_schema(tmp_path):
+     assert table.to_pylist()[0]["done"] is True
+ 
+ 
+-def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+-    """Adjacent LeRobot rows must be adjacent obs decisions (every obs_stride raw
+-    frames), with contiguous frame_index, so KV-memory rollouts unroll at the eval
+-    observation cadence rather than at every raw frame."""
++@pytest.mark.parametrize(
++    ("env_name", "obs_fps", "latency", "action_values", "expected_action", "active_action_dim"),
++    [
++        ("flappy", 30, 3, (1, [0.0, 1.0], "FLAP"), [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0], 2),
++        (
++            "demon_attack",
++            15,
++            6,
++            (5, [0.0, 0.0, 0.0, 0.0, 0.0, 1.0], "LEFTFIRE"),
++            [0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0],
++            6,
++        ),
++        (
++            "deadly_corridor",
++            8.75,
++            6,
++            (11, None, "MOVE_FORWARD + MOVE_RIGHT + ATTACK"),
++            [1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 1.0],
++            7,
++        ),
++    ],
++)
++def test_starvla_lerobot_exporter_writes_bridge_carrier_with_raw_latency_prompt(
++    tmp_path,
++    env_name,
++    obs_fps,
++    latency,
++    action_values,
++    expected_action,
++    active_action_dim,
++):
++    from latency_bench.data.action_layouts import action_spec_for_env
++    from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
++    from latency_bench.data.hf_parquet import write_parquet_splits
++
++    image_path = tmp_path / "image.png"
++    save_test_image(image_path, 9)
++    raw_export = tmp_path / "raw"
++    action_spec = action_spec_for_env(env_name)
++    action_id, action, action_text = action_values
++    row = {
++        "image": str(image_path),
++        "env_name": env_name,
++        "episode_idx": 0,
++        "trajectory_idx": 0,
++        "decision_step": 0,
++        "split": "train",
++        "seed": 0,
++        "prompt": f"The action latency is fixed at {latency} raw environment frames.",
++        "latency_raw_frames": latency,
++        "latency_ms": latency * 1000 / {"flappy": 30, "demon_attack": 60, "deadly_corridor": 35}[env_name],
++        "raw_reward": 1.0,
++        "action_id": action_id,
++        "action_text": action_text,
++    }
++    if action is not None:
++        row["action"] = action
++    write_parquet_splits(raw_export, [row], [], action_spec=action_spec)
++    (raw_export / "metadata.json").write_text(
++        json.dumps(
++            {
++                "env_name": env_name,
++                "action_layout": action_spec.layout,
++                "reward_field": action_spec.reward_field,
++                "action_labels": list(action_spec.labels),
++                "obs_fps": obs_fps,
++            }
++        ),
++        encoding="utf-8",
++    )
++
++    manifest = convert_canonical_rollout_to_lerobot(
++        raw_export,
++        tmp_path / "lerobot",
++        action_carrier="bridge",
++        force=True,
++    )
++
++    table = pq.read_table(tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000000.parquet")
++    info = json.loads((tmp_path / "lerobot" / "meta" / "info.json").read_text(encoding="utf-8"))
++    task_lines = (tmp_path / "lerobot" / "meta" / "tasks.jsonl").read_text(
++        encoding="utf-8"
++    ).splitlines()
++    tasks = [json.loads(line) for line in task_lines]
++    converted_row = table.to_pylist()[0]
++    assert converted_row["action"] == expected_action
++    assert converted_row["observation.state"] == [0.0] * 7
++    assert info["features"]["action"]["shape"] == [7]
++    assert info["features"]["observation.state"]["shape"] == [7]
++    assert tasks == [{"task_index": 0, "task": row["prompt"]}]
++    assert manifest["action_carrier"] == "bridge"
++    assert manifest["active_action_dim"] == active_action_dim
++
++
++def test_starvla_lerobot_exporter_keeps_every_observation_decision(tmp_path):
++    """Canonical rows already occur at the policy observation cadence."""
+     from latency_bench.data.action_layouts import action_spec_for_env
+     from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
+     from latency_bench.data.hf_parquet import write_parquet_splits
+@@ -66,12 +161,13 @@ def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+     save_test_image(image_path, 9)
+     raw_export = tmp_path / "raw"
+     action_spec = action_spec_for_env("deadly_corridor")
+-    # One episode, 8 contiguous per-raw-frame decision steps; obs_stride = 4.
++    # One trajectory with eight contiguous policy-decision observations.
+     train_rows = [
+         {
+             "image": str(image_path),
+             "env_name": "deadly_corridor",
+             "episode_idx": 0,
++            "trajectory_idx": 0,
+             "decision_step": step,
+             "split": "train",
+             "seed": 7,
+@@ -92,6 +188,7 @@ def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+                 "action_layout": "deadly_corridor_joint_54",
+                 "reward_field": "raw_reward",
+                 "action_labels": list(action_spec.labels),
++                "obs_fps": 8.75,
+                 "obs_stride_raw_frames": 4,
+             }
+         ),
+@@ -102,12 +199,73 @@ def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+ 
+     table = pq.read_table(tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000000.parquet")
+     rows = table.to_pylist()
+-    # Only decision_steps 0 and 4 survive; frame_index is contiguous obs index.
+-    assert [r["frame_index"] for r in rows] == [0, 1]
+-    assert [r["episode_index"] for r in rows] == [0, 0]
++    assert [r["frame_index"] for r in rows] == list(range(8))
++    assert [r["episode_index"] for r in rows] == [0] * 8
+     assert rows[-1]["done"] is True
+ 
+ 
++def test_starvla_lerobot_exporter_splits_life_trajectories(tmp_path):
++    from latency_bench.data.action_layouts import action_spec_for_env
++    from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
++    from latency_bench.data.hf_parquet import write_parquet_splits
++
++    image_path = tmp_path / "image.png"
++    save_test_image(image_path, 9)
++    raw_export = tmp_path / "raw"
++    action_spec = action_spec_for_env("demon_attack")
++    train_rows = [
++        {
++            "image": str(image_path),
++            "env_name": "demon_attack",
++            "episode_idx": 0,
++            "trajectory_idx": trajectory_idx,
++            "decision_step": decision_step,
++            "split": "train",
++            "seed": 7,
++            "prompt": "pick action",
++            "latency_raw_frames": 6,
++            "latency_ms": 100.0,
++            "raw_reward": 1.0,
++            "action_id": 0,
++            "action": [1.0, 0.0, 0.0, 0.0, 0.0, 0.0],
++            "action_text": "NOOP",
++        }
++        for trajectory_idx, decision_step in ((0, 0), (0, 1), (1, 2), (1, 3))
++    ]
++    write_parquet_splits(raw_export, train_rows, [], action_spec=action_spec)
++    (raw_export / "metadata.json").write_text(
++        json.dumps(
++            {
++                "env_name": "demon_attack",
++                "action_layout": action_spec.layout,
++                "reward_field": action_spec.reward_field,
++                "action_labels": list(action_spec.labels),
++                "obs_fps": 15,
++            }
++        ),
++        encoding="utf-8",
++    )
++
++    manifest = convert_canonical_rollout_to_lerobot(
++        raw_export,
++        tmp_path / "lerobot",
++        action_carrier="bridge",
++        force=True,
++    )
++
++    first = pq.read_table(
++        tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000000.parquet"
++    ).to_pylist()
++    second = pq.read_table(
++        tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000001.parquet"
++    ).to_pylist()
++    assert manifest["episodes"] == 2
++    assert [row["frame_index"] for row in first] == [0, 1]
++    assert [row["frame_index"] for row in second] == [0, 1]
++    assert [row["done"] for row in first] == [False, True]
++    assert [row["done"] for row in second] == [False, True]
++
++
+ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+     from datasets import Dataset, Features, Image as HFImage, Sequence as HFSequence, Value
+     from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
+@@ -121,6 +279,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+             "image": HFImage(),
+             "env_name": Value("string"),
+             "episode_idx": Value("int64"),
++            "trajectory_idx": Value("int64"),
+             "decision_step": Value("int64"),
+             "split": Value("string"),
+             "seed": Value("int64"),
+@@ -138,6 +297,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+             "image": [str(image_path)],
+             "env_name": ["flappy"],
+             "episode_idx": [0],
++            "trajectory_idx": [0],
+             "decision_step": [0],
+             "split": ["train"],
+             "seed": [11],
+@@ -156,6 +316,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+             "image": [str(image_path)],
+             "env_name": ["flappy"],
+             "episode_idx": [1],
++            "trajectory_idx": [0],
+             "decision_step": [0],
+             "split": ["val"],
+             "seed": [12],
+@@ -176,6 +337,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+                 "action_layout": "flappy_discrete",
+                 "reward_field": "raw_reward",
+                 "action_labels": ["NOOP", "FLAP"],
++                "obs_fps": 30,
+             }
+         ),
+         encoding="utf-8",
+diff --git a/tests/scripts/test_generate_sweep_configs.py b/tests/scripts/test_generate_sweep_configs.py
+index 23d79cb..438b296 100644
+--- a/tests/scripts/test_generate_sweep_configs.py
++++ b/tests/scripts/test_generate_sweep_configs.py
+@@ -23,6 +23,75 @@ def _read_manifest(path: Path) -> list[dict[str, str]]:
+     return [dict(zip(header, line.split("\t"))) for line in lines[1:]]
+ 
+ 
++@pytest.mark.parametrize(
++    ("env_name", "latency_label", "latency_frames", "obs_fps", "obs_fps_label"),
++    [
++        ("flappy", "fixed_l3", 3, 30, "30"),
++        ("demon_attack", "fixed_l6", 6, 15, "15"),
++        ("deadly_corridor", "fixed_l6", 6, 8.75, "8p75"),
++    ],
++)
++def test_memory_rollout_teacher_spec_emits_single_uploadable_job(
++    tmp_path,
++    env_name,
++    latency_label,
++    latency_frames,
++    obs_fps,
++    obs_fps_label,
++):
++    generated_root = tmp_path / "configs"
++    output_root = tmp_path / "outputs"
++    train_dir = tmp_path / "checkpoints"
++
++    generate_sweep_configs(
++        [
++            "--spec",
++            f"configs/experiments/memory/rollout_teacher/{env_name}.yaml",
++            "--generated-root",
++            str(generated_root),
++            "--output-root",
++            str(output_root),
++            "--train-dir",
++            str(train_dir),
++        ]
++    )
++
++    rows = _read_manifest(generated_root / "manifest.tsv")
++    assert len(rows) == 1
++    row = rows[0]
++    assert row["env"] == env_name
++    assert row["experiment_family"] == "memory_rollout_teacher"
++    assert row["sweep_name"] == "teacher"
++    assert row["seed"] == "0"
++    assert row["frame_stack"] == "4"
++    assert row["obs_fps"] == obs_fps_label
++    assert row["latency_label"] == latency_label
++    assert row["latency_kind"] == "fixed"
++    assert row["latency_fixed_frames"] == str(latency_frames)
++    assert row["train_upload_repo_id"] == "latency-sensitive-bench/paper-experiment-models"
++    assert row["train_upload_repo_type"] == "model"
++    assert row["eval_upload_repo_id"] == ""
++    assert row["train_artifact_repo_path"].startswith(
++        f"memory/rollout_teachers/{env_name}/train/"
++    )
++
++    train_cfg = yaml.safe_load(Path(row["train_config"]).read_text(encoding="utf-8"))
++    assert train_cfg["experiment"]["seed"] == 0
++    assert train_cfg["env"]["frame_stack"] == 4
++    assert train_cfg["env"]["obs_fps"] == pytest.approx(obs_fps)
++    if env_name == "demon_attack":
++        assert train_cfg["env"]["reset_mode"] == "fire"
++    assert train_cfg["latency"]["method"] == "fixed"
++    assert train_cfg["latency"]["fixed_latency_ms"] == pytest.approx(
++        latency_frames * 1000 / train_cfg["env"]["env_fps"]
++    )
++    assert train_cfg["evaluation"]["eval_suites"] == {
++        "fixed": [[latency_frames]],
++        "normal": [],
++        "uniform": [],
++    }
++
++
+ def test_latency_bench_eval_config_disables_full_trace_records(tmp_path):
+     base_config = {
+         "backend": {
+diff --git a/tests/scripts/test_manifest_jobs.py b/tests/scripts/test_manifest_jobs.py
+index bd7bf03..74e4f73 100644
+--- a/tests/scripts/test_manifest_jobs.py
++++ b/tests/scripts/test_manifest_jobs.py
+@@ -1,8 +1,10 @@
+ from __future__ import annotations
+ 
++import json
+ import os
+-import subprocess
++import shlex
+ import signal
++import subprocess
+ import sys
+ import threading
+ import time
+@@ -10,6 +12,7 @@ import types
+ from pathlib import Path
+ 
+ import pytest
++from helpers import REPO_ROOT
+ 
+ import scripts.experiment_runner.run_manifest_jobs as run_manifest_jobs
+ from scripts.experiment_runner.job_manifest import (
+@@ -68,6 +71,26 @@ def test_train_stage_only_trains_and_resumes_from_success(tmp_path):
+     assert [event["stage"] for event in events] == ["train_started", "train_succeeded"]
+ 
+ 
++@pytest.mark.parametrize(
++    ("recorded_stage", "expected_stage"),
++    [
++        ("train_started", "train"),
++        ("train_succeeded", "eval"),
++        ("eval_started", "eval"),
++        ("eval_succeeded", "upload"),
++        ("upload_started", "upload"),
++    ],
++)
++def test_resume_stage_keeps_interrupted_job_at_its_current_pipeline_stage(
++    recorded_stage,
++    expected_stage,
++):
++    assert run_manifest_jobs._resume_stage(
++        {"job_id": "job1"},
++        {"job1": recorded_stage},
++    ) == expected_stage
++
++
+ def test_eval_stage_does_not_train_or_upload(tmp_path):
+     manifest_path, train_mark, eval_mark = _write_single_job(tmp_path)
+     uploads: list[str] = []
+@@ -134,6 +157,26 @@ def test_upload_stage_needs_no_gpu_and_retries_failed_upload(tmp_path):
+     ]
+ 
+ 
++def test_upload_stage_resumes_interrupted_upload(tmp_path):
++    manifest_path, _, _ = _write_single_job(tmp_path)
++    append_event(
++        event_log_path_for_manifest(manifest_path),
++        {"job_id": "job1", "stage": "upload_started", "ts": 1.0},
++    )
++    uploads: list[str] = []
++
++    assert run_jobs(
++        manifest_path=manifest_path,
++        gpus=[],
++        upload_workers=1,
++        max_upload_attempts=1,
++        stage="upload",
++        uploader=lambda job: uploads.append(job["job_id"]),
++    ) == 0
++
++    assert uploads == ["job1"]
++
++
+ def test_dry_run_does_not_execute_or_record_success(tmp_path, capsys):
+     manifest_path, train_mark, eval_mark = _write_single_job(tmp_path)
+ 
+@@ -310,6 +353,186 @@ def test_run_command_cleanup_terminates_child_process_group(tmp_path):
+         os.kill(child_pid, 0)
+ 
+ 
++def test_sigint_cancels_queue_and_kills_all_active_process_groups(tmp_path):
++    helper_path = tmp_path / "ignore_term_job.py"
++    helper_path.write_text(
++        """from __future__ import annotations
++
++import json
++import os
++from pathlib import Path
++import signal
++import subprocess
++import sys
++import time
++
++signal.signal(signal.SIGTERM, signal.SIG_IGN)
++child = subprocess.Popen(
++    [
++        sys.executable,
++        "-c",
++        "import signal, time; signal.signal(signal.SIGTERM, signal.SIG_IGN); time.sleep(60)",
++    ]
++)
++Path(sys.argv[1]).write_text(
++    json.dumps(
++        {
++            "leader_pid": os.getpid(),
++            "leader_pgid": os.getpgid(0),
++            "child_pid": child.pid,
++            "child_pgid": os.getpgid(child.pid),
++        }
++    ),
++    encoding="utf-8",
++)
++time.sleep(60)
++""",
++        encoding="utf-8",
++    )
++    manifest_path = tmp_path / "manifest.tsv"
++    markers = [tmp_path / f"job_{idx}.json" for idx in range(4)]
++    rows = []
++    for idx, marker in enumerate(markers):
++        row = {column: "" for column in MANIFEST_V2_COLUMNS}
++        row.update(
++            {
++                "schema_version": "v2",
++                "job_id": f"job{idx}",
++                "train_command": f"exec {shlex.join([sys.executable, str(helper_path), str(marker)])}",
++            }
++        )
++        rows.append(row)
++    write_manifest(manifest_path, rows)
++
++    runner = subprocess.Popen(
++        [
++            sys.executable,
++            "-m",
++            "scripts.experiment_runner.run_manifest_jobs",
++            "--manifest",
++            str(manifest_path),
++            "--stage",
++            "train",
++            "--gpus",
++            "0,1",
++        ],
++        cwd=REPO_ROOT,
++        stdout=subprocess.PIPE,
++        stderr=subprocess.PIPE,
++        text=True,
++    )
++    process_groups: set[int] = set()
++    try:
++        deadline = time.time() + 10.0
++        while sum(marker.is_file() for marker in markers) < 2 and time.time() < deadline:
++            time.sleep(0.05)
++        assert [marker.is_file() for marker in markers] == [True, True, False, False]
++        for marker in markers[:2]:
++            process_info = json.loads(marker.read_text(encoding="utf-8"))
++            assert process_info["leader_pgid"] == process_info["child_pgid"]
++            process_groups.add(int(process_info["leader_pgid"]))
++
++        runner.send_signal(signal.SIGINT)
++        time.sleep(0.1)
++        runner.send_signal(signal.SIGINT)
++        stdout, stderr = runner.communicate(timeout=10.0)
++        assert runner.returncode != 0, (stdout, stderr)
++        assert [marker.is_file() for marker in markers] == [True, True, False, False]
++
++        deadline = time.time() + 5.0
++        while process_groups and time.time() < deadline:
++            for process_group in list(process_groups):
++                try:
++                    os.killpg(process_group, 0)
++                except ProcessLookupError:
++                    process_groups.remove(process_group)
++            time.sleep(0.05)
++        assert not process_groups
++
++        events = load_events(event_log_path_for_manifest(manifest_path))
++        assert [event["stage"] for event in events] == ["train_started", "train_started"]
++    finally:
++        if runner.poll() is None:
++            runner.kill()
++            runner.wait()
++        for process_group in process_groups:
++            try:
++                os.killpg(process_group, signal.SIGKILL)
++            except ProcessLookupError:
++                pass
++
++
++def test_sigint_does_not_wait_for_in_flight_upload(tmp_path):
++    manifest_path = tmp_path / "manifest.tsv"
++    row = {column: "" for column in MANIFEST_V2_COLUMNS}
++    row.update(
++        {
++            "schema_version": "v2",
++            "job_id": "job0",
++            "train_upload_repo_id": "repo-model",
++            "train_upload_repo_type": "model",
++        }
++    )
++    write_manifest(manifest_path, [row])
++    append_event(
++        event_log_path_for_manifest(manifest_path),
++        {"job_id": "job0", "stage": "eval_succeeded"},
++    )
++    upload_started = tmp_path / "upload_started"
++    helper_path = tmp_path / "blocking_upload_runner.py"
++    helper_path.write_text(
++        """from pathlib import Path
++import sys
++import time
++
++from scripts.experiment_runner.run_manifest_jobs import run_jobs
++
++manifest_path = Path(sys.argv[1])
++upload_started = Path(sys.argv[2])
++
++def blocking_uploader(job):
++    del job
++    upload_started.write_text("started", encoding="utf-8")
++    time.sleep(60)
++
++raise SystemExit(
++    run_jobs(
++        manifest_path=manifest_path,
++        gpus=[],
++        upload_workers=1,
++        max_upload_attempts=1,
++        stage="upload",
++        uploader=blocking_uploader,
++    )
++)
++""",
++        encoding="utf-8",
++    )
++    runner = subprocess.Popen(
++        [sys.executable, str(helper_path), str(manifest_path), str(upload_started)],
++        cwd=REPO_ROOT,
++        env={**os.environ, "PYTHONPATH": str(REPO_ROOT)},
++        stdout=subprocess.PIPE,
++        stderr=subprocess.PIPE,
++        text=True,
++    )
++    try:
++        deadline = time.time() + 10.0
++        while not upload_started.is_file() and time.time() < deadline:
++            time.sleep(0.05)
++        assert upload_started.is_file()
++
++        started_at = time.monotonic()
++        runner.send_signal(signal.SIGINT)
++        stdout, stderr = runner.communicate(timeout=5.0)
++        assert runner.returncode != 0, (stdout, stderr)
++        assert time.monotonic() - started_at < 3.0
++    finally:
++        if runner.poll() is None:
++            runner.kill()
++            runner.wait()
++
++
+ def test_default_uploader_prefers_env_token_over_dotenv(monkeypatch, tmp_path):
+     monkeypatch.chdir(tmp_path)
+     (tmp_path / ".env").write_text("HF_TOKEN=dotenv-token\n", encoding="utf-8")

evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 1440.0, "episode_return_env": 1440.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 42, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 1786, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 7141}
+{"episode_id": 1, "episode_return": 7070.0, "episode_return_env": 7070.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 43, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 2, "episode_return": 5170.0, "episode_return_env": 5170.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 44, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 3, "episode_return": 4970.0, "episode_return_env": 4970.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 45, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 4, "episode_return": 6300.0, "episode_return_env": 6300.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 46, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 5, "episode_return": 1600.0, "episode_return_env": 1600.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 47, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 2233, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 8930}
+{"episode_id": 6, "episode_return": 2560.0, "episode_return_env": 2560.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 48, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 2387, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 9548}
+{"episode_id": 7, "episode_return": 5920.0, "episode_return_env": 5920.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 49, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3390, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 13557}
+{"episode_id": 8, "episode_return": 5070.0, "episode_return_env": 5070.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 50, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 9, "episode_return": 5670.0, "episode_return_env": 5670.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 51, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 10, "episode_return": 3250.0, "episode_return_env": 3250.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 52, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 2804, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 11215}
+{"episode_id": 11, "episode_return": 5060.0, "episode_return_env": 5060.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 53, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3387, "num_dropped_actions": 1, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 13546}
+{"episode_id": 12, "episode_return": 4500.0, "episode_return_env": 4500.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 54, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 13, "episode_return": 4650.0, "episode_return_env": 4650.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 55, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 14, "episode_return": 2585.0, "episode_return_env": 2585.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 56, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 2358, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 9431}
+{"episode_id": 15, "episode_return": 5570.0, "episode_return_env": 5570.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 57, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3103, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 12412}
+{"episode_id": 16, "episode_return": 5300.0, "episode_return_env": 5300.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 58, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 17, "episode_return": 5450.0, "episode_return_env": 5450.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 59, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 18, "episode_return": 5030.0, "episode_return_env": 5030.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 60, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}
+{"episode_id": 19, "episode_return": 5010.0, "episode_return_env": 5010.0, "game_score": null, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_name": "starvla_wan_oft_demon_attack_raw6", "configured_latency_ms": 100.0, "env_fps": 60.0, "env_id": "ALE/DemonAttack-v5", "episode_seed": 61, "eval_latency_steps": 6, "frame_ms": 16.666666666666668, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 15.0, "output_dir": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_demon_attack_raw6", "suite_name": "fixed_6", "target_device": null, "train_latency_steps": 6}, "num_actions": 3600, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 14400}

evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/queue_eval_latency_6.json

@@ -0,0 +1,59 @@
+{
+  "checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt",
+  "configured_latency_ms": 100.0,
+  "eval_latency_steps": 6,
+  "experiment_name": "starvla_wan_oft_demon_attack_raw6",
+  "latency": 6,
+  "latency_type": "fixed_sweep",
+  "lengths": [
+    7141,
+    14400,
+    14400,
+    14400,
+    14400,
+    8930,
+    9548,
+    13557,
+    14400,
+    14400,
+    11215,
+    13546,
+    14400,
+    14400,
+    9431,
+    12412,
+    14400,
+    14400,
+    14400,
+    14400
+  ],
+  "mean_length": 12929.0,
+  "mean_return": 4608.75,
+  "returns": [
+    1440.0,
+    7070.0,
+    5170.0,
+    4970.0,
+    6300.0,
+    1600.0,
+    2560.0,
+    5920.0,
+    5070.0,
+    5670.0,
+    3250.0,
+    5060.0,
+    4500.0,
+    4650.0,
+    2585.0,
+    5570.0,
+    5300.0,
+    5450.0,
+    5030.0,
+    5010.0
+  ],
+  "seed": 42,
+  "std_return": 1488.6200614999113,
+  "suite_name": "fixed_6",
+  "timestamp_utc": "2026-07-31T03:38:14.217776+00:00",
+  "train_latency_steps": 6
+}

evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/queue_eval_results.jsonl

@@ -0,0 +1,2 @@
+{"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "configured_latency_ms": 100.0, "eval_latency_steps": 6, "experiment_name": "starvla_wan_oft_demon_attack_raw6", "latency": 6, "latency_type": "fixed_sweep", "lengths": [7141, 14400, 14400, 14400, 14400, 8930, 9548, 13557, 14400, 14400, 11215, 13546, 14400, 14400, 9431, 12412, 14400, 14400, 14400, 14400], "mean_length": 12929.0, "mean_return": 4608.75, "returns": [1440.0, 7070.0, 5170.0, 4970.0, 6300.0, 1600.0, 2560.0, 5920.0, 5070.0, 5670.0, 3250.0, 5060.0, 4500.0, 4650.0, 2585.0, 5570.0, 5300.0, 5450.0, 5030.0, 5010.0], "seed": 42, "std_return": 1488.6200614999113, "suite_name": "fixed_6", "timestamp_utc": "2026-07-31T03:38:07.619471+00:00", "train_latency_steps": 6}
+{"checkpoint_path": "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "configured_latency_ms": 100.0, "eval_latency_steps": 6, "experiment_name": "starvla_wan_oft_demon_attack_raw6", "latency": 6, "latency_type": "fixed_sweep", "lengths": [7141, 14400, 14400, 14400, 14400, 8930, 9548, 13557, 14400, 14400, 11215, 13546, 14400, 14400, 9431, 12412, 14400, 14400, 14400, 14400], "mean_length": 12929.0, "mean_return": 4608.75, "returns": [1440.0, 7070.0, 5170.0, 4970.0, 6300.0, 1600.0, 2560.0, 5920.0, 5070.0, 5670.0, 3250.0, 5060.0, 4500.0, 4650.0, 2585.0, 5570.0, 5300.0, 5450.0, 5030.0, 5010.0], "seed": 42, "std_return": 1488.6200614999113, "suite_name": "fixed_6", "timestamp_utc": "2026-07-31T03:38:14.217776+00:00", "train_latency_steps": 6}

evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/.hydra/hydra.yaml

@@ -0,0 +1,379 @@
+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=wan_oft
+    - env=demon_attack
+    - init=wan_oft_libero
+    - mode=single
+    - ++framework.qwenvl.base_vlm=playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    - ++framework.qwenvl.attn_implementation=sdpa
+    - ++framework.qwenvl.flex_backend=triton
+    - ++framework.qwenvl.enable_gradient_checkpointing=true
+    - ++framework.action_model.state_dim=7
+    - ++framework.action_model.loss_type=current_discrete_ce
+    - ++framework.action_model.action_horizon=8
+    - ++framework.action_model.future_action_window_size=7
+    - ++framework.action_model.past_action_window_size=0
+    - ++framework.action_model.action_model_type=MLP
+    - ++framework.action_model.action_dim=7
+    - ++framework.action_model.action_env_dim=6
+    - ++framework.action_model.action_query_source=mean
+    - ++framework.action_model.action_query_num_heads=24
+    - ++framework.kv_memory.enabled=false
+    - ++framework.kv_memory.window=4
+    - ++framework.kv_memory.rollout_len=8
+    - ++framework.kv_memory.packed_train=false
+    - ++framework.kv_memory.rebased_sink=true
+    - ++framework.name=WanOFT
+    - ++framework.world_model.base_wm=playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    - ++framework.world_model.extract_layers=[-1]
+    - ++framework.world_model.num_frames=5
+    - ++datasets.vla_data.dataset_py=lerobot_datasets
+    - ++datasets.vla_data.include_state=true
+    - ++datasets.vla_data.data_root_dir=playground/Datasets/rl_games
+    - ++datasets.vla_data.data_mix=demon_attack_train__bridge
+    - ++datasets.vla_data.eval_data_mix=null
+    - ++datasets.vla_data.custom_mixtures_path=null
+    - ++datasets.vla_data.action_type=discrete
+    - ++datasets.vla_data.sequential_step_sampling=false
+    - ++datasets.vla_data.eval_sequential_step_sampling=null
+    - ++datasets.vla_data.num_workers=8
+    - ++datasets.vla_data.eval_num_workers=8
+    - ++datasets.vla_data.prefetch_factor=4
+    - ++datasets.vla_data.persistent_workers=true
+    - ++datasets.vla_data.pin_memory=true
+    - ++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.latency_curriculum.latencies=null
+    - ++datasets.vla_data.latency_curriculum.phase_steps=null
+    - ++datasets.vla_data.latency_curriculum.phase_distributions=null
+    - ++datasets.vla_data.latency_curriculum.new_latency_passes=1.0
+    - ++datasets.vla_data.latency_curriculum.replay_passes=0.25
+    - ++datasets.vla_data.latency_curriculum.target_total_passes=2.0
+    - ++datasets.vla_data.latency_curriculum.final_equalization=true
+    - ++datasets.vla_data.latency_curriculum.step_budget_mode=auto
+    - ++datasets.vla_data.latency_curriculum.eval_at_phase_end=false
+    - ++datasets.vla_data.latency_curriculum.save_at_phase_end=false
+    - ++datasets.vla_data.latency_curriculum.computed_plan=null
+    - ++datasets.vla_data.per_device_batch_size=1
+    - ++datasets.vla_data.load_all_data_for_training=true
+    - ++datasets.vla_data.num_obs_frames=1
+    - ++datasets.vla_data.image_mode=single
+    - ++datasets.vla_data.prompt_mode=raw
+    - ++datasets.vla_data.stitch_grid=[2,2]
+    - ++datasets.vla_data.obs_image_size=[224,224]
+    - ++datasets.vla_data.video_backend=torchvision_av
+    - ++datasets.vla_data.observation_indices=[-4,-3,-2,-1,0]
+    - ++datasets.vla_data.language_indices=[0]
+    - ++datasets.vla_data.state_indices=[0]
+    - ++datasets.vla_data.action_indices=[0,1,2,3,4,5,6,7]
+    - ++datasets.vla_data.pack_image_sequence=true
+    - ++datasets.vla_data.image_sequence_length=5
+    - ++datasets.vla_data.context_images_column=observation.context_images
+    - ++dataset.source_hf=latency-sensitive-bench/memory-rollouts
+    - ++dataset.config_name=demon_attack_fixed_latency_6_200ep_7k2steps
+    - ++dataset.source_subdir=null
+    - ++dataset.converted_name=demon_attack_train__bridge
+    - ++dataset.single_source_hf=
+    - ++dataset.mixed_source_hf=
+    - ++dataset.single_converted_name=demon_attack_train__bridge
+    - ++dataset.mixed_converted_name=demon_attack_mixed_latency_train__bridge
+    - ++dataset.single_latency_filter=null
+    - ++dataset.mixed_latency_filter=null
+    - ++dataset.force_download=false
+    - ++dataset.setup_force=false
+    - ++dataset.skip_verification=false
+    - ++dataset.target_latency_unit=observation_steps
+    - ++dataset.verify_rows=200
+    - ++dataset.max_episodes=200
+    - ++dataset.episodes_per_latency=null
+    - ++dataset.latency_filter=null
+    - ++dataset.debug_subset.enabled=false
+    - ++dataset.debug_subset.max_episodes=5
+    - ++dataset.debug_subset.suffix=debug
+    - ++base_model.repo_id=Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    - ++initialization.checkpoint_local_dir=playground/Pretrained_models/WM4A-Wan2d2-OFT-LIBERO-4in1
+    - ++initialization.checkpoint_hf_repo_id=StarVLA/WM4A-Wan2d2-OFT-LIBERO-4in1
+    - ++initialization.checkpoint_filename=checkpoints/steps_60000_pytorch_model.pt
+    - ++trainer.max_train_steps=5000
+    - ++trainer.num_warmup_steps=0
+    - ++trainer.save_interval=500
+    - ++trainer.eval_interval=100
+    - ++trainer.eval_num_batches=100
+    - ++trainer.per_latency_eval_num_batches=null
+    - ++trainer.eval_action_classification=false
+    - ++trainer.eval_action_classification_interval=null
+    - ++trainer.cc_f1_tolerance=1
+    - ++trainer.learning_rate.base=2.5e-05
+    - ++trainer.learning_rate.qwen_vl_interface=1e-05
+    - ++trainer.learning_rate.action_model=0.0001
+    - ++trainer.learning_rate.action_query_proj=0.0001
+    - ++trainer.lr_scheduler_type=cosine_with_min_lr
+    - ++trainer.scheduler_specific_kwargs.min_lr=1e-06
+    - ++trainer.freeze_modules=
+    - ++trainer.freeze_vit=false
+    - ++trainer.freeze_tied_embedding=false
+    - ++trainer.freeze_llm_layers=[]
+    - ++trainer.loss_scale.vla=1.0
+    - ++trainer.loss_scale.vlm=0.1
+    - ++trainer.max_grad_norm=1.0
+    - ++trainer.weight_decay=0.0
+    - ++trainer.logging_frequency=1
+    - ++trainer.profile_timing.enabled=false
+    - ++trainer.profile_timing.log_interval=10
+    - ++trainer.gradient_clipping=1.0
+    - ++trainer.gradient_accumulation_steps=16
+    - ++trainer.distributed_backend=deepspeed
+    - ++trainer.is_resume=false
+    - ++trainer.pretrained_checkpoint=null
+    - ++trainer.resume_step=0
+    - ++trainer.reload_modules=null
+    - ++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=true
+    - ++trainer.save_format=pt
+    - ++workspace_dir=WORKSPACE_DIR
+    - ++run_root_dir=results/Checkpoints
+    - ++seed=42
+    - ++auth.env_file=null
+    - ++auth.hf_token_env=HF_TOKEN
+    - ++auth.wandb_api_key_env=WANDB_API_KEY
+    - ++paths.run_root_dir=results/Checkpoints
+    - ++paths.dataset_local_dir=data/demon_attack_fix_latency_6_200ep_context5
+    - ++paths.dataset_cache_dir=null
+    - ++paths.base_model_dir=playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    - ++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml
+    - ++rl_games.model_alias=wan_oft
+    - ++rl_games.env_eval.image_size=224
+    - ++rl_games.env_eval.frameskip=4
+    - ++rl_games.env_eval.image_transform=raw_rgb
+    - ++rl_games.env_eval.prompt_mode=raw
+    - ++rl_games.env_eval.ghost_trail.history_frames=5
+    - ++rl_games.env_eval.ghost_trail.gamma=1.3
+    - ++rl_games.env_eval.ghost_trail.min_alpha=35
+    - ++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0
+    - ++rl_games.env_eval.ghost_trail.ground_fraction=0.22
+    - ++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 Demon Attack from a
+      single game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT,
+      RIGHTFIRE, LEFTFIRE.'''
+    - ++rl_games.env_eval.eval_parallel_envs=5
+    - ++rl_games.env_eval.action_chunk_execution.enabled=false
+    - ++rl_games.env_eval.action_chunk_execution.chunk_size=null
+    - ++rl_games.env_eval.demon_attack.noop_max=30
+    - ++rl_games.env_eval.enabled=true
+    - ++rl_games.env_eval.eval_backend=eval_core
+    - ++rl_games.env_eval.distributed_mode=rank_sharded
+    - ++rl_games.env_eval.vectorized.enabled=false
+    - ++rl_games.env_eval.vectorized.batch_size=1
+    - ++rl_games.env_eval.latency.prompt_map_path=data/demon_attack_fix_latency_6_200ep_context5/demon_attack_train__bridge/latency_prompt_map.json
+    - ++rl_games.env_eval.latency.mode=single
+    - ++rl_games.env_eval.latency.values=[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.num_episodes=5
+    - ++rl_games.env_eval.mid_train.max_steps_per_episode=3600
+    - ++rl_games.env_eval.post_train.enabled=false
+    - ++rl_games.env_eval.post_train.latencies=[6]
+    - ++rl_games.env_eval.post_train.num_episodes=20
+    - ++rl_games.env_eval.post_train.max_steps_per_episode=3600
+    - ++rl_games.task=demon_attack
+    - ++rl_games.initialization_mode=bridge
+    - ++rl_games.action_carrier=bridge
+    - ++checkpoint.load=auto
+    - ++checkpoint.hf_repo_id=null
+    - ++checkpoint.save_best_model=false
+    - ++checkpoint.save_final_model=true
+    - ++checkpoint.save_pt_file=true
+    - ++checkpoint.save_training_state=true
+    - ++checkpoint.save_safetensors_file=false
+    - ++checkpoint.local.keep_last_n=1
+    - ++checkpoint.sync.enabled=false
+    - ++checkpoint.sync.repo_id=null
+    - ++checkpoint.sync.keep_last_n=0
+    - ++checkpoint.sync.sync_every_n_checkpoints=1
+    - ++checkpoint.sync.resume_policy=local_latest
+    - ++run_id=wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce
+    - ++output_dir=null
+    - ++config_yaml=null
+    - ++is_debug=false
+    - ++version_id=0.21
+    - ++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints
+    - ++trainer.is_resume=true
+    - ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_4000_state
+    - ++trainer.resume_step=4000
+    - ++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/data/demon_attack_fix_latency_6_200ep_context5
+    - ++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    - ++framework.world_model.base_wm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+  job:
+    name: train_starvla_hydra
+    chdir: false
+    override_dirname: '++auth.env_file=null,++auth.hf_token_env=HF_TOKEN,++auth.wandb_api_key_env=WANDB_API_KEY,++base_model.repo_id=Wan-AI/Wan2.2-TI2V-5B-Diffusers,++checkpoint.hf_repo_id=null,++checkpoint.load=auto,++checkpoint.local.keep_last_n=1,++checkpoint.save_best_model=false,++checkpoint.save_final_model=true,++checkpoint.save_pt_file=true,++checkpoint.save_safetensors_file=false,++checkpoint.save_training_state=true,++checkpoint.sync.enabled=false,++checkpoint.sync.keep_last_n=0,++checkpoint.sync.repo_id=null,++checkpoint.sync.resume_policy=local_latest,++checkpoint.sync.sync_every_n_checkpoints=1,++config_yaml=null,++dataset.config_name=demon_attack_fixed_latency_6_200ep_7k2steps,++dataset.converted_name=demon_attack_train__bridge,++dataset.debug_subset.enabled=false,++dataset.debug_subset.max_episodes=5,++dataset.debug_subset.suffix=debug,++dataset.episodes_per_latency=null,++dataset.force_download=false,++dataset.latency_filter=null,++dataset.max_episodes=200,++dataset.mixed_converted_name=demon_attack_mixed_latency_train__bridge,++dataset.mixed_latency_filter=null,++dataset.mixed_source_hf=,++dataset.setup_force=false,++dataset.single_converted_name=demon_attack_train__bridge,++dataset.single_latency_filter=null,++dataset.single_source_hf=,++dataset.skip_verification=false,++dataset.source_hf=latency-sensitive-bench/memory-rollouts,++dataset.source_subdir=null,++dataset.target_latency_unit=observation_steps,++dataset.verify_rows=200,++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_indices=[0,1,2,3,4,5,6,7],++datasets.vla_data.action_type=discrete,++datasets.vla_data.context_images_column=observation.context_images,++datasets.vla_data.custom_mixtures_path=null,++datasets.vla_data.data_mix=demon_attack_train__bridge,++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/data/demon_attack_fix_latency_6_200ep_context5,++datasets.vla_data.data_root_dir=playground/Datasets/rl_games,++datasets.vla_data.dataset_py=lerobot_datasets,++datasets.vla_data.eval_data_mix=null,++datasets.vla_data.eval_num_workers=8,++datasets.vla_data.eval_sequential_step_sampling=null,++datasets.vla_data.image_mode=single,++datasets.vla_data.image_sequence_length=5,++datasets.vla_data.include_state=true,++datasets.vla_data.language_indices=[0],++datasets.vla_data.latency_curriculum.computed_plan=null,++datasets.vla_data.latency_curriculum.enabled=false,++datasets.vla_data.latency_curriculum.eval_at_phase_end=false,++datasets.vla_data.latency_curriculum.final_equalization=true,++datasets.vla_data.latency_curriculum.latencies=null,++datasets.vla_data.latency_curriculum.new_latency_passes=1.0,++datasets.vla_data.latency_curriculum.phase_distributions=null,++datasets.vla_data.latency_curriculum.phase_steps=null,++datasets.vla_data.latency_curriculum.replay_passes=0.25,++datasets.vla_data.latency_curriculum.save_at_phase_end=false,++datasets.vla_data.latency_curriculum.step_budget_mode=auto,++datasets.vla_data.latency_curriculum.strategy=exclusive,++datasets.vla_data.latency_curriculum.target_total_passes=2.0,++datasets.vla_data.load_all_data_for_training=true,++datasets.vla_data.num_obs_frames=1,++datasets.vla_data.num_workers=8,++datasets.vla_data.obs_image_size=[224,224],++datasets.vla_data.observation_indices=[-4,-3,-2,-1,0],++datasets.vla_data.pack_image_sequence=true,++datasets.vla_data.per_device_batch_size=1,++datasets.vla_data.persistent_workers=true,++datasets.vla_data.pin_memory=true,++datasets.vla_data.prefetch_factor=4,++datasets.vla_data.prompt_mode=raw,++datasets.vla_data.sequential_step_sampling=false,++datasets.vla_data.shuffle=true,++datasets.vla_data.state_indices=[0],++datasets.vla_data.stitch_grid=[2,2],++datasets.vla_data.video_backend=torchvision_av,++framework.action_model.action_dim=7,++framework.action_model.action_env_dim=6,++framework.action_model.action_horizon=8,++framework.action_model.action_model_type=MLP,++framework.action_model.action_query_num_heads=24,++framework.action_model.action_query_source=mean,++framework.action_model.future_action_window_size=7,++framework.action_model.loss_type=current_discrete_ce,++framework.action_model.past_action_window_size=0,++framework.action_model.state_dim=7,++framework.kv_memory.enabled=false,++framework.kv_memory.packed_train=false,++framework.kv_memory.rebased_sink=true,++framework.kv_memory.rollout_len=8,++framework.kv_memory.window=4,++framework.name=WanOFT,++framework.qwenvl.attn_implementation=sdpa,++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers,++framework.qwenvl.base_vlm=playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers,++framework.qwenvl.enable_gradient_checkpointing=true,++framework.qwenvl.flex_backend=triton,++framework.world_model.base_wm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers,++framework.world_model.base_wm=playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers,++framework.world_model.extract_layers=[-1],++framework.world_model.num_frames=5,++initialization.checkpoint_filename=checkpoints/steps_60000_pytorch_model.pt,++initialization.checkpoint_hf_repo_id=StarVLA/WM4A-Wan2d2-OFT-LIBERO-4in1,++initialization.checkpoint_local_dir=playground/Pretrained_models/WM4A-Wan2d2-OFT-LIBERO-4in1,++is_debug=false,++output_dir=null,++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml,++paths.base_model_dir=playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers,++paths.dataset_cache_dir=null,++paths.dataset_local_dir=data/demon_attack_fix_latency_6_200ep_context5,++paths.run_root_dir=results/Checkpoints,++rl_games.action_carrier=bridge,++rl_games.env_eval.action_chunk_execution.chunk_size=null,++rl_games.env_eval.action_chunk_execution.enabled=false,++rl_games.env_eval.demon_attack.noop_max=30,++rl_games.env_eval.distributed_mode=rank_sharded,++rl_games.env_eval.enabled=true,++rl_games.env_eval.eval_backend=eval_core,++rl_games.env_eval.eval_parallel_envs=5,++rl_games.env_eval.fixed_episode_seeds=true,++rl_games.env_eval.frameskip=4,++rl_games.env_eval.ghost_trail.gamma=1.3,++rl_games.env_eval.ghost_trail.ground_fraction=0.22,++rl_games.env_eval.ghost_trail.history_frames=5,++rl_games.env_eval.ghost_trail.min_alpha=35,++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0,++rl_games.env_eval.image_size=224,++rl_games.env_eval.image_transform=raw_rgb,++rl_games.env_eval.latency.mode=single,++rl_games.env_eval.latency.prompt_map_path=data/demon_attack_fix_latency_6_200ep_context5/demon_attack_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=[6],++rl_games.env_eval.post_train.max_steps_per_episode=3600,++rl_games.env_eval.post_train.num_episodes=20,++rl_games.env_eval.prompt_mode=raw,++rl_games.env_eval.seed=42,++rl_games.env_eval.task_description=''You
+      are playing Demon Attack from a single game image. Choose exactly one action
+      from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE, LEFTFIRE.'',++rl_games.env_eval.task_seed_stride=0,++rl_games.env_eval.vectorized.batch_size=1,++rl_games.env_eval.vectorized.enabled=false,++rl_games.initialization_mode=bridge,++rl_games.model_alias=wan_oft,++rl_games.task=demon_attack,++run_id=wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce,++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints,++run_root_dir=results/Checkpoints,++seed=42,++trainer.cc_f1_tolerance=1,++trainer.distributed_backend=deepspeed,++trainer.eval_action_classification=false,++trainer.eval_action_classification_interval=null,++trainer.eval_interval=100,++trainer.eval_num_batches=100,++trainer.freeze_llm_layers=[],++trainer.freeze_modules=,++trainer.freeze_tied_embedding=false,++trainer.freeze_vit=false,++trainer.gradient_accumulation_steps=16,++trainer.gradient_clipping=1.0,++trainer.is_resume=false,++trainer.is_resume=true,++trainer.learning_rate.action_model=0.0001,++trainer.learning_rate.action_query_proj=0.0001,++trainer.learning_rate.base=2.5e-05,++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=5000,++trainer.num_warmup_steps=0,++trainer.optimizer.betas=[0.9,0.95],++trainer.optimizer.eps=1e-08,++trainer.optimizer.fused=true,++trainer.optimizer.name=AdamW,++trainer.optimizer.weight_decay=1e-08,++trainer.per_latency_eval_num_batches=null,++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_4000_state,++trainer.pretrained_checkpoint=null,++trainer.profile_timing.enabled=false,++trainer.profile_timing.log_interval=10,++trainer.reload_modules=null,++trainer.resume_step=0,++trainer.resume_step=4000,++trainer.save_format=pt,++trainer.save_interval=500,++trainer.scheduler_specific_kwargs.min_lr=1e-06,++trainer.weight_decay=0.0,++version_id=0.21,++workspace_dir=WORKSPACE_DIR,env=demon_attack,init=wan_oft_libero,mode=single,model=wan_oft'
+    id: ???
+    num: ???
+    config_name: train
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.4
+    version_base: '1.1'
+    cwd: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/examples/rl_games/config
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra
+    choices:
+      cross_task_setup: null
+      checkpoint: default
+      mode: single
+      init: wan_oft_libero
+      env: demon_attack
+      model: wan_oft
+      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

evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/.hydra/overrides.yaml

@@ -0,0 +1,217 @@
+- model=wan_oft
+- env=demon_attack
+- init=wan_oft_libero
+- mode=single
+- ++framework.qwenvl.base_vlm=playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+- ++framework.qwenvl.attn_implementation=sdpa
+- ++framework.qwenvl.flex_backend=triton
+- ++framework.qwenvl.enable_gradient_checkpointing=true
+- ++framework.action_model.state_dim=7
+- ++framework.action_model.loss_type=current_discrete_ce
+- ++framework.action_model.action_horizon=8
+- ++framework.action_model.future_action_window_size=7
+- ++framework.action_model.past_action_window_size=0
+- ++framework.action_model.action_model_type=MLP
+- ++framework.action_model.action_dim=7
+- ++framework.action_model.action_env_dim=6
+- ++framework.action_model.action_query_source=mean
+- ++framework.action_model.action_query_num_heads=24
+- ++framework.kv_memory.enabled=false
+- ++framework.kv_memory.window=4
+- ++framework.kv_memory.rollout_len=8
+- ++framework.kv_memory.packed_train=false
+- ++framework.kv_memory.rebased_sink=true
+- ++framework.name=WanOFT
+- ++framework.world_model.base_wm=playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+- ++framework.world_model.extract_layers=[-1]
+- ++framework.world_model.num_frames=5
+- ++datasets.vla_data.dataset_py=lerobot_datasets
+- ++datasets.vla_data.include_state=true
+- ++datasets.vla_data.data_root_dir=playground/Datasets/rl_games
+- ++datasets.vla_data.data_mix=demon_attack_train__bridge
+- ++datasets.vla_data.eval_data_mix=null
+- ++datasets.vla_data.custom_mixtures_path=null
+- ++datasets.vla_data.action_type=discrete
+- ++datasets.vla_data.sequential_step_sampling=false
+- ++datasets.vla_data.eval_sequential_step_sampling=null
+- ++datasets.vla_data.num_workers=8
+- ++datasets.vla_data.eval_num_workers=8
+- ++datasets.vla_data.prefetch_factor=4
+- ++datasets.vla_data.persistent_workers=true
+- ++datasets.vla_data.pin_memory=true
+- ++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.latency_curriculum.latencies=null
+- ++datasets.vla_data.latency_curriculum.phase_steps=null
+- ++datasets.vla_data.latency_curriculum.phase_distributions=null
+- ++datasets.vla_data.latency_curriculum.new_latency_passes=1.0
+- ++datasets.vla_data.latency_curriculum.replay_passes=0.25
+- ++datasets.vla_data.latency_curriculum.target_total_passes=2.0
+- ++datasets.vla_data.latency_curriculum.final_equalization=true
+- ++datasets.vla_data.latency_curriculum.step_budget_mode=auto
+- ++datasets.vla_data.latency_curriculum.eval_at_phase_end=false
+- ++datasets.vla_data.latency_curriculum.save_at_phase_end=false
+- ++datasets.vla_data.latency_curriculum.computed_plan=null
+- ++datasets.vla_data.per_device_batch_size=1
+- ++datasets.vla_data.load_all_data_for_training=true
+- ++datasets.vla_data.num_obs_frames=1
+- ++datasets.vla_data.image_mode=single
+- ++datasets.vla_data.prompt_mode=raw
+- ++datasets.vla_data.stitch_grid=[2,2]
+- ++datasets.vla_data.obs_image_size=[224,224]
+- ++datasets.vla_data.video_backend=torchvision_av
+- ++datasets.vla_data.observation_indices=[-4,-3,-2,-1,0]
+- ++datasets.vla_data.language_indices=[0]
+- ++datasets.vla_data.state_indices=[0]
+- ++datasets.vla_data.action_indices=[0,1,2,3,4,5,6,7]
+- ++datasets.vla_data.pack_image_sequence=true
+- ++datasets.vla_data.image_sequence_length=5
+- ++datasets.vla_data.context_images_column=observation.context_images
+- ++dataset.source_hf=latency-sensitive-bench/memory-rollouts
+- ++dataset.config_name=demon_attack_fixed_latency_6_200ep_7k2steps
+- ++dataset.source_subdir=null
+- ++dataset.converted_name=demon_attack_train__bridge
+- ++dataset.single_source_hf=
+- ++dataset.mixed_source_hf=
+- ++dataset.single_converted_name=demon_attack_train__bridge
+- ++dataset.mixed_converted_name=demon_attack_mixed_latency_train__bridge
+- ++dataset.single_latency_filter=null
+- ++dataset.mixed_latency_filter=null
+- ++dataset.force_download=false
+- ++dataset.setup_force=false
+- ++dataset.skip_verification=false
+- ++dataset.target_latency_unit=observation_steps
+- ++dataset.verify_rows=200
+- ++dataset.max_episodes=200
+- ++dataset.episodes_per_latency=null
+- ++dataset.latency_filter=null
+- ++dataset.debug_subset.enabled=false
+- ++dataset.debug_subset.max_episodes=5
+- ++dataset.debug_subset.suffix=debug
+- ++base_model.repo_id=Wan-AI/Wan2.2-TI2V-5B-Diffusers
+- ++initialization.checkpoint_local_dir=playground/Pretrained_models/WM4A-Wan2d2-OFT-LIBERO-4in1
+- ++initialization.checkpoint_hf_repo_id=StarVLA/WM4A-Wan2d2-OFT-LIBERO-4in1
+- ++initialization.checkpoint_filename=checkpoints/steps_60000_pytorch_model.pt
+- ++trainer.max_train_steps=5000
+- ++trainer.num_warmup_steps=0
+- ++trainer.save_interval=500
+- ++trainer.eval_interval=100
+- ++trainer.eval_num_batches=100
+- ++trainer.per_latency_eval_num_batches=null
+- ++trainer.eval_action_classification=false
+- ++trainer.eval_action_classification_interval=null
+- ++trainer.cc_f1_tolerance=1
+- ++trainer.learning_rate.base=2.5e-05
+- ++trainer.learning_rate.qwen_vl_interface=1e-05
+- ++trainer.learning_rate.action_model=0.0001
+- ++trainer.learning_rate.action_query_proj=0.0001
+- ++trainer.lr_scheduler_type=cosine_with_min_lr
+- ++trainer.scheduler_specific_kwargs.min_lr=1e-06
+- ++trainer.freeze_modules=
+- ++trainer.freeze_vit=false
+- ++trainer.freeze_tied_embedding=false
+- ++trainer.freeze_llm_layers=[]
+- ++trainer.loss_scale.vla=1.0
+- ++trainer.loss_scale.vlm=0.1
+- ++trainer.max_grad_norm=1.0
+- ++trainer.weight_decay=0.0
+- ++trainer.logging_frequency=1
+- ++trainer.profile_timing.enabled=false
+- ++trainer.profile_timing.log_interval=10
+- ++trainer.gradient_clipping=1.0
+- ++trainer.gradient_accumulation_steps=16
+- ++trainer.distributed_backend=deepspeed
+- ++trainer.is_resume=false
+- ++trainer.pretrained_checkpoint=null
+- ++trainer.resume_step=0
+- ++trainer.reload_modules=null
+- ++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=true
+- ++trainer.save_format=pt
+- ++workspace_dir=WORKSPACE_DIR
+- ++run_root_dir=results/Checkpoints
+- ++seed=42
+- ++auth.env_file=null
+- ++auth.hf_token_env=HF_TOKEN
+- ++auth.wandb_api_key_env=WANDB_API_KEY
+- ++paths.run_root_dir=results/Checkpoints
+- ++paths.dataset_local_dir=data/demon_attack_fix_latency_6_200ep_context5
+- ++paths.dataset_cache_dir=null
+- ++paths.base_model_dir=playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+- ++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml
+- ++rl_games.model_alias=wan_oft
+- ++rl_games.env_eval.image_size=224
+- ++rl_games.env_eval.frameskip=4
+- ++rl_games.env_eval.image_transform=raw_rgb
+- ++rl_games.env_eval.prompt_mode=raw
+- ++rl_games.env_eval.ghost_trail.history_frames=5
+- ++rl_games.env_eval.ghost_trail.gamma=1.3
+- ++rl_games.env_eval.ghost_trail.min_alpha=35
+- ++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0
+- ++rl_games.env_eval.ghost_trail.ground_fraction=0.22
+- ++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 Demon Attack from a single
+  game image. Choose exactly one action from: NOOP, FIRE, RIGHT, LEFT, RIGHTFIRE,
+  LEFTFIRE.'''
+- ++rl_games.env_eval.eval_parallel_envs=5
+- ++rl_games.env_eval.action_chunk_execution.enabled=false
+- ++rl_games.env_eval.action_chunk_execution.chunk_size=null
+- ++rl_games.env_eval.demon_attack.noop_max=30
+- ++rl_games.env_eval.enabled=true
+- ++rl_games.env_eval.eval_backend=eval_core
+- ++rl_games.env_eval.distributed_mode=rank_sharded
+- ++rl_games.env_eval.vectorized.enabled=false
+- ++rl_games.env_eval.vectorized.batch_size=1
+- ++rl_games.env_eval.latency.prompt_map_path=data/demon_attack_fix_latency_6_200ep_context5/demon_attack_train__bridge/latency_prompt_map.json
+- ++rl_games.env_eval.latency.mode=single
+- ++rl_games.env_eval.latency.values=[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.num_episodes=5
+- ++rl_games.env_eval.mid_train.max_steps_per_episode=3600
+- ++rl_games.env_eval.post_train.enabled=false
+- ++rl_games.env_eval.post_train.latencies=[6]
+- ++rl_games.env_eval.post_train.num_episodes=20
+- ++rl_games.env_eval.post_train.max_steps_per_episode=3600
+- ++rl_games.task=demon_attack
+- ++rl_games.initialization_mode=bridge
+- ++rl_games.action_carrier=bridge
+- ++checkpoint.load=auto
+- ++checkpoint.hf_repo_id=null
+- ++checkpoint.save_best_model=false
+- ++checkpoint.save_final_model=true
+- ++checkpoint.save_pt_file=true
+- ++checkpoint.save_training_state=true
+- ++checkpoint.save_safetensors_file=false
+- ++checkpoint.local.keep_last_n=1
+- ++checkpoint.sync.enabled=false
+- ++checkpoint.sync.repo_id=null
+- ++checkpoint.sync.keep_last_n=0
+- ++checkpoint.sync.sync_every_n_checkpoints=1
+- ++checkpoint.sync.resume_policy=local_latest
+- ++run_id=wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce
+- ++output_dir=null
+- ++config_yaml=null
+- ++is_debug=false
+- ++version_id=0.21
+- ++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints
+- ++trainer.is_resume=true
+- ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_4000_state
+- ++trainer.resume_step=4000
+- ++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/data/demon_attack_fix_latency_6_200ep_context5
+- ++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+- ++framework.world_model.base_wm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers

evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/summary.jsonl

@@ -0,0 +1,10 @@
+{"steps": 500}
+{"steps": 1000}
+{"steps": 1500}
+{"steps": 2000}
+{"steps": 2500}
+{"steps": 3000}
+{"steps": 3500}
+{"steps": 4000}
+{"steps": 4500}
+{"steps": 5000}

evaluations/flappy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/train/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/episode_metrics.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db8769831ceffad9d0fa3a2970035339e038fdc549b9de3253d1dd05f5f9fb06
+size 23536650

evaluations/flappy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/train/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/git.diff

@@ -0,0 +1,1947 @@
+diff --git a/configs/experiments/fixed_latency_curve.yaml b/configs/experiments/fixed_latency_curve.yaml
+deleted file mode 100644
+index 3ab184f..0000000
+--- a/configs/experiments/fixed_latency_curve.yaml
++++ /dev/null
+@@ -1,17 +0,0 @@
+-__base__: ../base/dummy.yaml
+-
+-experiment:
+-  name: fixed_latency_curve
+-
+-executor:
+-  mode: simulated
+-
+-latency:
+-  method: fixed
+-  fixed_latency_ms: 100
+-
+-evaluation:
+-  eval_episodes: 20
+-
+-logging:
+-  output_dir: outputs/fixed_latency_curve
+diff --git a/configs/experiments/flappy/frame_stack/base.yaml b/configs/experiments/flappy/frame_stack/base.yaml
+deleted file mode 100644
+index f1557d5..0000000
+--- a/configs/experiments/flappy/frame_stack/base.yaml
++++ /dev/null
+@@ -1,123 +0,0 @@
+-experiment:
+-  name: flappy_base
+-  seed: 0
+-
+-backend:
+-  type: sample_factory
+-  algo: APPO
+-  device: cuda
+-  train_dir: results/checkpoints
+-  restart_behavior: resume
+-  run_mode: train
+-
+-executor:
+-  mode: simulated
+-
+-env:
+-  name: flappy
+-  # note that env_fps must be an integer multiple of obs_fps!
+-  env_fps: 30
+-  obs_fps: 30
+-  gym_id: FlappyBird-v0
+-  render_mode: rgb_array
+-  observation_mode: image
+-  use_lidar: false
+-  normalize_obs: true
+-  audio_on: false
+-  frame_stack: 4
+-  obs_resize:
+-    - 84
+-    - 84
+-  # for training it could be true, but for realtime mode, THIS MUST BE FALSE!
+-  use_gpu_render: true
+-  simulator: gpu
+-  gpu_render_device: auto
+-  gpu_render_profile: false
+-  gpu_render_profile_interval: 200
+-  noop_action: noop
+-  action_map:
+-    noop: 0
+-    flap: 1
+-  oneshot_actions:
+-    - flap
+-
+-latency:
+-  method: normal
+-  mean_latency_ms: 66.7
+-  std_latency_ms: 33.3
+-  min_latency_ms: 0
+-  max_latency_ms: 500.0
+-
+-
+-scheduler:
+-  # [one_frame_then_noop, hold]
+-  # if hold, default hold duration is: env_fps / obs_fps
+-  hold_policy: one_frame_then_noop
+-  # can be either `issue_order_fifo` or `latest ready`
+-  ordering_policy: latest_ready
+-
+-policy:
+-  type: random
+-  actions:
+-    - noop
+-    - flap
+-
+-training:
+-  train_for_env_steps: 25000000
+-  num_workers: 2
+-  num_envs_per_worker: 128
+-  worker_num_splits: 1
+-  num_policies: 1
+-  batch_size: 4096
+-  rollout: 128
+-  recurrence: 1
+-  num_epochs: 8
+-  num_batches_per_epoch: 8
+-  max_policy_lag: 400
+-  learning_rate: 0.00025
+-  gamma: 0.99
+-  gae_lambda: 0.95
+-  ppo_clip_ratio: 0.1
+-  ppo_clip_value: 0.2
+-  value_loss_coeff: 0.5
+-  max_grad_norm: 0.5
+-  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: 1000000
+-  eval_episodes: 100
+-  eval_parallel_envs: 100
+-  eval_max_steps: 3600
+-  eval_deterministic: true
+-  eval_raw_reward: false
+-  eval_suites:
+-    fixed:
+-      - [0, 8]    # [start, end]
+-    normal:
+-      - [2.0, 0.5, 0, 15]   # low latency, low jitter
+-      - [2.0, 1.0, 0, 15]   # low latency, medium jitter
+-      - [4.0, 0.5, 0, 15]   # high latency, low jitter
+-      - [4.0, 1.0, 0, 15]   # high latency, medium jitter
+-      - [4.0, 2.0, 0, 15]   # high latency, high jitter
+-    uniform:
+-      - [1, 3]  # low latency, low jitter
+-      - [0, 4]  # low latency, high jitter
+-      - [3, 5]  # high latency, low jitter
+-      - [2, 6]  # high latency, high jitter
+-
+-logging:
+-  output_dir: outputs/flappy_base
+-  save_step_records: true
+-  save_action_records: true
+-  save_latency_records: true
+-  wandb_project: latency-sensitive-bench
+-  wandb_group: flappy-frame_stack
+-  wandb_job_type: sample_factory
+-  wandb_tags: ""
+diff --git a/latency_bench/data/exporters/starvla_lerobot.py b/latency_bench/data/exporters/starvla_lerobot.py
+index c693460..4ad276a 100644
+--- a/latency_bench/data/exporters/starvla_lerobot.py
++++ b/latency_bench/data/exporters/starvla_lerobot.py
+@@ -18,12 +18,17 @@ from latency_bench.utils.io import write_jsonl
+ from latency_bench.data.action_layouts import action_spec_for_env
+ from latency_bench.data.filters import density_weights_clean_v1
+ 
+-FPS_BY_ENV = {
+-    "flappy": 30,
+-    "demon_attack": 60,
+-    "deadly_corridor": 35,
+-}
+ STATE_DIM = 1
++BRIDGE_DIM = 7
++DEADLY_CORRIDOR_BRIDGE_ACTION_LABELS = (
++    "MOVE_FORWARD",
++    "MOVE_BACKWARD",
++    "MOVE_LEFT",
++    "MOVE_RIGHT",
++    "TURN_LEFT",
++    "TURN_RIGHT",
++    "ATTACK",
++)
+ 
+ 
+ def _png_bytes(image: Any) -> bytes:
+@@ -64,30 +69,38 @@ def _load_rows(raw_export_dir: Path, split: str) -> list[dict[str, Any]]:
+ 
+ 
+ def _episode_groups(rows: list[dict[str, Any]]) -> list[list[dict[str, Any]]]:
+-    """Group canonical rows by episode and order by decision step.
++    """Group canonical rows by trajectory and order by decision step.
+ 
+     Args:
+         rows: Canonical rollout rows from one split.
+ 
+     Returns:
+-        Episode row groups sorted by episode index.
++        Trajectory row groups sorted by episode and trajectory index.
+     """
+-    grouped: dict[int, list[dict[str, Any]]] = {}
++    grouped: dict[tuple[int, int], list[dict[str, Any]]] = {}
+     for row in rows:
+-        grouped.setdefault(int(row["episode_idx"]), []).append(row)
++        key = (int(row["episode_idx"]), int(row["trajectory_idx"]))
++        grouped.setdefault(key, []).append(row)
+     return [
+         sorted(episode_rows, key=lambda row: int(row["decision_step"]))
+         for _, episode_rows in sorted(grouped.items())
+     ]
+ 
+ 
+-def _write_episode(path: Path, rows: list[dict[str, Any]], *, action_dim: int) -> None:
++def _write_episode(
++    path: Path,
++    rows: list[dict[str, Any]],
++    *,
++    action_dim: int,
++    state_dim: int,
++) -> None:
+     """Write one LeRobot episode parquet file.
+ 
+     Args:
+         path: Destination episode parquet path.
+         rows: Converted LeRobot frame rows.
+         action_dim: Expected action vector dimension.
++        state_dim: Expected state carrier dimension.
+     """
+     path.parent.mkdir(parents=True, exist_ok=True)
+     table = pa.table(
+@@ -97,8 +110,8 @@ def _write_episode(path: Path, rows: list[dict[str, Any]], *, action_dim: int) -
+                 type=pa.struct([("bytes", pa.binary()), ("path", pa.string())]),
+             ),
+             "observation.state": pa.array(
+-                [[0.0] * STATE_DIM for _ in rows],
+-                type=pa.list_(pa.float32(), STATE_DIM),
++                [[0.0] * state_dim for _ in rows],
++                type=pa.list_(pa.float32(), state_dim),
+             ),
+             "action": pa.array(
+                 [row["action"] for row in rows],
+@@ -131,11 +144,47 @@ def _row_action_vector(row: dict[str, Any], *, action_dim: int) -> list[float]:
+     return [1.0 if idx == action_id else 0.0 for idx in range(action_dim)]
+ 
+ 
++def _bridge_action_vector(
++    row: dict[str, Any],
++    *,
++    env_name: str,
++    active_action_dim: int,
++) -> list[float]:
++    if env_name == "deadly_corridor":
++        action_text = str(row["action_text"]).upper()
++        return [
++            1.0 if label in action_text else 0.0
++            for label in DEADLY_CORRIDOR_BRIDGE_ACTION_LABELS
++        ]
++    action = _row_action_vector(row, action_dim=active_action_dim)
++    return [*action, *([0.0] * (BRIDGE_DIM - active_action_dim))]
++
++
++def _target_action_labels(
++    *,
++    env_name: str,
++    source_action_labels: list[str],
++    action_carrier: str,
++) -> list[str]:
++    if action_carrier == "native":
++        return source_action_labels
++    if env_name == "deadly_corridor":
++        return list(DEADLY_CORRIDOR_BRIDGE_ACTION_LABELS)
++    return [
++        *source_action_labels,
++        *[
++            f"BRIDGE_PAD_{idx}"
++            for idx in range(len(source_action_labels), BRIDGE_DIM)
++        ],
++    ]
++
++
+ def _write_metadata(
+     dataset_dir: Path,
+     *,
+-    fps: int,
++    fps: float,
+     action_labels: list[str],
++    state_dim: int,
+     episode_lengths: list[int],
+     task_prompts: list[str],
+ ) -> None:
+@@ -144,7 +193,8 @@ def _write_metadata(
+     Args:
+         dataset_dir: LeRobot dataset split directory.
+         fps: Frame rate stored in metadata.
+-        action_labels: Canonical action labels in vector order.
++        action_labels: Target action labels in vector order.
++        state_dim: State carrier dimension.
+         episode_lengths: Number of frames per converted episode.
+         task_prompts: Unique prompts used as LeRobot tasks.
+     """
+@@ -155,10 +205,10 @@ def _write_metadata(
+         json.dumps(
+             {
+                 "codebase_version": "v2.0",
+-                "fps": int(fps),
++                "fps": fps,
+                 "features": {
+                     "observation.image": {"dtype": "image"},
+-                    "observation.state": {"dtype": "float32", "shape": [STATE_DIM]},
++                    "observation.state": {"dtype": "float32", "shape": [state_dim]},
+                     "action": {"dtype": "float32", "shape": [action_dim], "names": action_labels},
+                     "timestamp": {"dtype": "float64", "shape": [1]},
+                     "episode_index": {"dtype": "int64", "shape": [1]},
+@@ -183,9 +233,11 @@ def _convert_split(
+     *,
+     env_name: str,
+     action_labels: list[str],
+-    fps: int,
++    fps: float,
+     reward_field: str,
+-    obs_stride: int = 1,
++    action_carrier: str,
++    active_action_dim: int,
++    state_dim: int,
+     density_weighting: bool = False,
+ ) -> dict[str, Any]:
+     """Convert one canonical rollout split into a LeRobot dataset directory.
+@@ -195,13 +247,11 @@ def _convert_split(
+         output_dir: Destination LeRobot split directory.
+         env_name: Canonical benchmark environment name.
+         action_labels: Canonical action labels in vector order.
+-        fps: Frame rate used for timestamps.
++        fps: Observation frame rate used for timestamps.
+         reward_field: Source canonical reward column.
+-        obs_stride: Raw frames per observation decision (``round(env_fps/obs_fps)``).
+-            Canonical rows are recorded once per raw frame; keeping only the obs
+-            decisions (``decision_step % obs_stride == 0``) makes adjacent LeRobot
+-            rows adjacent observations, so KV-memory rollouts unroll at the eval
+-            observation cadence.
++        action_carrier: Target action/state carrier.
++        active_action_dim: Number of active environment action dimensions.
++        state_dim: Target state carrier dimension.
+         density_weighting: When True, write per-row ``density_weight`` from clean_v1's
+             event balancing (train split). When False, weights are uniform 1.0 (val
+             split keeps an unweighted held-out loss).
+@@ -214,7 +264,6 @@ def _convert_split(
+     task_prompts: list[str] = []
+     episode_lengths: list[int] = []
+     for new_episode_idx, episode_rows in enumerate(_episode_groups(rows)):
+-        episode_rows = [row for row in episode_rows if int(row["decision_step"]) % obs_stride == 0]
+         density_weights = (
+             density_weights_clean_v1(env_name, episode_rows)
+             if density_weighting
+@@ -231,8 +280,14 @@ def _convert_split(
+             out_rows.append(
+                 {
+                     "image_bytes": _png_bytes(row["image"]),
+-                    "action": _row_action_vector(row, action_dim=len(action_labels)),
+-                    "timestamp": float(frame_idx) / int(fps),
++                    "action": _bridge_action_vector(
++                        row,
++                        env_name=env_name,
++                        active_action_dim=active_action_dim,
++                    )
++                    if action_carrier == "bridge"
++                    else _row_action_vector(row, action_dim=len(action_labels)),
++                    "timestamp": float(frame_idx) / fps,
+                     "episode_index": new_episode_idx,
+                     "frame_index": frame_idx,
+                     "task_index": prompt_to_task_index[prompt],
+@@ -248,11 +303,13 @@ def _convert_split(
+             output_dir / f"data/chunk-{episode_chunk:03d}/episode_{new_episode_idx:06d}.parquet",
+             out_rows,
+             action_dim=len(action_labels),
++            state_dim=state_dim,
+         )
+     _write_metadata(
+         output_dir,
+         fps=fps,
+         action_labels=action_labels,
++        state_dim=state_dim,
+         episode_lengths=episode_lengths,
+         task_prompts=task_prompts,
+     )
+@@ -269,6 +326,7 @@ def convert_canonical_rollout_to_lerobot(
+     raw_export_dir: Path | str,
+     output_dir: Path | str,
+     *,
++    action_carrier: str = "native",
+     force: bool = False,
+ ) -> dict[str, Any]:
+     """Convert a canonical rollout export into train/val LeRobot datasets.
+@@ -276,6 +334,7 @@ def convert_canonical_rollout_to_lerobot(
+     Args:
+         raw_export_dir: Directory with canonical rollout parquet files and metadata.
+         output_dir: Destination LeRobot train dataset directory.
++        action_carrier: Target action/state carrier, ``native`` or ``bridge``.
+         force: Remove existing train/val output directories before writing.
+ 
+     Returns:
+@@ -286,11 +345,20 @@ def convert_canonical_rollout_to_lerobot(
+     metadata = json.loads((raw_export_dir / "metadata.json").read_text(encoding="utf-8"))
+     env_name = str(metadata["env_name"])
+     action_spec = action_spec_for_env(env_name)
+-    action_labels = list(metadata["action_labels"])
++    source_action_labels = list(metadata["action_labels"])
++    action_labels = _target_action_labels(
++        env_name=env_name,
++        source_action_labels=source_action_labels,
++        action_carrier=action_carrier,
++    )
++    active_action_dim = {
++        "flappy": 2,
++        "demon_attack": 6,
++        "deadly_corridor": 7,
++    }[env_name]
++    state_dim = {"native": STATE_DIM, "bridge": BRIDGE_DIM}[action_carrier]
+     reward_field = str(metadata["reward_field"])
+-    obs_stride = int(metadata.get("obs_stride_raw_frames", 1))
+-    # Rows survive at the observation cadence, so timestamps advance at obs_fps.
+-    fps = max(1, int(FPS_BY_ENV[env_name]) // obs_stride)
++    fps = metadata["obs_fps"]
+     val_output_dir = output_dir.with_name(f"{output_dir.name}__val")
+     if output_dir.exists() and force:
+         shutil.rmtree(output_dir)
+@@ -303,7 +371,9 @@ def convert_canonical_rollout_to_lerobot(
+         action_labels=action_labels,
+         fps=fps,
+         reward_field=reward_field,
+-        obs_stride=obs_stride,
++        action_carrier=action_carrier,
++        active_action_dim=active_action_dim,
++        state_dim=state_dim,
+         density_weighting=True,
+     )
+     val_manifest = _convert_split(
+@@ -313,7 +383,9 @@ def convert_canonical_rollout_to_lerobot(
+         action_labels=action_labels,
+         fps=fps,
+         reward_field=reward_field,
+-        obs_stride=obs_stride,
++        action_carrier=action_carrier,
++        active_action_dim=active_action_dim,
++        state_dim=state_dim,
+         density_weighting=False,
+     )
+     manifest = {
+@@ -321,9 +393,13 @@ def convert_canonical_rollout_to_lerobot(
+         "format": "starvla_lerobot_v2_image_parquet",
+         "source": str(raw_export_dir),
+         "action_layout": str(metadata["action_layout"]),
++        "action_carrier": action_carrier,
+         "action_labels": action_labels,
+         "action_dim": len(action_labels),
+-        "active_action_dim": len(action_spec.labels),
++        "active_action_dim": active_action_dim
++        if action_carrier == "bridge"
++        else len(action_spec.labels),
++        "state_dim": state_dim,
+         "validation_dataset_name": val_output_dir.name,
+         "validation_episodes": val_manifest["episodes"],
+         "validation_frames": val_manifest["frames"],
+diff --git a/latency_bench/data/hf_parquet.py b/latency_bench/data/hf_parquet.py
+index 7a26426..6163c0d 100644
+--- a/latency_bench/data/hf_parquet.py
++++ b/latency_bench/data/hf_parquet.py
+@@ -205,7 +205,9 @@ def completed_episode_indices_from_shards(output_dir: Path) -> set[int]:
+             with shard_path.open("r", encoding="utf-8") as handle:
+                 for line in handle:
+                     if line.strip():
+-                        completed_episode_indices.add(int(json.loads(line)["episode_idx"]))
++                        row = json.loads(line)
++                        row["trajectory_idx"]
++                        completed_episode_indices.add(int(row["episode_idx"]))
+                         break
+     return completed_episode_indices
+ 
+diff --git a/latency_bench/data/rollout_schema.py b/latency_bench/data/rollout_schema.py
+index 11426b9..21d2bf5 100644
+--- a/latency_bench/data/rollout_schema.py
++++ b/latency_bench/data/rollout_schema.py
+@@ -7,11 +7,12 @@ from datasets import Features, Image as HFImage, Sequence as HFSequence, Value
+ from latency_bench.data.action_layouts import ActionSpec
+ 
+ 
+-SCHEMA_VERSION = 2
++SCHEMA_VERSION = 3
+ CANONICAL_ROW_FIELDS = (
+     "image",
+     "env_name",
+     "episode_idx",
++    "trajectory_idx",
+     "decision_step",
+     "split",
+     "seed",
+@@ -26,6 +27,7 @@ def _rollout_feature_map(action_spec: ActionSpec) -> dict[str, Any]:
+         "image": HFImage(),
+         "env_name": Value("string"),
+         "episode_idx": Value("int64"),
++        "trajectory_idx": Value("int64"),
+         "decision_step": Value("int64"),
+         "split": Value("string"),
+         "seed": Value("int64"),
+diff --git a/latency_bench/data/sf_teacher_rollout.py b/latency_bench/data/sf_teacher_rollout.py
+index af2b64f..bbbd157 100644
+--- a/latency_bench/data/sf_teacher_rollout.py
++++ b/latency_bench/data/sf_teacher_rollout.py
+@@ -355,6 +355,7 @@ def _write_filtered_episode_images(
+     image_writer: RolloutImageWriter,
+     rows: list[dict[str, Any]],
+     frames_by_decision_step: Mapping[int, np.ndarray],
++    trajectory_by_decision_step: Mapping[int, int],
+     *,
+     context_window: int,
+ ) -> list[dict[str, Any]]:
+@@ -364,6 +365,7 @@ def _write_filtered_episode_images(
+         image_writer: Image writer shared by the rollout run.
+         rows: Final rows selected for materialization.
+         frames_by_decision_step: Captured frames keyed by decision step.
++        trajectory_by_decision_step: Trajectory index keyed by decision step.
+         context_window: Total image window length including the current ``image`` field.
+ 
+     Returns:
+@@ -371,10 +373,15 @@ def _write_filtered_episode_images(
+     """
+     if not rows:
+         return []
+-    available_decision_steps = sorted(frames_by_decision_step)
+     required_steps: set[int] = set()
+     for row in rows:
+         decision_step = int(row["decision_step"])
++        trajectory_idx = int(row["trajectory_idx"])
++        available_decision_steps = sorted(
++            step
++            for step, step_trajectory_idx in trajectory_by_decision_step.items()
++            if step_trajectory_idx == trajectory_idx
++        )
+         required_steps.add(decision_step)
+         if int(context_window) > 1:
+             required_steps.update(
+@@ -396,6 +403,12 @@ def _write_filtered_episode_images(
+     for row in rows:
+         written = dict(row)
+         decision_step = int(written["decision_step"])
++        trajectory_idx = int(written["trajectory_idx"])
++        available_decision_steps = sorted(
++            step
++            for step, step_trajectory_idx in trajectory_by_decision_step.items()
++            if step_trajectory_idx == trajectory_idx
++        )
+         written["image"] = written_paths_by_step[decision_step]
+         if int(context_window) > 1:
+             written["context_images"] = [
+@@ -804,6 +817,7 @@ def _replay_accepted_specs(
+     active_episode_rows: list[list[dict[str, Any]]] = [[] for _ in range(driver.slot_count)]
+     active_episode_frames: list[list[np.ndarray]] = [[] for _ in range(driver.slot_count)]
+     active_decision_steps = [0 for _ in range(driver.slot_count)]
++    active_trajectory_indices = [0 for _ in range(driver.slot_count)]
+     active_ghost_streams: list[Any | None] = [None for _ in range(driver.slot_count)]
+     rnn_states = torch.zeros([driver.slot_count, rnn_size], dtype=torch.float32, device=device)
+     image_writer = RolloutImageWriter(
+@@ -844,6 +858,7 @@ def _replay_accepted_specs(
+                 active_episode_rows[slot] = []
+                 active_episode_frames[slot] = []
+                 active_decision_steps[slot] = 0
++                active_trajectory_indices[slot] = 0
+                 rnn_states[slot].zero_()
+ 
+         with tqdm(total=len(specs_to_replay), initial=0, unit="episode") as progress:
+@@ -881,21 +896,21 @@ def _replay_accepted_specs(
+                         next_decision_step=next_decision_step,
+                         max_steps_per_episode=int(max_steps_per_episode),
+                     )
+-                    if not life_loss_boundary:
+-                        active_episode_frames[slot].append(np.array(image_frame, copy=True))
+-                        row = teacher_dataset_row(
+-                            env_name=env_name,
+-                            spec=spec,
+-                            decision_step=decision_step,
+-                            image=str(image_path),
+-                            raw_action=raw_actions[batch_idx],
+-                            info=step_result.info,
+-                            reward=step_result.reward,
+-                            env_fps=env_fps,
+-                            obs_fps=obs_fps,
+-                            image_transform=image_transform_config.normalized_transform(),
+-                        )
+-                        active_episode_rows[slot].append(row)
++                    active_episode_frames[slot].append(np.array(image_frame, copy=True))
++                    row = teacher_dataset_row(
++                        env_name=env_name,
++                        spec=spec,
++                        trajectory_idx=active_trajectory_indices[slot],
++                        decision_step=decision_step,
++                        image=str(image_path),
++                        raw_action=raw_actions[batch_idx],
++                        info=step_result.info,
++                        reward=step_result.reward,
++                        env_fps=env_fps,
++                        obs_fps=obs_fps,
++                        image_transform=image_transform_config.normalized_transform(),
++                    )
++                    active_episode_rows[slot].append(row)
+                     rnn_states[slot] = new_rnn_states[batch_idx]
+                     if episode_done:
+                         final_rows = _finalize_attempt_rows(
+@@ -931,6 +946,10 @@ def _replay_accepted_specs(
+                             image_writer,
+                             shard_rows,
+                             frames_by_decision_step,
++                            {
++                                int(row["decision_step"]): int(row["trajectory_idx"])
++                                for row in final_rows
++                            },
+                             context_window=int(context_window),
+                         )
+                         write_episode_row_shard(output_dir, shard_rows)
+@@ -943,12 +962,14 @@ def _replay_accepted_specs(
+                         active_episode_rows[slot] = []
+                         active_episode_frames[slot] = []
+                         active_decision_steps[slot] = 0
++                        active_trajectory_indices[slot] = 0
+                         if _is_demon_ghost_trail:
+                             active_ghost_streams[slot] = None
+                         rnn_states[slot].zero_()
+                     else:
+                         if life_loss_boundary:
+                             active_obs[slot], reset_info = driver.continue_after_life_loss(slot)
++                            active_trajectory_indices[slot] += 1
+                             if _is_demon_ghost_trail:
+                                 active_ghost_streams[slot].clear()
+                                 active_image_frames[slot] = frame_from_env_raw_rgb_info(reset_info).copy()
+@@ -981,6 +1002,7 @@ def _replay_accepted_specs(
+                         active_episode_rows[slot] = []
+                         active_episode_frames[slot] = []
+                         active_decision_steps[slot] = 0
++                        active_trajectory_indices[slot] = 0
+                         rnn_states[slot].zero_()
+                 else:
+                     if isinstance(driver, BatchedTeacherRolloutDriver):
+diff --git a/latency_bench/data/sf_teacher_rows.py b/latency_bench/data/sf_teacher_rows.py
+index efa6f3e..4e7ce9a 100644
+--- a/latency_bench/data/sf_teacher_rows.py
++++ b/latency_bench/data/sf_teacher_rows.py
+@@ -14,6 +14,7 @@ def teacher_dataset_row(
+     *,
+     env_name: str,
+     spec: Mapping[str, Any],
++    trajectory_idx: int,
+     decision_step: int,
+     image: str,
+     raw_action: Any,
+@@ -28,6 +29,7 @@ def teacher_dataset_row(
+     Args:
+         env_name: Canonical benchmark environment name.
+         spec: Episode spec containing index, split, and seed.
++        trajectory_idx: Zero-based trajectory index within the full-game episode.
+         decision_step: Zero-based policy decision step within the episode.
+         image: Path to the saved pre-step image.
+         raw_action: Raw teacher action for this decision.
+@@ -48,6 +50,7 @@ def teacher_dataset_row(
+         "image": image,
+         "env_name": str(env_name),
+         "episode_idx": int(spec["episode_idx"]),
++        "trajectory_idx": int(trajectory_idx),
+         "decision_step": int(decision_step),
+         "split": str(spec["split"]),
+         "seed": int(spec["seed"]),
+diff --git a/latency_bench/integrations/sample_factory/runtime.py b/latency_bench/integrations/sample_factory/runtime.py
+index e1efadb..aa4f941 100644
+--- a/latency_bench/integrations/sample_factory/runtime.py
++++ b/latency_bench/integrations/sample_factory/runtime.py
+@@ -58,7 +58,10 @@ class SampleFactoryIntegrationModule:
+         getattr(self.module, self._register_attr)()
+ 
+     def build_sf_cfg_from_config(self, config: dict[str, Any]):
+-        return self.module.parse_cfg(self.module.build_cli_args_from_config(config))
++        return self.module.parse_cfg(
++            self.module.build_cli_args_from_config(config),
++            evaluation=True,
++        )
+ 
+     def make_reference_env(self, cfg: Any):
+         return self.module.make_reference_env(cfg)
+diff --git a/scripts/experiment_runner/run_manifest_jobs.py b/scripts/experiment_runner/run_manifest_jobs.py
+index e2d1b5e..133ccdc 100644
+--- a/scripts/experiment_runner/run_manifest_jobs.py
++++ b/scripts/experiment_runner/run_manifest_jobs.py
+@@ -73,11 +73,6 @@ def _event_writer(event_log: Path) -> Callable[[dict[str, Any]], None]:
+     return write
+ 
+ 
+-def _register_process(process: subprocess.Popen) -> None:
+-    with _ACTIVE_PROCESSES_LOCK:
+-        _ACTIVE_PROCESSES.add(process)
+-
+-
+ def _forget_process(process: subprocess.Popen) -> None:
+     with _ACTIVE_PROCESSES_LOCK:
+         _ACTIVE_PROCESSES.discard(process)
+@@ -89,11 +84,10 @@ def _active_processes() -> list[subprocess.Popen]:
+ 
+ 
+ def _signal_process_group(process: subprocess.Popen, sig: int) -> None:
+-    if process.poll() is None:
+-        try:
+-            os.killpg(process.pid, sig)
+-        except ProcessLookupError:
+-            pass
++    try:
++        os.killpg(process.pid, sig)
++    except ProcessLookupError:
++        pass
+ 
+ 
+ def _terminate_active_processes(sig: int = signal.SIGTERM) -> None:
+@@ -118,15 +112,24 @@ def _terminate_active_processes(sig: int = signal.SIGTERM) -> None:
+             pass
+ 
+ 
+-def _run_command(command: str, *, gpu: str) -> None:
++def _run_command(
++    command: str,
++    *,
++    gpu: str,
++    cancel_event: threading.Event | None = None,
++) -> bool:
+     env = os.environ.copy()
+     env["CUDA_VISIBLE_DEVICES"] = str(gpu)
+-    process = subprocess.Popen(command, shell=True, env=env, start_new_session=True)
+-    _register_process(process)
++    with _ACTIVE_PROCESSES_LOCK:
++        if cancel_event is not None and cancel_event.is_set():
++            return False
++        process = subprocess.Popen(command, shell=True, env=env, start_new_session=True)
++        _ACTIVE_PROCESSES.add(process)
+     try:
+         returncode = process.wait()
+         if returncode != 0:
+             raise subprocess.CalledProcessError(returncode, command)
++        return True
+     finally:
+         _forget_process(process)
+ 
+@@ -201,14 +204,11 @@ def _resume_stage(job: dict[str, str], states: dict[str, str]) -> str:
+     stage = states.get(str(job["job_id"]), "")
+     if stage in {"job_complete", "upload_skipped", "upload_succeeded"}:
+         return "complete"
+-    if stage in {"eval_succeeded", "upload_failed"}:
++    if stage in {"eval_succeeded", "upload_started", "upload_failed"}:
+         return "upload"
+-    if stage == "train_succeeded":
+-        return "eval"
+-    if stage == "eval_failed":
+-        # Training already succeeded before eval ran; re-eval from the checkpoint
+-        # instead of retraining from scratch.
+-        return "eval"
++    if stage in {"train_succeeded", "eval_started", "eval_failed"}:
++        train_config = Path(job["train_artifact_dir"]) / "config.json"
++        return "eval" if train_config.is_file() else "train"
+     return "train"
+ 
+ 
+@@ -223,16 +223,6 @@ def run_jobs(
+     dry_run: bool = False,
+     uploader: Callable[[dict[str, str]], None] | None = None,
+ ) -> int:
+-    previous_sigint = signal.getsignal(signal.SIGINT)
+-    previous_sigterm = signal.getsignal(signal.SIGTERM)
+-
+-    def stop_active_processes(signum, frame):
+-        del frame
+-        _terminate_active_processes(signal.SIGTERM)
+-        raise KeyboardInterrupt(signum)
+-
+-    signal.signal(signal.SIGINT, stop_active_processes)
+-    signal.signal(signal.SIGTERM, stop_active_processes)
+     manifest_file = Path(manifest_path).expanduser().resolve()
+     event_log = event_log_path_for_manifest(manifest_file)
+     write_event = _event_writer(event_log)
+@@ -253,12 +243,16 @@ def run_jobs(
+         jobs = [
+             row
+             for row in rows
+-            if states.get(row["job_id"], "") in {"eval_succeeded", "upload_failed"}
++            if states.get(row["job_id"], "")
++            in {"eval_succeeded", "upload_started", "upload_failed"}
+         ]
+     job_queue: queue.Queue[dict[str, str] | None] = queue.Queue()
+     upload_queue: queue.Queue[UploadTask | None] = queue.Queue()
+     failures: list[tuple[str, str]] = []
+     failure_lock = threading.Lock()
++    cancel_event = threading.Event()
++    signal_event = threading.Event()
++    cleanup_started = False
+     uploader_fn = _default_uploader if uploader is None else uploader
+ 
+     if stage != "upload":
+@@ -280,28 +274,49 @@ def run_jobs(
+             if job is None:
+                 job_queue.task_done()
+                 break
++            if cancel_event.is_set():
++                job_queue.task_done()
++                continue
+             next_stage = _resume_stage(job, states) if stage == "all" else stage
+             job_id = str(job["job_id"])
+             try:
+                 if next_stage == "train":
++                    if cancel_event.is_set():
++                        continue
+                     if dry_run:
+                         print(f"[dry-run] CUDA_VISIBLE_DEVICES={gpu} {job['train_command']}", flush=True)
+                     else:
+                         write_event({"job_id": job_id, "stage": "train_started", "gpu": gpu, "ts": _now()})
+-                        _run_command(str(job["train_command"]), gpu=gpu)
++                        started = _run_command(
++                            str(job["train_command"]),
++                            gpu=gpu,
++                            cancel_event=cancel_event,
++                        )
++                        if not started or cancel_event.is_set():
++                            continue
+                         write_event({"job_id": job_id, "stage": "train_succeeded", "gpu": gpu, "ts": _now()})
+                     next_stage = "eval" if stage == "all" else "complete"
+                 if next_stage == "eval":
++                    if cancel_event.is_set():
++                        continue
+                     if dry_run:
+                         print(f"[dry-run] CUDA_VISIBLE_DEVICES={gpu} {job['eval_command']}", flush=True)
+                     else:
+                         write_event({"job_id": job_id, "stage": "eval_started", "gpu": gpu, "ts": _now()})
+-                        _run_command(str(job["eval_command"]), gpu=gpu)
++                        started = _run_command(
++                            str(job["eval_command"]),
++                            gpu=gpu,
++                            cancel_event=cancel_event,
++                        )
++                        if not started or cancel_event.is_set():
++                            continue
+                         write_event({"job_id": job_id, "stage": "eval_succeeded", "gpu": gpu, "ts": _now()})
+                     next_stage = "upload" if stage == "all" else "complete"
+-                if next_stage == "upload":
++                if next_stage == "upload" and not cancel_event.is_set():
+                     upload_queue.put(UploadTask(job))
+             except subprocess.CalledProcessError:
++                if cancel_event.is_set():
++                    continue
+                 failed_stage = "train_failed" if next_stage == "train" else "eval_failed"
+                 write_event({"job_id": job_id, "stage": failed_stage, "gpu": gpu, "ts": _now()})
+                 record_failure(job_id, failed_stage)
+@@ -311,48 +326,52 @@ def run_jobs(
+     def upload_worker() -> None:
+         while True:
+             task = upload_queue.get()
+-            if task is None:
+-                upload_queue.task_done()
+-                break
+-            job = task.job
+-            job_id = str(job["job_id"])
+-            train_repo_id = str(job["train_upload_repo_id"])
+-            eval_repo_id = str(job["eval_upload_repo_id"])
+-            if dry_run:
+-                targets = [repo_id for repo_id in (eval_repo_id, train_repo_id) if repo_id]
+-                print(f"[dry-run] upload {job_id}: {', '.join(targets) or 'no upload target'}", flush=True)
+-                upload_queue.task_done()
+-                continue
+-            if not train_repo_id and not eval_repo_id:
+-                write_event({"job_id": job_id, "stage": "upload_skipped", "ts": _now()})
+-                write_event({"job_id": job_id, "stage": "job_complete", "ts": _now()})
+-                upload_queue.task_done()
+-                continue
+-            attempt = 0
+-            success = False
+-            while attempt < max_upload_attempts and not success:
+-                attempt += 1
+-                write_event({"job_id": job_id, "stage": "upload_started", "attempt": attempt, "ts": _now()})
+-                try:
+-                    if not dry_run:
+-                        uploader_fn(job)
+-                    write_event({"job_id": job_id, "stage": "upload_succeeded", "attempt": attempt, "ts": _now()})
++            try:
++                if task is None:
++                    break
++                if cancel_event.is_set():
++                    continue
++                job = task.job
++                job_id = str(job["job_id"])
++                train_repo_id = str(job["train_upload_repo_id"])
++                eval_repo_id = str(job["eval_upload_repo_id"])
++                if dry_run:
++                    targets = [repo_id for repo_id in (eval_repo_id, train_repo_id) if repo_id]
++                    print(f"[dry-run] upload {job_id}: {', '.join(targets) or 'no upload target'}", flush=True)
++                    continue
++                if not train_repo_id and not eval_repo_id:
++                    write_event({"job_id": job_id, "stage": "upload_skipped", "ts": _now()})
+                     write_event({"job_id": job_id, "stage": "job_complete", "ts": _now()})
+-                    success = True
+-                except Exception as error:
+-                    write_event(
+-                        {
+-                            "job_id": job_id,
+-                            "stage": "upload_failed",
+-                            "attempt": attempt,
+-                            "error": repr(error),
+-                            "ts": _now(),
+-                        }
+-                    )
+-                    print(f"upload failed for {job_id} (attempt {attempt}): {error!r}", file=sys.stderr)
+-            if not success:
+-                record_failure(job_id, "upload_failed")
+-            upload_queue.task_done()
++                    continue
++                attempt = 0
++                success = False
++                while attempt < max_upload_attempts and not success and not cancel_event.is_set():
++                    attempt += 1
++                    write_event({"job_id": job_id, "stage": "upload_started", "attempt": attempt, "ts": _now()})
++                    try:
++                        uploader_fn(job)
++                        if cancel_event.is_set():
++                            continue
++                        write_event({"job_id": job_id, "stage": "upload_succeeded", "attempt": attempt, "ts": _now()})
++                        write_event({"job_id": job_id, "stage": "job_complete", "ts": _now()})
++                        success = True
++                    except Exception as error:
++                        if cancel_event.is_set():
++                            continue
++                        write_event(
++                            {
++                                "job_id": job_id,
++                                "stage": "upload_failed",
++                                "attempt": attempt,
++                                "error": repr(error),
++                                "ts": _now(),
++                            }
++                        )
++                        print(f"upload failed for {job_id} (attempt {attempt}): {error!r}", file=sys.stderr)
++                if not success and not cancel_event.is_set():
++                    record_failure(job_id, "upload_failed")
++            finally:
++                upload_queue.task_done()
+ 
+     upload_threads = [
+         threading.Thread(target=upload_worker, name=f"upload-{idx}", daemon=True)
+@@ -362,17 +381,44 @@ def run_jobs(
+         threading.Thread(target=train_worker, args=(gpu,), name=f"train-{gpu}", daemon=True)
+         for gpu in gpus if stage != "upload"
+     ]
++    previous_sigint = signal.getsignal(signal.SIGINT)
++    previous_sigterm = signal.getsignal(signal.SIGTERM)
++
++    def stop_active_processes(signum, frame):
++        nonlocal cleanup_started
++        del frame
++        if signal_event.is_set():
++            return
++        signal_event.set()
++        cancel_event.set()
++        if cleanup_started:
++            return
++        raise KeyboardInterrupt(signum)
++
++    signal.signal(signal.SIGINT, stop_active_processes)
++    signal.signal(signal.SIGTERM, stop_active_processes)
+     try:
+         for thread in upload_threads + train_threads:
+             thread.start()
+         job_queue.join()
+-        for _ in upload_threads:
+-            upload_queue.put(None)
+         upload_queue.join()
+-        for thread in train_threads + upload_threads:
+-            thread.join()
+         return 1 if failures else 0
+     finally:
++        cleanup_started = True
++        cancel_event.set()
++        _terminate_active_processes(signal.SIGTERM)
++        for thread in train_threads:
++            thread.join()
++        _terminate_active_processes(signal.SIGTERM)
++        if signal_event.is_set():
++            for _ in upload_threads:
++                upload_queue.put(None)
++        else:
++            upload_queue.join()
++            for _ in upload_threads:
++                upload_queue.put(None)
++            for thread in upload_threads:
++                thread.join()
+         _terminate_active_processes(signal.SIGTERM)
+         signal.signal(signal.SIGINT, previous_sigint)
+         signal.signal(signal.SIGTERM, previous_sigterm)
+diff --git a/scripts/rollout_data/prepare_starvla_lerobot_dataset.py b/scripts/rollout_data/prepare_starvla_lerobot_dataset.py
+index e202657..2521899 100644
+--- a/scripts/rollout_data/prepare_starvla_lerobot_dataset.py
++++ b/scripts/rollout_data/prepare_starvla_lerobot_dataset.py
+@@ -26,6 +26,7 @@ def convert_raw_export_to_lerobot(
+     raw_export_dir: Path,
+     lerobot_root: Path,
+     data_mix: str,
++    action_carrier: str = "native",
+ ) -> dict:
+     """Convert a raw rollout export into a named LeRobot data mix.
+ 
+@@ -33,6 +34,7 @@ def convert_raw_export_to_lerobot(
+         raw_export_dir: Canonical rollout export directory.
+         lerobot_root: Directory that will contain the LeRobot datasets.
+         data_mix: Train dataset directory name under ``lerobot_root``.
++        action_carrier: Target action/state carrier.
+ 
+     Returns:
+         Conversion manifest returned by the LeRobot adapter.
+@@ -41,6 +43,7 @@ def convert_raw_export_to_lerobot(
+     return convert_canonical_rollout_to_lerobot(
+         raw_export_dir,
+         output_dir,
++        action_carrier=action_carrier,
+         force=True,
+     )
+ 
+@@ -92,6 +95,7 @@ def build_parser() -> argparse.ArgumentParser:
+     parser.add_argument("raw_export_dir")
+     parser.add_argument("lerobot_root")
+     parser.add_argument("data_mix")
++    parser.add_argument("--action-carrier", choices=("native", "bridge"), default="native")
+     parser.add_argument("--push-to-hub", default="")
+     parser.add_argument("--private", action="store_true")
+     return parser
+@@ -115,6 +119,7 @@ def main(argv: Sequence[str] | None = None) -> int:
+         raw_export_dir=raw_export_dir,
+         lerobot_root=lerobot_root,
+         data_mix=data_mix,
++        action_carrier=args.action_carrier,
+     )
+     if args.push_to_hub:
+         push_lerobot_dataset(
+diff --git a/scripts/sweeps/generate_sweep_configs.py b/scripts/sweeps/generate_sweep_configs.py
+index 8074c79..bd5fafe 100644
+--- a/scripts/sweeps/generate_sweep_configs.py
++++ b/scripts/sweeps/generate_sweep_configs.py
+@@ -116,6 +116,7 @@ ENV_KEYS_BY_NAME = {
+         "obs_fps",
+         "screen_size",
+         "noop_max",
++        "reset_mode",
+         "frame_stack",
+         "terminal_on_life_loss",
+         "clip_reward",
+diff --git a/tests/data/test_rollout_dataset_schema.py b/tests/data/test_rollout_dataset_schema.py
+index 06a6091..55ccebf 100644
+--- a/tests/data/test_rollout_dataset_schema.py
++++ b/tests/data/test_rollout_dataset_schema.py
+@@ -8,6 +8,7 @@ def test_canonical_rollout_fields_exclude_legacy_aliases():
+         "image",
+         "env_name",
+         "episode_idx",
++        "trajectory_idx",
+         "decision_step",
+         "split",
+         "seed",
+@@ -36,6 +37,7 @@ def test_rollout_features_include_env_specific_action_fields():
+     assert "action_tuple" not in deadly_features
+     assert "action" not in deadly_features
+     for features in (flappy_features, deadly_features):
++        assert "trajectory_idx" in features
+         assert "decision_step" in features
+         assert "latency_raw_frames" in features
+         assert "raw_reward" in features
+diff --git a/tests/data/test_rollout_hf_parquet.py b/tests/data/test_rollout_hf_parquet.py
+index 9fa4c48..4e56de6 100644
+--- a/tests/data/test_rollout_hf_parquet.py
++++ b/tests/data/test_rollout_hf_parquet.py
+@@ -15,6 +15,7 @@ def test_hf_parquet_writes_canonical_columns_only(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -56,6 +57,7 @@ def test_hf_parquet_shards_sort_by_decision_step(tmp_path):
+             "image": str(image_path),
+             "env_name": "flappy",
+             "episode_idx": 0,
++            "trajectory_idx": 0,
+             "decision_step": decision_step,
+             "split": "train",
+             "seed": 11,
+@@ -92,6 +94,7 @@ def test_hf_parquet_embeds_context_images_when_rows_provide_them(tmp_path):
+         "context_images": [str(path) for path in context_paths],
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -116,3 +119,18 @@ def test_hf_parquet_embeds_context_images_when_rows_provide_them(tmp_path):
+     assert "context_images" in table.column_names
+     assert len(exported["context_images"]) == 3
+     assert exported["context_images"][0]["bytes"].startswith(b"\x89PNG\r\n\x1a\n")
++
++
++def test_completed_episode_shards_require_trajectory_identity(tmp_path):
++    import json
++
++    import pytest
++
++    from latency_bench.data.hf_parquet import completed_episode_indices_from_shards
++
++    shard_path = tmp_path / "episode_rows" / "train" / "episode_000000.jsonl"
++    shard_path.parent.mkdir(parents=True)
++    shard_path.write_text(json.dumps({"episode_idx": 0}) + "\n", encoding="utf-8")
++
++    with pytest.raises(KeyError, match="trajectory_idx"):
++        completed_episode_indices_from_shards(tmp_path)
+diff --git a/tests/data/test_sf_teacher_dataset_export.py b/tests/data/test_sf_teacher_dataset_export.py
+index 62b30a6..ba4644b 100644
+--- a/tests/data/test_sf_teacher_dataset_export.py
++++ b/tests/data/test_sf_teacher_dataset_export.py
+@@ -1299,7 +1299,7 @@ def test_write_filtered_episode_images_preserves_empty_filtered_episode(tmp_path
+ 
+     image_writer = RolloutImageWriter(max_workers=0, row_buffer_size=1, export_image_size=None)
+ 
+-    rows = _write_filtered_episode_images(image_writer, [], {}, context_window=5)
++    rows = _write_filtered_episode_images(image_writer, [], {}, {}, context_window=5)
+ 
+     assert rows == []
+ 
+@@ -1314,6 +1314,7 @@ def test_write_filtered_episode_images_resets_context_after_decision_step_gap(tm
+         {
+             "image": str(tmp_path / "images" / "train" / "episode_000000" / f"decision_step_{step:06d}.png"),
+             "decision_step": step,
++            "trajectory_idx": 0,
+         }
+         for step in [0, 1, 4, 5]
+     ]
+@@ -1326,6 +1327,7 @@ def test_write_filtered_episode_images_resets_context_after_decision_step_gap(tm
+         image_writer,
+         rows,
+         frames_by_decision_step,
++        {step: 0 for step in frames_by_decision_step},
+         context_window=4,
+     )
+ 
+@@ -1733,13 +1735,24 @@ def test_collect_teacher_dataset_demon_attack_exports_full_game_across_life_loss
+         image_writer_workers=0,
+         row_buffer_size=1,
+         filter_preset="none",
++        context_window=3,
+     )
+     train_rows = _read_shard_rows(output_dir, "train")
+     state = json.loads((output_dir / "probe_state.json").read_text(encoding="utf-8"))
+ 
+     assert ScriptedDemonEnv.reset_seeds == [100, None, None, 100, None, None]
+-    assert [row["decision_step"] for row in train_rows] == [2]
+-    assert train_rows[0]["raw_reward"] == 30.0
++    assert [row["decision_step"] for row in train_rows] == [0, 1, 2]
++    assert [row["trajectory_idx"] for row in train_rows] == [0, 1, 2]
++    assert [row["action_id"] for row in train_rows] == [1, 1, 1]
++    assert [row["raw_reward"] for row in train_rows] == [10.0, 20.0, 30.0]
++    from PIL import Image
++
++    context_values = [
++        [np.asarray(Image.open(path))[0, 0, 0].item() for path in row["context_images"]]
++        for row in train_rows
++    ]
++    current_values = [np.asarray(Image.open(row["image"]))[0, 0, 0].item() for row in train_rows]
++    assert context_values == [[value, value] for value in current_values]
+     assert state["accepted_specs"] == [
+         {
+             "attempt_idx": 0,
+@@ -1900,6 +1913,7 @@ def test_collect_teacher_dataset_skips_runtime_when_probe_and_shards_complete(mo
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 100,
+@@ -2264,6 +2278,7 @@ def test_write_parquet_splits_handles_empty_validation_rows(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -2300,6 +2315,7 @@ def test_write_parquet_splits_embeds_image_bytes(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -2336,6 +2352,7 @@ def test_write_parquet_splits_preserves_rows_across_large_exports(tmp_path):
+             "image": str(image_path),
+             "env_name": "flappy",
+             "episode_idx": episode_idx,
++            "trajectory_idx": 0,
+             "decision_step": 0,
+             "split": "train",
+             "seed": 11 + episode_idx,
+@@ -2390,6 +2407,7 @@ def  test_write_parquet_splits_from_shards_writes_train_and_val_rows(tmp_path):
+                 "image": str(image_path),
+                 "env_name": "flappy",
+                 "episode_idx": episode_idx,
++                "trajectory_idx": 0,
+                 "decision_step": decision_step,
+                 "split": split,
+                 "seed": 11 + episode_idx,
+@@ -2432,6 +2450,7 @@ def test_write_parquet_splits_from_shards_replaces_stale_partial_file(tmp_path):
+         "image": str(stale_image_path),
+         "env_name": "flappy",
+         "episode_idx": 99,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 99,
+@@ -2452,6 +2471,7 @@ def test_write_parquet_splits_from_shards_replaces_stale_partial_file(tmp_path):
+         "image": str(image_path),
+         "env_name": "flappy",
+         "episode_idx": 0,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 11,
+@@ -2536,6 +2556,8 @@ def test_prepare_starvla_lerobot_dataset_converts_from_raw_export(monkeypatch, t
+         str(raw_export),
+         str(tmp_path / "lerobot"),
+         "demon_attack_train",
++        "--action-carrier",
++        "bridge",
+     ])
+ 
+     assert result == 0
+@@ -2543,7 +2565,7 @@ def test_prepare_starvla_lerobot_dataset_converts_from_raw_export(monkeypatch, t
+         (
+             raw_export,
+             tmp_path / "lerobot" / "demon_attack_train",
+-            {"force": True},
++            {"action_carrier": "bridge", "force": True},
+         )
+     ]
+ 
+diff --git a/tests/data/test_starvla_lerobot_exporter.py b/tests/data/test_starvla_lerobot_exporter.py
+index 64b5f47..58391c7 100644
+--- a/tests/data/test_starvla_lerobot_exporter.py
++++ b/tests/data/test_starvla_lerobot_exporter.py
+@@ -3,6 +3,7 @@ from __future__ import annotations
+ import json
+ 
+ import pyarrow.parquet as pq
++import pytest
+ 
+ from helpers import save_test_image
+ 
+@@ -20,6 +21,7 @@ def test_starvla_lerobot_exporter_reads_canonical_rollout_schema(tmp_path):
+         "image": str(image_path),
+         "env_name": "deadly_corridor",
+         "episode_idx": 3,
++        "trajectory_idx": 0,
+         "decision_step": 0,
+         "split": "train",
+         "seed": 14,
+@@ -38,6 +40,7 @@ def test_starvla_lerobot_exporter_reads_canonical_rollout_schema(tmp_path):
+                 "action_layout": "deadly_corridor_joint_54",
+                 "reward_field": "raw_reward",
+                 "action_labels": list(action_spec.labels),
++                "obs_fps": 8.75,
+             }
+         ),
+         encoding="utf-8",
+@@ -54,10 +57,102 @@ def test_starvla_lerobot_exporter_reads_canonical_rollout_schema(tmp_path):
+     assert table.to_pylist()[0]["done"] is True
+ 
+ 
+-def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+-    """Adjacent LeRobot rows must be adjacent obs decisions (every obs_stride raw
+-    frames), with contiguous frame_index, so KV-memory rollouts unroll at the eval
+-    observation cadence rather than at every raw frame."""
++@pytest.mark.parametrize(
++    ("env_name", "obs_fps", "latency", "action_values", "expected_action", "active_action_dim"),
++    [
++        ("flappy", 30, 3, (1, [0.0, 1.0], "FLAP"), [0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0], 2),
++        (
++            "demon_attack",
++            15,
++            6,
++            (5, [0.0, 0.0, 0.0, 0.0, 0.0, 1.0], "LEFTFIRE"),
++            [0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0],
++            6,
++        ),
++        (
++            "deadly_corridor",
++            8.75,
++            6,
++            (11, None, "MOVE_FORWARD + MOVE_RIGHT + ATTACK"),
++            [1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 1.0],
++            7,
++        ),
++    ],
++)
++def test_starvla_lerobot_exporter_writes_bridge_carrier_with_raw_latency_prompt(
++    tmp_path,
++    env_name,
++    obs_fps,
++    latency,
++    action_values,
++    expected_action,
++    active_action_dim,
++):
++    from latency_bench.data.action_layouts import action_spec_for_env
++    from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
++    from latency_bench.data.hf_parquet import write_parquet_splits
++
++    image_path = tmp_path / "image.png"
++    save_test_image(image_path, 9)
++    raw_export = tmp_path / "raw"
++    action_spec = action_spec_for_env(env_name)
++    action_id, action, action_text = action_values
++    row = {
++        "image": str(image_path),
++        "env_name": env_name,
++        "episode_idx": 0,
++        "trajectory_idx": 0,
++        "decision_step": 0,
++        "split": "train",
++        "seed": 0,
++        "prompt": f"The action latency is fixed at {latency} raw environment frames.",
++        "latency_raw_frames": latency,
++        "latency_ms": latency * 1000 / {"flappy": 30, "demon_attack": 60, "deadly_corridor": 35}[env_name],
++        "raw_reward": 1.0,
++        "action_id": action_id,
++        "action_text": action_text,
++    }
++    if action is not None:
++        row["action"] = action
++    write_parquet_splits(raw_export, [row], [], action_spec=action_spec)
++    (raw_export / "metadata.json").write_text(
++        json.dumps(
++            {
++                "env_name": env_name,
++                "action_layout": action_spec.layout,
++                "reward_field": action_spec.reward_field,
++                "action_labels": list(action_spec.labels),
++                "obs_fps": obs_fps,
++            }
++        ),
++        encoding="utf-8",
++    )
++
++    manifest = convert_canonical_rollout_to_lerobot(
++        raw_export,
++        tmp_path / "lerobot",
++        action_carrier="bridge",
++        force=True,
++    )
++
++    table = pq.read_table(tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000000.parquet")
++    info = json.loads((tmp_path / "lerobot" / "meta" / "info.json").read_text(encoding="utf-8"))
++    task_lines = (tmp_path / "lerobot" / "meta" / "tasks.jsonl").read_text(
++        encoding="utf-8"
++    ).splitlines()
++    tasks = [json.loads(line) for line in task_lines]
++    converted_row = table.to_pylist()[0]
++    assert converted_row["action"] == expected_action
++    assert converted_row["observation.state"] == [0.0] * 7
++    assert info["features"]["action"]["shape"] == [7]
++    assert info["features"]["observation.state"]["shape"] == [7]
++    assert tasks == [{"task_index": 0, "task": row["prompt"]}]
++    assert manifest["action_carrier"] == "bridge"
++    assert manifest["active_action_dim"] == active_action_dim
++
++
++def test_starvla_lerobot_exporter_keeps_every_observation_decision(tmp_path):
++    """Canonical rows already occur at the policy observation cadence."""
+     from latency_bench.data.action_layouts import action_spec_for_env
+     from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
+     from latency_bench.data.hf_parquet import write_parquet_splits
+@@ -66,12 +161,13 @@ def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+     save_test_image(image_path, 9)
+     raw_export = tmp_path / "raw"
+     action_spec = action_spec_for_env("deadly_corridor")
+-    # One episode, 8 contiguous per-raw-frame decision steps; obs_stride = 4.
++    # One trajectory with eight contiguous policy-decision observations.
+     train_rows = [
+         {
+             "image": str(image_path),
+             "env_name": "deadly_corridor",
+             "episode_idx": 0,
++            "trajectory_idx": 0,
+             "decision_step": step,
+             "split": "train",
+             "seed": 7,
+@@ -92,6 +188,7 @@ def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+                 "action_layout": "deadly_corridor_joint_54",
+                 "reward_field": "raw_reward",
+                 "action_labels": list(action_spec.labels),
++                "obs_fps": 8.75,
+                 "obs_stride_raw_frames": 4,
+             }
+         ),
+@@ -102,12 +199,73 @@ def test_starvla_lerobot_exporter_subsamples_to_obs_cadence(tmp_path):
+ 
+     table = pq.read_table(tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000000.parquet")
+     rows = table.to_pylist()
+-    # Only decision_steps 0 and 4 survive; frame_index is contiguous obs index.
+-    assert [r["frame_index"] for r in rows] == [0, 1]
+-    assert [r["episode_index"] for r in rows] == [0, 0]
++    assert [r["frame_index"] for r in rows] == list(range(8))
++    assert [r["episode_index"] for r in rows] == [0] * 8
+     assert rows[-1]["done"] is True
+ 
+ 
++def test_starvla_lerobot_exporter_splits_life_trajectories(tmp_path):
++    from latency_bench.data.action_layouts import action_spec_for_env
++    from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
++    from latency_bench.data.hf_parquet import write_parquet_splits
++
++    image_path = tmp_path / "image.png"
++    save_test_image(image_path, 9)
++    raw_export = tmp_path / "raw"
++    action_spec = action_spec_for_env("demon_attack")
++    train_rows = [
++        {
++            "image": str(image_path),
++            "env_name": "demon_attack",
++            "episode_idx": 0,
++            "trajectory_idx": trajectory_idx,
++            "decision_step": decision_step,
++            "split": "train",
++            "seed": 7,
++            "prompt": "pick action",
++            "latency_raw_frames": 6,
++            "latency_ms": 100.0,
++            "raw_reward": 1.0,
++            "action_id": 0,
++            "action": [1.0, 0.0, 0.0, 0.0, 0.0, 0.0],
++            "action_text": "NOOP",
++        }
++        for trajectory_idx, decision_step in ((0, 0), (0, 1), (1, 2), (1, 3))
++    ]
++    write_parquet_splits(raw_export, train_rows, [], action_spec=action_spec)
++    (raw_export / "metadata.json").write_text(
++        json.dumps(
++            {
++                "env_name": "demon_attack",
++                "action_layout": action_spec.layout,
++                "reward_field": action_spec.reward_field,
++                "action_labels": list(action_spec.labels),
++                "obs_fps": 15,
++            }
++        ),
++        encoding="utf-8",
++    )
++
++    manifest = convert_canonical_rollout_to_lerobot(
++        raw_export,
++        tmp_path / "lerobot",
++        action_carrier="bridge",
++        force=True,
++    )
++
++    first = pq.read_table(
++        tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000000.parquet"
++    ).to_pylist()
++    second = pq.read_table(
++        tmp_path / "lerobot" / "data" / "chunk-000" / "episode_000001.parquet"
++    ).to_pylist()
++    assert manifest["episodes"] == 2
++    assert [row["frame_index"] for row in first] == [0, 1]
++    assert [row["frame_index"] for row in second] == [0, 1]
++    assert [row["done"] for row in first] == [False, True]
++    assert [row["done"] for row in second] == [False, True]
++
++
+ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+     from datasets import Dataset, Features, Image as HFImage, Sequence as HFSequence, Value
+     from latency_bench.data.exporters.starvla_lerobot import convert_canonical_rollout_to_lerobot
+@@ -121,6 +279,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+             "image": HFImage(),
+             "env_name": Value("string"),
+             "episode_idx": Value("int64"),
++            "trajectory_idx": Value("int64"),
+             "decision_step": Value("int64"),
+             "split": Value("string"),
+             "seed": Value("int64"),
+@@ -138,6 +297,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+             "image": [str(image_path)],
+             "env_name": ["flappy"],
+             "episode_idx": [0],
++            "trajectory_idx": [0],
+             "decision_step": [0],
+             "split": ["train"],
+             "seed": [11],
+@@ -156,6 +316,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+             "image": [str(image_path)],
+             "env_name": ["flappy"],
+             "episode_idx": [1],
++            "trajectory_idx": [0],
+             "decision_step": [0],
+             "split": ["val"],
+             "seed": [12],
+@@ -176,6 +337,7 @@ def test_starvla_lerobot_exporter_does_not_accept_latency_alias(tmp_path):
+                 "action_layout": "flappy_discrete",
+                 "reward_field": "raw_reward",
+                 "action_labels": ["NOOP", "FLAP"],
++                "obs_fps": 30,
+             }
+         ),
+         encoding="utf-8",
+diff --git a/tests/policies/test_sample_factory_policy.py b/tests/policies/test_sample_factory_policy.py
+index 0cfe0e1..e61b69c 100644
+--- a/tests/policies/test_sample_factory_policy.py
++++ b/tests/policies/test_sample_factory_policy.py
+@@ -18,6 +18,21 @@ from latency_bench.policy import sample_factory as sample_factory_policy
+ from latency_bench.policy import sample_factory_common
+ 
+ 
++def test_sample_factory_integration_parses_policy_config_for_evaluation():
++    calls = []
++    parsed_cfg = object()
++    module = SimpleNamespace(
++        build_cli_args_from_config=lambda config: ["--experiment", config["experiment"]],
++        parse_cfg=lambda argv, *, evaluation: calls.append((argv, evaluation)) or parsed_cfg,
++    )
++    integration = sf_runtime.SampleFactoryIntegrationModule(module, "register_components")
++
++    result = integration.build_sf_cfg_from_config({"experiment": "teacher"})
++
++    assert result is parsed_cfg
++    assert calls == [(["--experiment", "teacher"], True)]
++
++
+ @pytest.mark.parametrize(
+     ("policy_type", "integration_module_name", "register_attr"),
+     [
+diff --git a/tests/scripts/test_generate_sweep_configs.py b/tests/scripts/test_generate_sweep_configs.py
+index 23d79cb..438b296 100644
+--- a/tests/scripts/test_generate_sweep_configs.py
++++ b/tests/scripts/test_generate_sweep_configs.py
+@@ -23,6 +23,75 @@ def _read_manifest(path: Path) -> list[dict[str, str]]:
+     return [dict(zip(header, line.split("\t"))) for line in lines[1:]]
+ 
+ 
++@pytest.mark.parametrize(
++    ("env_name", "latency_label", "latency_frames", "obs_fps", "obs_fps_label"),
++    [
++        ("flappy", "fixed_l3", 3, 30, "30"),
++        ("demon_attack", "fixed_l6", 6, 15, "15"),
++        ("deadly_corridor", "fixed_l6", 6, 8.75, "8p75"),
++    ],
++)
++def test_memory_rollout_teacher_spec_emits_single_uploadable_job(
++    tmp_path,
++    env_name,
++    latency_label,
++    latency_frames,
++    obs_fps,
++    obs_fps_label,
++):
++    generated_root = tmp_path / "configs"
++    output_root = tmp_path / "outputs"
++    train_dir = tmp_path / "checkpoints"
++
++    generate_sweep_configs(
++        [
++            "--spec",
++            f"configs/experiments/memory/rollout_teacher/{env_name}.yaml",
++            "--generated-root",
++            str(generated_root),
++            "--output-root",
++            str(output_root),
++            "--train-dir",
++            str(train_dir),
++        ]
++    )
++
++    rows = _read_manifest(generated_root / "manifest.tsv")
++    assert len(rows) == 1
++    row = rows[0]
++    assert row["env"] == env_name
++    assert row["experiment_family"] == "memory_rollout_teacher"
++    assert row["sweep_name"] == "teacher"
++    assert row["seed"] == "0"
++    assert row["frame_stack"] == "4"
++    assert row["obs_fps"] == obs_fps_label
++    assert row["latency_label"] == latency_label
++    assert row["latency_kind"] == "fixed"
++    assert row["latency_fixed_frames"] == str(latency_frames)
++    assert row["train_upload_repo_id"] == "latency-sensitive-bench/paper-experiment-models"
++    assert row["train_upload_repo_type"] == "model"
++    assert row["eval_upload_repo_id"] == ""
++    assert row["train_artifact_repo_path"].startswith(
++        f"memory/rollout_teachers/{env_name}/train/"
++    )
++
++    train_cfg = yaml.safe_load(Path(row["train_config"]).read_text(encoding="utf-8"))
++    assert train_cfg["experiment"]["seed"] == 0
++    assert train_cfg["env"]["frame_stack"] == 4
++    assert train_cfg["env"]["obs_fps"] == pytest.approx(obs_fps)
++    if env_name == "demon_attack":
++        assert train_cfg["env"]["reset_mode"] == "fire"
++    assert train_cfg["latency"]["method"] == "fixed"
++    assert train_cfg["latency"]["fixed_latency_ms"] == pytest.approx(
++        latency_frames * 1000 / train_cfg["env"]["env_fps"]
++    )
++    assert train_cfg["evaluation"]["eval_suites"] == {
++        "fixed": [[latency_frames]],
++        "normal": [],
++        "uniform": [],
++    }
++
++
+ def test_latency_bench_eval_config_disables_full_trace_records(tmp_path):
+     base_config = {
+         "backend": {
+diff --git a/tests/scripts/test_manifest_jobs.py b/tests/scripts/test_manifest_jobs.py
+index bd7bf03..c9dd3c6 100644
+--- a/tests/scripts/test_manifest_jobs.py
++++ b/tests/scripts/test_manifest_jobs.py
+@@ -1,8 +1,10 @@
+ from __future__ import annotations
+ 
++import json
+ import os
+-import subprocess
++import shlex
+ import signal
++import subprocess
+ import sys
+ import threading
+ import time
+@@ -10,6 +12,7 @@ import types
+ from pathlib import Path
+ 
+ import pytest
++from helpers import REPO_ROOT
+ 
+ import scripts.experiment_runner.run_manifest_jobs as run_manifest_jobs
+ from scripts.experiment_runner.job_manifest import (
+@@ -68,6 +71,39 @@ def test_train_stage_only_trains_and_resumes_from_success(tmp_path):
+     assert [event["stage"] for event in events] == ["train_started", "train_succeeded"]
+ 
+ 
++@pytest.mark.parametrize(
++    ("recorded_stage", "expected_stage"),
++    [
++        ("train_started", "train"),
++        ("train_succeeded", "eval"),
++        ("eval_started", "eval"),
++        ("eval_failed", "eval"),
++        ("eval_succeeded", "upload"),
++        ("upload_started", "upload"),
++    ],
++)
++def test_resume_stage_keeps_interrupted_job_at_its_current_pipeline_stage(
++    tmp_path,
++    recorded_stage,
++    expected_stage,
++):
++    train_artifact_dir = tmp_path / "train_artifact"
++    train_artifact_dir.mkdir()
++    (train_artifact_dir / "config.json").write_text("{}")
++    assert run_manifest_jobs._resume_stage(
++        {"job_id": "job1", "train_artifact_dir": str(train_artifact_dir)},
++        {"job1": recorded_stage},
++    ) == expected_stage
++
++
++@pytest.mark.parametrize("recorded_stage", ["train_succeeded", "eval_started", "eval_failed"])
++def test_resume_stage_retrains_when_train_artifact_was_deleted(tmp_path, recorded_stage):
++    assert run_manifest_jobs._resume_stage(
++        {"job_id": "job1", "train_artifact_dir": str(tmp_path / "deleted_train_artifact")},
++        {"job1": recorded_stage},
++    ) == "train"
++
++
+ def test_eval_stage_does_not_train_or_upload(tmp_path):
+     manifest_path, train_mark, eval_mark = _write_single_job(tmp_path)
+     uploads: list[str] = []
+@@ -134,6 +170,26 @@ def test_upload_stage_needs_no_gpu_and_retries_failed_upload(tmp_path):
+     ]
+ 
+ 
++def test_upload_stage_resumes_interrupted_upload(tmp_path):
++    manifest_path, _, _ = _write_single_job(tmp_path)
++    append_event(
++        event_log_path_for_manifest(manifest_path),
++        {"job_id": "job1", "stage": "upload_started", "ts": 1.0},
++    )
++    uploads: list[str] = []
++
++    assert run_jobs(
++        manifest_path=manifest_path,
++        gpus=[],
++        upload_workers=1,
++        max_upload_attempts=1,
++        stage="upload",
++        uploader=lambda job: uploads.append(job["job_id"]),
++    ) == 0
++
++    assert uploads == ["job1"]
++
++
+ def test_dry_run_does_not_execute_or_record_success(tmp_path, capsys):
+     manifest_path, train_mark, eval_mark = _write_single_job(tmp_path)
+ 
+@@ -310,6 +366,186 @@ def test_run_command_cleanup_terminates_child_process_group(tmp_path):
+         os.kill(child_pid, 0)
+ 
+ 
++def test_sigint_cancels_queue_and_kills_all_active_process_groups(tmp_path):
++    helper_path = tmp_path / "ignore_term_job.py"
++    helper_path.write_text(
++        """from __future__ import annotations
++
++import json
++import os
++from pathlib import Path
++import signal
++import subprocess
++import sys
++import time
++
++signal.signal(signal.SIGTERM, signal.SIG_IGN)
++child = subprocess.Popen(
++    [
++        sys.executable,
++        "-c",
++        "import signal, time; signal.signal(signal.SIGTERM, signal.SIG_IGN); time.sleep(60)",
++    ]
++)
++Path(sys.argv[1]).write_text(
++    json.dumps(
++        {
++            "leader_pid": os.getpid(),
++            "leader_pgid": os.getpgid(0),
++            "child_pid": child.pid,
++            "child_pgid": os.getpgid(child.pid),
++        }
++    ),
++    encoding="utf-8",
++)
++time.sleep(60)
++""",
++        encoding="utf-8",
++    )
++    manifest_path = tmp_path / "manifest.tsv"
++    markers = [tmp_path / f"job_{idx}.json" for idx in range(4)]
++    rows = []
++    for idx, marker in enumerate(markers):
++        row = {column: "" for column in MANIFEST_V2_COLUMNS}
++        row.update(
++            {
++                "schema_version": "v2",
++                "job_id": f"job{idx}",
++                "train_command": f"exec {shlex.join([sys.executable, str(helper_path), str(marker)])}",
++            }
++        )
++        rows.append(row)
++    write_manifest(manifest_path, rows)
++
++    runner = subprocess.Popen(
++        [
++            sys.executable,
++            "-m",
++            "scripts.experiment_runner.run_manifest_jobs",
++            "--manifest",
++            str(manifest_path),
++            "--stage",
++            "train",
++            "--gpus",
++            "0,1",
++        ],
++        cwd=REPO_ROOT,
++        stdout=subprocess.PIPE,
++        stderr=subprocess.PIPE,
++        text=True,
++    )
++    process_groups: set[int] = set()
++    try:
++        deadline = time.time() + 10.0
++        while sum(marker.is_file() for marker in markers) < 2 and time.time() < deadline:
++            time.sleep(0.05)
++        assert [marker.is_file() for marker in markers] == [True, True, False, False]
++        for marker in markers[:2]:
++            process_info = json.loads(marker.read_text(encoding="utf-8"))
++            assert process_info["leader_pgid"] == process_info["child_pgid"]
++            process_groups.add(int(process_info["leader_pgid"]))
++
++        runner.send_signal(signal.SIGINT)
++        time.sleep(0.1)
++        runner.send_signal(signal.SIGINT)
++        stdout, stderr = runner.communicate(timeout=10.0)
++        assert runner.returncode != 0, (stdout, stderr)
++        assert [marker.is_file() for marker in markers] == [True, True, False, False]
++
++        deadline = time.time() + 5.0
++        while process_groups and time.time() < deadline:
++            for process_group in list(process_groups):
++                try:
++                    os.killpg(process_group, 0)
++                except ProcessLookupError:
++                    process_groups.remove(process_group)
++            time.sleep(0.05)
++        assert not process_groups
++
++        events = load_events(event_log_path_for_manifest(manifest_path))
++        assert [event["stage"] for event in events] == ["train_started", "train_started"]
++    finally:
++        if runner.poll() is None:
++            runner.kill()
++            runner.wait()
++        for process_group in process_groups:
++            try:
++                os.killpg(process_group, signal.SIGKILL)
++            except ProcessLookupError:
++                pass
++
++
++def test_sigint_does_not_wait_for_in_flight_upload(tmp_path):
++    manifest_path = tmp_path / "manifest.tsv"
++    row = {column: "" for column in MANIFEST_V2_COLUMNS}
++    row.update(
++        {
++            "schema_version": "v2",
++            "job_id": "job0",
++            "train_upload_repo_id": "repo-model",
++            "train_upload_repo_type": "model",
++        }
++    )
++    write_manifest(manifest_path, [row])
++    append_event(
++        event_log_path_for_manifest(manifest_path),
++        {"job_id": "job0", "stage": "eval_succeeded"},
++    )
++    upload_started = tmp_path / "upload_started"
++    helper_path = tmp_path / "blocking_upload_runner.py"
++    helper_path.write_text(
++        """from pathlib import Path
++import sys
++import time
++
++from scripts.experiment_runner.run_manifest_jobs import run_jobs
++
++manifest_path = Path(sys.argv[1])
++upload_started = Path(sys.argv[2])
++
++def blocking_uploader(job):
++    del job
++    upload_started.write_text("started", encoding="utf-8")
++    time.sleep(60)
++
++raise SystemExit(
++    run_jobs(
++        manifest_path=manifest_path,
++        gpus=[],
++        upload_workers=1,
++        max_upload_attempts=1,
++        stage="upload",
++        uploader=blocking_uploader,
++    )
++)
++""",
++        encoding="utf-8",
++    )
++    runner = subprocess.Popen(
++        [sys.executable, str(helper_path), str(manifest_path), str(upload_started)],
++        cwd=REPO_ROOT,
++        env={**os.environ, "PYTHONPATH": str(REPO_ROOT)},
++        stdout=subprocess.PIPE,
++        stderr=subprocess.PIPE,
++        text=True,
++    )
++    try:
++        deadline = time.time() + 10.0
++        while not upload_started.is_file() and time.time() < deadline:
++            time.sleep(0.05)
++        assert upload_started.is_file()
++
++        started_at = time.monotonic()
++        runner.send_signal(signal.SIGINT)
++        stdout, stderr = runner.communicate(timeout=5.0)
++        assert runner.returncode != 0, (stdout, stderr)
++        assert time.monotonic() - started_at < 3.0
++    finally:
++        if runner.poll() is None:
++            runner.kill()
++            runner.wait()
++
++
+ def test_default_uploader_prefers_env_token_over_dotenv(monkeypatch, tmp_path):
+     monkeypatch.chdir(tmp_path)
+     (tmp_path / ".env").write_text("HF_TOKEN=dotenv-token\n", encoding="utf-8")
+diff --git a/training/deadly_corridor_sf/integration.py b/training/deadly_corridor_sf/integration.py
+index 08fc63e..6405b7c 100644
+--- a/training/deadly_corridor_sf/integration.py
++++ b/training/deadly_corridor_sf/integration.py
+@@ -287,7 +287,7 @@ def add_latency_args(parser: argparse.ArgumentParser) -> None:
+     parser.add_argument("--vizdoom_init_retry_jitter_sec", type=float, default=0.25)
+ 
+ 
+-def parse_cfg(argv=None):
++def parse_cfg(argv=None, *, evaluation: bool = True):
+     from sample_factory.cfg.arguments import parse_full_cfg, parse_sf_args
+     from sf_examples.vizdoom.doom.doom_params import (
+         add_doom_env_args,
+@@ -295,7 +295,7 @@ def parse_cfg(argv=None):
+         doom_override_defaults,
+     )
+ 
+-    parser, _ = parse_sf_args(argv=argv, evaluation=True)
++    parser, _ = parse_sf_args(argv=argv, evaluation=evaluation)
+     add_doom_env_args(parser)
+     add_doom_env_eval_args(parser)
+     add_latency_args(parser)
+diff --git a/training/demon_attack_sf/integration.py b/training/demon_attack_sf/integration.py
+index 1ee9060..9097048 100644
+--- a/training/demon_attack_sf/integration.py
++++ b/training/demon_attack_sf/integration.py
+@@ -318,10 +318,10 @@ def demon_attack_override_defaults(parser: argparse.ArgumentParser) -> None:
+     )
+ 
+ 
+-def parse_cfg(argv=None):
++def parse_cfg(argv=None, *, evaluation: bool = False):
+     from sample_factory.cfg.arguments import parse_full_cfg, parse_sf_args
+ 
+-    parser, _ = parse_sf_args(argv=argv, evaluation=False)
++    parser, _ = parse_sf_args(argv=argv, evaluation=evaluation)
+     add_demon_attack_env_args(parser)
+     demon_attack_override_defaults(parser)
+     cfg = parse_full_cfg(parser, argv)
+diff --git a/training/flappy_sf/integration.py b/training/flappy_sf/integration.py
+index 6f666be..c342ba7 100644
+--- a/training/flappy_sf/integration.py
++++ b/training/flappy_sf/integration.py
+@@ -541,10 +541,10 @@ def flappy_override_defaults(parser: argparse.ArgumentParser) -> None:
+     )
+ 
+ 
+-def parse_cfg(argv=None):
++def parse_cfg(argv=None, *, evaluation: bool = False):
+     from sample_factory.cfg.arguments import parse_full_cfg, parse_sf_args
+ 
+-    parser, _ = parse_sf_args(argv=argv, evaluation=False)
++    parser, _ = parse_sf_args(argv=argv, evaluation=evaluation)
+     add_flappy_env_args(parser)
+     flappy_override_defaults(parser)
+     cfg = parse_full_cfg(parser, argv)

evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 42, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 1, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 43, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 2, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 44, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 3, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 45, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 4, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 46, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 5, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 47, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 6, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 48, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 7, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 49, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 8, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 50, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 9, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 51, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 10, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 52, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 11, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 53, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 12, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 54, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 13, "episode_return": 132.80000157654285, "episode_return_env": 132.80000157654285, "game_score": 28.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 55, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 1087, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 1087}
+{"episode_id": 14, "episode_return": 370.80000436306, "episode_return_env": 370.80000436306, "game_score": 79.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 56, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3008, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3008}
+{"episode_id": 15, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 57, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 16, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 58, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 17, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 59, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 18, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 60, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 19, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_raw3", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 61, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_raw3", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}

evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/queue_eval_latency_3.json

@@ -0,0 +1,59 @@
+{
+  "checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt",
+  "configured_latency_ms": 100.0,
+  "eval_latency_steps": 3,
+  "experiment_name": "starvla_wan_oft_flappy_raw3",
+  "latency": 3,
+  "latency_type": "fixed_sweep",
+  "lengths": [
+    3600,
+    3600,
+    3600,
+    3600,
+    3600,
+    3600,
+    3600,
+    3600,
+    3600,
+    3600,
+    3600,
+    3600,
+    3600,
+    1087,
+    3008,
+    3600,
+    3600,
+    3600,
+    3600,
+    3600
+  ],
+  "mean_length": 3444.75,
+  "mean_return": 425.32000499889256,
+  "returns": [
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    132.80000157654285,
+    370.80000436306,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471
+  ],
+  "seed": 42,
+  "std_return": 69.00409923621429,
+  "suite_name": "fixed_3",
+  "timestamp_utc": "2026-07-29T22:31:52.064243+00:00",
+  "train_latency_steps": 3
+}

evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/queue_eval_results.jsonl

@@ -0,0 +1 @@
+{"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "configured_latency_ms": 100.0, "eval_latency_steps": 3, "experiment_name": "starvla_wan_oft_flappy_raw3", "latency": 3, "latency_type": "fixed_sweep", "lengths": [3600, 3600, 3600, 3600, 3600, 3600, 3600, 3600, 3600, 3600, 3600, 3600, 3600, 1087, 3008, 3600, 3600, 3600, 3600, 3600], "mean_length": 3444.75, "mean_return": 425.32000499889256, "returns": [444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 132.80000157654285, 370.80000436306, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471], "seed": 42, "std_return": 69.00409923621429, "suite_name": "fixed_3", "timestamp_utc": "2026-07-29T22:31:52.064243+00:00", "train_latency_steps": 3}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/eval/post_train/step_5000.json

@@ -0,0 +1,318 @@
+{
+  "per_latency": {
+    "gymnasium/latency_0": {
+      "latency": 0,
+      "num_episodes": 20,
+      "mean_reward": 87.4,
+      "mean_length": 87.4,
+      "std_reward": 69.0328907116021,
+      "std_length": 69.0328907116021,
+      "episode_rewards": [
+        72.0,
+        276.0,
+        158.0,
+        31.0,
+        59.0,
+        28.0,
+        31.0,
+        50.0,
+        31.0,
+        55.0,
+        228.0,
+        109.0,
+        59.0,
+        22.0,
+        90.0,
+        45.0,
+        33.0,
+        112.0,
+        174.0,
+        85.0
+      ],
+      "episode_lengths": [
+        72,
+        276,
+        158,
+        31,
+        59,
+        28,
+        31,
+        50,
+        31,
+        55,
+        228,
+        109,
+        59,
+        22,
+        90,
+        45,
+        33,
+        112,
+        174,
+        85
+      ],
+      "decoded_action_hist": {},
+      "fixed_episode_seeds": true,
+      "eval_seed": 42,
+      "episode_seeds": [
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null
+      ],
+      "episode_indices": [
+        0,
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19
+      ]
+    },
+    "gymnasium/latency_2": {
+      "latency": 2,
+      "num_episodes": 20,
+      "mean_reward": 76.95,
+      "mean_length": 76.95,
+      "std_reward": 23.97389204947749,
+      "std_length": 23.97389204947749,
+      "episode_rewards": [
+        92.0,
+        36.0,
+        53.0,
+        86.0,
+        56.0,
+        98.0,
+        82.0,
+        83.0,
+        98.0,
+        46.0,
+        120.0,
+        83.0,
+        80.0,
+        100.0,
+        45.0,
+        112.0,
+        97.0,
+        66.0,
+        42.0,
+        64.0
+      ],
+      "episode_lengths": [
+        92,
+        36,
+        53,
+        86,
+        56,
+        98,
+        82,
+        83,
+        98,
+        46,
+        120,
+        83,
+        80,
+        100,
+        45,
+        112,
+        97,
+        66,
+        42,
+        64
+      ],
+      "decoded_action_hist": {},
+      "fixed_episode_seeds": true,
+      "eval_seed": 42,
+      "episode_seeds": [
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null
+      ],
+      "episode_indices": [
+        0,
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19
+      ]
+    },
+    "gymnasium/latency_4": {
+      "latency": 4,
+      "num_episodes": 20,
+      "mean_reward": 27.45,
+      "mean_length": 27.45,
+      "std_reward": 4.271709259769443,
+      "std_length": 4.271709259769443,
+      "episode_rewards": [
+        37.0,
+        22.0,
+        29.0,
+        31.0,
+        22.0,
+        28.0,
+        25.0,
+        30.0,
+        25.0,
+        26.0,
+        30.0,
+        26.0,
+        33.0,
+        26.0,
+        22.0,
+        36.0,
+        27.0,
+        26.0,
+        22.0,
+        26.0
+      ],
+      "episode_lengths": [
+        37,
+        22,
+        29,
+        31,
+        22,
+        28,
+        25,
+        30,
+        25,
+        26,
+        30,
+        26,
+        33,
+        26,
+        22,
+        36,
+        27,
+        26,
+        22,
+        26
+      ],
+      "decoded_action_hist": {},
+      "fixed_episode_seeds": true,
+      "eval_seed": 42,
+      "episode_seeds": [
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null
+      ],
+      "episode_indices": [
+        0,
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19
+      ]
+    }
+  },
+  "aggregate": {
+    "stage": "post_train",
+    "step": 5000,
+    "task": "gymnasium",
+    "model_alias": "openvla",
+    "fixed_episode_seeds": true,
+    "eval_seed": 42,
+    "total_episodes": 60,
+    "mean_reward": 63.93333333333333,
+    "mean_length": 63.93333333333333,
+    "std_reward": 49.69804377996739,
+    "std_length": 49.69804377996739,
+    "task_count": 1,
+    "macro_mean_reward": 63.93333333333334,
+    "macro_mean_length": 63.93333333333334,
+    "distributed_eval": false
+  }
+}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 72.0, "episode_return_env": 72.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 42, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 72, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 72}
+{"episode_id": 1, "episode_return": 276.0, "episode_return_env": 276.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 43, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 276, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 276}
+{"episode_id": 2, "episode_return": 158.0, "episode_return_env": 158.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 44, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 158, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 158}
+{"episode_id": 3, "episode_return": 31.0, "episode_return_env": 31.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 45, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 31, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 31}
+{"episode_id": 4, "episode_return": 59.0, "episode_return_env": 59.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 46, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 59, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 59}
+{"episode_id": 5, "episode_return": 28.0, "episode_return_env": 28.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 47, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 28, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 28}
+{"episode_id": 6, "episode_return": 31.0, "episode_return_env": 31.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 48, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 31, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 31}
+{"episode_id": 7, "episode_return": 50.0, "episode_return_env": 50.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 49, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 50, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 50}
+{"episode_id": 8, "episode_return": 31.0, "episode_return_env": 31.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 50, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 31, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 31}
+{"episode_id": 9, "episode_return": 55.0, "episode_return_env": 55.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 51, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 55, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 55}
+{"episode_id": 10, "episode_return": 228.0, "episode_return_env": 228.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 52, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 228, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 228}
+{"episode_id": 11, "episode_return": 109.0, "episode_return_env": 109.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 53, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 109, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 109}
+{"episode_id": 12, "episode_return": 59.0, "episode_return_env": 59.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 54, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 59, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 59}
+{"episode_id": 13, "episode_return": 22.0, "episode_return_env": 22.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 55, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 22}
+{"episode_id": 14, "episode_return": 90.0, "episode_return_env": 90.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 56, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 90, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 90}
+{"episode_id": 15, "episode_return": 45.0, "episode_return_env": 45.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 57, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 45, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 45}
+{"episode_id": 16, "episode_return": 33.0, "episode_return_env": 33.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 58, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 33, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 33}
+{"episode_id": 17, "episode_return": 112.0, "episode_return_env": 112.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 59, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 112, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 112}
+{"episode_id": 18, "episode_return": 174.0, "episode_return_env": 174.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 60, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 174, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 174}
+{"episode_id": 19, "episode_return": 85.0, "episode_return_env": 85.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 61, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 85, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 85}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/resolved_config.yaml

@@ -0,0 +1,138 @@
+experiment:
+  name: starvla_gymnasium_train_eval
+  seed: 42
+backend:
+  run_mode: eval
+executor:
+  mode: simulated
+  simulated_inference_pool: false
+  simulated_worker_capacity: null
+env:
+  name: gymnasium
+  task_name: inverted_pendulum_rgb_state
+  env_id: LatencyBench/InvertedPendulumDiscrete-v0
+  registration_imports:
+  - latency_bench.envs.gymnasium_mujoco
+  env_fps: 25.0
+  obs_fps: 25.0
+  frame_stack: 1
+  base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose exactly
+    one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+    push_right_hard.'
+  action_labels:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_values:
+  - 0
+  - 1
+  - 2
+  - 3
+  - 4
+  noop_action_id: 2
+  make_kwargs:
+    base_env_id: InvertedPendulum-v4
+    render_mode: rgb_array
+    force_values:
+    - -3.0
+    - -1.5
+    - 0.0
+    - 1.5
+    - 3.0
+    base_make_kwargs: {}
+  state_space:
+    labels:
+    - cart_position
+    - pole_angle
+    - cart_velocity
+    - pole_angular_velocity
+  action_order:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_map:
+    push_left_hard: 0
+    push_left_soft: 1
+    zero_force: 2
+    push_right_soft: 3
+    push_right_hard: 4
+  noop_action: zero_force
+  obs_resize:
+  - 224
+  - 224
+latency:
+  method: zero
+  add_latency_info: false
+scheduler:
+  hold_policy: hold
+  ordering_policy: issue_order_fifo
+policy:
+  type: starvla
+  checkpoint_path: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0
+  device: cuda
+  task_manifest_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/manifest.json
+  gymnasium_training_task_contract:
+    task_name: inverted_pendulum_rgb_state
+    env_id: LatencyBench/InvertedPendulumDiscrete-v0
+    registration_imports:
+    - latency_bench.envs.gymnasium_mujoco
+    env_fps: 25.0
+    obs_fps: 25.0
+    frame_stack: 1
+    base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose
+      exactly one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+      push_right_hard.'
+    action_labels:
+    - push_left_hard
+    - push_left_soft
+    - zero_force
+    - push_right_soft
+    - push_right_hard
+    action_values:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    noop_action_id: 2
+    make_kwargs:
+      base_env_id: InvertedPendulum-v4
+      render_mode: rgb_array
+      force_values:
+      - -3.0
+      - -1.5
+      - 0.0
+      - 1.5
+      - 3.0
+      base_make_kwargs: {}
+    state_space:
+      labels:
+      - cart_position
+      - pole_angle
+      - cart_velocity
+      - pole_angular_velocity
+  latency_prompt_map_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_prompt_map.json
+  latency_prompt_key: 0
+  action_prefix:
+    mode: none
+evaluation:
+  eval_episodes: 20
+  eval_parallel_envs: 5
+  eval_max_steps: 3600
+  eval_max_decision_steps: 3600
+  eval_deterministic: true
+  eval_raw_reward: false
+  eval_latency_values: null
+logging:
+  output_dir: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0
+  video:
+    enabled: false
+    num_bins: 1
+  save_step_records: false
+  save_action_records: false
+  save_latency_records: false
+  simulated_pipeline_profile: false

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 92.0, "episode_return_env": 92.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 42, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 92, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 92}
+{"episode_id": 1, "episode_return": 36.0, "episode_return_env": 36.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 43, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 36, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 36}
+{"episode_id": 2, "episode_return": 53.0, "episode_return_env": 53.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 44, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 53, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 53}
+{"episode_id": 3, "episode_return": 86.0, "episode_return_env": 86.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 45, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 86, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 86}
+{"episode_id": 4, "episode_return": 56.0, "episode_return_env": 56.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 46, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 56, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 56}
+{"episode_id": 5, "episode_return": 98.0, "episode_return_env": 98.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 47, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 98, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 98}
+{"episode_id": 6, "episode_return": 82.0, "episode_return_env": 82.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 48, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 82, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 82}
+{"episode_id": 7, "episode_return": 83.0, "episode_return_env": 83.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 49, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 83, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 83}
+{"episode_id": 8, "episode_return": 98.0, "episode_return_env": 98.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 50, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 98, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 98}
+{"episode_id": 9, "episode_return": 46.0, "episode_return_env": 46.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 51, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 46, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 46}
+{"episode_id": 10, "episode_return": 120.0, "episode_return_env": 120.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 52, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 120, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 120}
+{"episode_id": 11, "episode_return": 83.0, "episode_return_env": 83.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 53, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 83, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 83}
+{"episode_id": 12, "episode_return": 80.0, "episode_return_env": 80.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 54, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 80, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 80}
+{"episode_id": 13, "episode_return": 100.0, "episode_return_env": 100.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 55, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 100, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 100}
+{"episode_id": 14, "episode_return": 45.0, "episode_return_env": 45.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 56, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 45, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 45}
+{"episode_id": 15, "episode_return": 112.0, "episode_return_env": 112.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 57, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 112, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 112}
+{"episode_id": 16, "episode_return": 97.0, "episode_return_env": 97.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 58, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 97, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 97}
+{"episode_id": 17, "episode_return": 66.0, "episode_return_env": 66.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 59, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 66, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 66}
+{"episode_id": 18, "episode_return": 42.0, "episode_return_env": 42.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 60, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 42, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 42}
+{"episode_id": 19, "episode_return": 64.0, "episode_return_env": 64.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 61, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 64, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 64}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/resolved_config.yaml

@@ -0,0 +1,139 @@
+experiment:
+  name: starvla_gymnasium_train_eval
+  seed: 42
+backend:
+  run_mode: eval
+executor:
+  mode: simulated
+  simulated_inference_pool: false
+  simulated_worker_capacity: null
+env:
+  name: gymnasium
+  task_name: inverted_pendulum_rgb_state
+  env_id: LatencyBench/InvertedPendulumDiscrete-v0
+  registration_imports:
+  - latency_bench.envs.gymnasium_mujoco
+  env_fps: 25.0
+  obs_fps: 25.0
+  frame_stack: 1
+  base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose exactly
+    one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+    push_right_hard.'
+  action_labels:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_values:
+  - 0
+  - 1
+  - 2
+  - 3
+  - 4
+  noop_action_id: 2
+  make_kwargs:
+    base_env_id: InvertedPendulum-v4
+    render_mode: rgb_array
+    force_values:
+    - -3.0
+    - -1.5
+    - 0.0
+    - 1.5
+    - 3.0
+    base_make_kwargs: {}
+  state_space:
+    labels:
+    - cart_position
+    - pole_angle
+    - cart_velocity
+    - pole_angular_velocity
+  action_order:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_map:
+    push_left_hard: 0
+    push_left_soft: 1
+    zero_force: 2
+    push_right_soft: 3
+    push_right_hard: 4
+  noop_action: zero_force
+  obs_resize:
+  - 224
+  - 224
+latency:
+  method: fixed
+  add_latency_info: false
+  fixed_latency_ms: 80.0
+scheduler:
+  hold_policy: hold
+  ordering_policy: issue_order_fifo
+policy:
+  type: starvla
+  checkpoint_path: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2
+  device: cuda
+  task_manifest_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/manifest.json
+  gymnasium_training_task_contract:
+    task_name: inverted_pendulum_rgb_state
+    env_id: LatencyBench/InvertedPendulumDiscrete-v0
+    registration_imports:
+    - latency_bench.envs.gymnasium_mujoco
+    env_fps: 25.0
+    obs_fps: 25.0
+    frame_stack: 1
+    base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose
+      exactly one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+      push_right_hard.'
+    action_labels:
+    - push_left_hard
+    - push_left_soft
+    - zero_force
+    - push_right_soft
+    - push_right_hard
+    action_values:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    noop_action_id: 2
+    make_kwargs:
+      base_env_id: InvertedPendulum-v4
+      render_mode: rgb_array
+      force_values:
+      - -3.0
+      - -1.5
+      - 0.0
+      - 1.5
+      - 3.0
+      base_make_kwargs: {}
+    state_space:
+      labels:
+      - cart_position
+      - pole_angle
+      - cart_velocity
+      - pole_angular_velocity
+  latency_prompt_map_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_prompt_map.json
+  latency_prompt_key: 2
+  action_prefix:
+    mode: none
+evaluation:
+  eval_episodes: 20
+  eval_parallel_envs: 5
+  eval_max_steps: 3600
+  eval_max_decision_steps: 3600
+  eval_deterministic: true
+  eval_raw_reward: false
+  eval_latency_values: null
+logging:
+  output_dir: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2
+  video:
+    enabled: false
+    num_bins: 1
+  save_step_records: false
+  save_action_records: false
+  save_latency_records: false
+  simulated_pipeline_profile: false

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 37.0, "episode_return_env": 37.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 42, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 37, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 37}
+{"episode_id": 1, "episode_return": 22.0, "episode_return_env": 22.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 43, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 22}
+{"episode_id": 2, "episode_return": 29.0, "episode_return_env": 29.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 44, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 29, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 29}
+{"episode_id": 3, "episode_return": 31.0, "episode_return_env": 31.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 45, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 31, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 31}
+{"episode_id": 4, "episode_return": 22.0, "episode_return_env": 22.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 46, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 22}
+{"episode_id": 5, "episode_return": 28.0, "episode_return_env": 28.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 47, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 28, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 28}
+{"episode_id": 6, "episode_return": 25.0, "episode_return_env": 25.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 48, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 25, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 25}
+{"episode_id": 7, "episode_return": 30.0, "episode_return_env": 30.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 49, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 30, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 30}
+{"episode_id": 8, "episode_return": 25.0, "episode_return_env": 25.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 50, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 25, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 25}
+{"episode_id": 9, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 51, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 26}
+{"episode_id": 10, "episode_return": 30.0, "episode_return_env": 30.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 52, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 30, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 30}
+{"episode_id": 11, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 53, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 26}
+{"episode_id": 12, "episode_return": 33.0, "episode_return_env": 33.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 54, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 33, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 33}
+{"episode_id": 13, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 55, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 26}
+{"episode_id": 14, "episode_return": 22.0, "episode_return_env": 22.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 56, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 22}
+{"episode_id": 15, "episode_return": 36.0, "episode_return_env": 36.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 57, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 36, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 36}
+{"episode_id": 16, "episode_return": 27.0, "episode_return_env": 27.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 58, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 27, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 27}
+{"episode_id": 17, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 59, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 26}
+{"episode_id": 18, "episode_return": 22.0, "episode_return_env": 22.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 60, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 22}
+{"episode_id": 19, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 61, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 26}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/resolved_config.yaml

@@ -0,0 +1,139 @@
+experiment:
+  name: starvla_gymnasium_train_eval
+  seed: 42
+backend:
+  run_mode: eval
+executor:
+  mode: simulated
+  simulated_inference_pool: false
+  simulated_worker_capacity: null
+env:
+  name: gymnasium
+  task_name: inverted_pendulum_rgb_state
+  env_id: LatencyBench/InvertedPendulumDiscrete-v0
+  registration_imports:
+  - latency_bench.envs.gymnasium_mujoco
+  env_fps: 25.0
+  obs_fps: 25.0
+  frame_stack: 1
+  base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose exactly
+    one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+    push_right_hard.'
+  action_labels:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_values:
+  - 0
+  - 1
+  - 2
+  - 3
+  - 4
+  noop_action_id: 2
+  make_kwargs:
+    base_env_id: InvertedPendulum-v4
+    render_mode: rgb_array
+    force_values:
+    - -3.0
+    - -1.5
+    - 0.0
+    - 1.5
+    - 3.0
+    base_make_kwargs: {}
+  state_space:
+    labels:
+    - cart_position
+    - pole_angle
+    - cart_velocity
+    - pole_angular_velocity
+  action_order:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_map:
+    push_left_hard: 0
+    push_left_soft: 1
+    zero_force: 2
+    push_right_soft: 3
+    push_right_hard: 4
+  noop_action: zero_force
+  obs_resize:
+  - 224
+  - 224
+latency:
+  method: fixed
+  add_latency_info: false
+  fixed_latency_ms: 160.0
+scheduler:
+  hold_policy: hold
+  ordering_policy: issue_order_fifo
+policy:
+  type: starvla
+  checkpoint_path: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4
+  device: cuda
+  task_manifest_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/manifest.json
+  gymnasium_training_task_contract:
+    task_name: inverted_pendulum_rgb_state
+    env_id: LatencyBench/InvertedPendulumDiscrete-v0
+    registration_imports:
+    - latency_bench.envs.gymnasium_mujoco
+    env_fps: 25.0
+    obs_fps: 25.0
+    frame_stack: 1
+    base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose
+      exactly one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+      push_right_hard.'
+    action_labels:
+    - push_left_hard
+    - push_left_soft
+    - zero_force
+    - push_right_soft
+    - push_right_hard
+    action_values:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    noop_action_id: 2
+    make_kwargs:
+      base_env_id: InvertedPendulum-v4
+      render_mode: rgb_array
+      force_values:
+      - -3.0
+      - -1.5
+      - 0.0
+      - 1.5
+      - 3.0
+      base_make_kwargs: {}
+    state_space:
+      labels:
+      - cart_position
+      - pole_angle
+      - cart_velocity
+      - pole_angular_velocity
+  latency_prompt_map_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_prompt_map.json
+  latency_prompt_key: 4
+  action_prefix:
+    mode: none
+evaluation:
+  eval_episodes: 20
+  eval_parallel_envs: 5
+  eval_max_steps: 3600
+  eval_max_decision_steps: 3600
+  eval_deterministic: true
+  eval_raw_reward: false
+  eval_latency_values: null
+logging:
+  output_dir: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4
+  video:
+    enabled: false
+    num_bins: 1
+  save_step_records: false
+  save_action_records: false
+  save_latency_records: false
+  simulated_pipeline_profile: false

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/eval/post_train/step_5000.json

@@ -0,0 +1,318 @@
+{
+  "per_latency": {
+    "gymnasium/latency_0": {
+      "latency": 0,
+      "num_episodes": 20,
+      "mean_reward": 539.85,
+      "mean_length": 539.85,
+      "std_reward": 388.3377750103639,
+      "std_length": 388.3377750103639,
+      "episode_rewards": [
+        94.0,
+        1000.0,
+        969.0,
+        1000.0,
+        1000.0,
+        344.0,
+        1000.0,
+        225.0,
+        138.0,
+        954.0,
+        98.0,
+        572.0,
+        1000.0,
+        90.0,
+        487.0,
+        311.0,
+        317.0,
+        92.0,
+        106.0,
+        1000.0
+      ],
+      "episode_lengths": [
+        94,
+        1000,
+        969,
+        1000,
+        1000,
+        344,
+        1000,
+        225,
+        138,
+        954,
+        98,
+        572,
+        1000,
+        90,
+        487,
+        311,
+        317,
+        92,
+        106,
+        1000
+      ],
+      "decoded_action_hist": {},
+      "fixed_episode_seeds": true,
+      "eval_seed": 42,
+      "episode_seeds": [
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null
+      ],
+      "episode_indices": [
+        0,
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19
+      ]
+    },
+    "gymnasium/latency_2": {
+      "latency": 2,
+      "num_episodes": 20,
+      "mean_reward": 80.1,
+      "mean_length": 80.1,
+      "std_reward": 31.30319472513948,
+      "std_length": 31.30319472513948,
+      "episode_rewards": [
+        91.0,
+        41.0,
+        48.0,
+        86.0,
+        42.0,
+        98.0,
+        93.0,
+        152.0,
+        98.0,
+        46.0,
+        114.0,
+        83.0,
+        52.0,
+        101.0,
+        26.0,
+        112.0,
+        97.0,
+        94.0,
+        42.0,
+        86.0
+      ],
+      "episode_lengths": [
+        91,
+        41,
+        48,
+        86,
+        42,
+        98,
+        93,
+        152,
+        98,
+        46,
+        114,
+        83,
+        52,
+        101,
+        26,
+        112,
+        97,
+        94,
+        42,
+        86
+      ],
+      "decoded_action_hist": {},
+      "fixed_episode_seeds": true,
+      "eval_seed": 42,
+      "episode_seeds": [
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null
+      ],
+      "episode_indices": [
+        0,
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19
+      ]
+    },
+    "gymnasium/latency_4": {
+      "latency": 4,
+      "num_episodes": 20,
+      "mean_reward": 27.65,
+      "mean_length": 27.65,
+      "std_reward": 4.735768153108849,
+      "std_length": 4.735768153108849,
+      "episode_rewards": [
+        37.0,
+        22.0,
+        29.0,
+        31.0,
+        22.0,
+        28.0,
+        25.0,
+        30.0,
+        25.0,
+        26.0,
+        30.0,
+        26.0,
+        33.0,
+        26.0,
+        22.0,
+        40.0,
+        27.0,
+        26.0,
+        22.0,
+        26.0
+      ],
+      "episode_lengths": [
+        37,
+        22,
+        29,
+        31,
+        22,
+        28,
+        25,
+        30,
+        25,
+        26,
+        30,
+        26,
+        33,
+        26,
+        22,
+        40,
+        27,
+        26,
+        22,
+        26
+      ],
+      "decoded_action_hist": {},
+      "fixed_episode_seeds": true,
+      "eval_seed": 42,
+      "episode_seeds": [
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null
+      ],
+      "episode_indices": [
+        0,
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19
+      ]
+    }
+  },
+  "aggregate": {
+    "stage": "post_train",
+    "step": 5000,
+    "task": "gymnasium",
+    "model_alias": "openvla",
+    "fixed_episode_seeds": true,
+    "eval_seed": 42,
+    "total_episodes": 60,
+    "mean_reward": 215.86666666666667,
+    "mean_length": 215.86666666666667,
+    "std_reward": 321.7824558438338,
+    "std_length": 321.7824558438338,
+    "task_count": 1,
+    "macro_mean_reward": 215.86666666666667,
+    "macro_mean_length": 215.86666666666667,
+    "distributed_eval": false
+  }
+}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 94.0, "episode_return_env": 94.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 42, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 94, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 94}
+{"episode_id": 1, "episode_return": 1000.0, "episode_return_env": 1000.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 43, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 1000, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 1000}
+{"episode_id": 2, "episode_return": 969.0, "episode_return_env": 969.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 44, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 969, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 969}
+{"episode_id": 3, "episode_return": 1000.0, "episode_return_env": 1000.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 45, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 1000, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 1000}
+{"episode_id": 4, "episode_return": 1000.0, "episode_return_env": 1000.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 46, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 1000, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 1000}
+{"episode_id": 5, "episode_return": 344.0, "episode_return_env": 344.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 47, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 344, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 344}
+{"episode_id": 6, "episode_return": 1000.0, "episode_return_env": 1000.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 48, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 1000, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 1000}
+{"episode_id": 7, "episode_return": 225.0, "episode_return_env": 225.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 49, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 225, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 225}
+{"episode_id": 8, "episode_return": 138.0, "episode_return_env": 138.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 50, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 138, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 138}
+{"episode_id": 9, "episode_return": 954.0, "episode_return_env": 954.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 51, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 954, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 954}
+{"episode_id": 10, "episode_return": 98.0, "episode_return_env": 98.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 52, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 98, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 98}
+{"episode_id": 11, "episode_return": 572.0, "episode_return_env": 572.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 53, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 572, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 572}
+{"episode_id": 12, "episode_return": 1000.0, "episode_return_env": 1000.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 54, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 1000, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 1000}
+{"episode_id": 13, "episode_return": 90.0, "episode_return_env": 90.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 55, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 90, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 90}
+{"episode_id": 14, "episode_return": 487.0, "episode_return_env": 487.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 56, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 487, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 487}
+{"episode_id": 15, "episode_return": 311.0, "episode_return_env": 311.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 57, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 311, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 311}
+{"episode_id": 16, "episode_return": 317.0, "episode_return_env": 317.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 58, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 317, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 317}
+{"episode_id": 17, "episode_return": 92.0, "episode_return_env": 92.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 59, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 92, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 92}
+{"episode_id": 18, "episode_return": 106.0, "episode_return_env": 106.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 60, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 106, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 106}
+{"episode_id": 19, "episode_return": 1000.0, "episode_return_env": 1000.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 61, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 1000, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 1000}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/resolved_config.yaml

@@ -0,0 +1,138 @@
+experiment:
+  name: starvla_gymnasium_train_eval
+  seed: 42
+backend:
+  run_mode: eval
+executor:
+  mode: simulated
+  simulated_inference_pool: false
+  simulated_worker_capacity: null
+env:
+  name: gymnasium
+  task_name: inverted_pendulum_rgb_state
+  env_id: LatencyBench/InvertedPendulumDiscrete-v0
+  registration_imports:
+  - latency_bench.envs.gymnasium_mujoco
+  env_fps: 25.0
+  obs_fps: 25.0
+  frame_stack: 1
+  base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose exactly
+    one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+    push_right_hard.'
+  action_labels:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_values:
+  - 0
+  - 1
+  - 2
+  - 3
+  - 4
+  noop_action_id: 2
+  make_kwargs:
+    base_env_id: InvertedPendulum-v4
+    render_mode: rgb_array
+    force_values:
+    - -3.0
+    - -1.5
+    - 0.0
+    - 1.5
+    - 3.0
+    base_make_kwargs: {}
+  state_space:
+    labels:
+    - cart_position
+    - pole_angle
+    - cart_velocity
+    - pole_angular_velocity
+  action_order:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_map:
+    push_left_hard: 0
+    push_left_soft: 1
+    zero_force: 2
+    push_right_soft: 3
+    push_right_hard: 4
+  noop_action: zero_force
+  obs_resize:
+  - 224
+  - 224
+latency:
+  method: zero
+  add_latency_info: false
+scheduler:
+  hold_policy: hold
+  ordering_policy: issue_order_fifo
+policy:
+  type: starvla
+  checkpoint_path: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0
+  device: cuda
+  task_manifest_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/manifest.json
+  gymnasium_training_task_contract:
+    task_name: inverted_pendulum_rgb_state
+    env_id: LatencyBench/InvertedPendulumDiscrete-v0
+    registration_imports:
+    - latency_bench.envs.gymnasium_mujoco
+    env_fps: 25.0
+    obs_fps: 25.0
+    frame_stack: 1
+    base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose
+      exactly one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+      push_right_hard.'
+    action_labels:
+    - push_left_hard
+    - push_left_soft
+    - zero_force
+    - push_right_soft
+    - push_right_hard
+    action_values:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    noop_action_id: 2
+    make_kwargs:
+      base_env_id: InvertedPendulum-v4
+      render_mode: rgb_array
+      force_values:
+      - -3.0
+      - -1.5
+      - 0.0
+      - 1.5
+      - 3.0
+      base_make_kwargs: {}
+    state_space:
+      labels:
+      - cart_position
+      - pole_angle
+      - cart_velocity
+      - pole_angular_velocity
+  latency_prompt_map_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_prompt_map.json
+  latency_prompt_key: 0
+  action_prefix:
+    mode: none
+evaluation:
+  eval_episodes: 20
+  eval_parallel_envs: 5
+  eval_max_steps: 3600
+  eval_max_decision_steps: 3600
+  eval_deterministic: true
+  eval_raw_reward: false
+  eval_latency_values: null
+logging:
+  output_dir: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0
+  video:
+    enabled: false
+    num_bins: 1
+  save_step_records: false
+  save_action_records: false
+  save_latency_records: false
+  simulated_pipeline_profile: false

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 91.0, "episode_return_env": 91.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 42, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 91, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 91}
+{"episode_id": 1, "episode_return": 41.0, "episode_return_env": 41.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 43, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 41, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 41}
+{"episode_id": 2, "episode_return": 48.0, "episode_return_env": 48.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 44, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 48, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 48}
+{"episode_id": 3, "episode_return": 86.0, "episode_return_env": 86.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 45, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 86, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 86}
+{"episode_id": 4, "episode_return": 42.0, "episode_return_env": 42.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 46, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 42, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 42}
+{"episode_id": 5, "episode_return": 98.0, "episode_return_env": 98.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 47, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 98, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 98}
+{"episode_id": 6, "episode_return": 93.0, "episode_return_env": 93.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 48, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 93, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 93}
+{"episode_id": 7, "episode_return": 152.0, "episode_return_env": 152.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 49, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 152, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 152}
+{"episode_id": 8, "episode_return": 98.0, "episode_return_env": 98.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 50, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 98, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 98}
+{"episode_id": 9, "episode_return": 46.0, "episode_return_env": 46.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 51, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 46, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 46}
+{"episode_id": 10, "episode_return": 114.0, "episode_return_env": 114.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 52, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 114, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 114}
+{"episode_id": 11, "episode_return": 83.0, "episode_return_env": 83.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 53, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 83, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 83}
+{"episode_id": 12, "episode_return": 52.0, "episode_return_env": 52.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 54, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 52, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 52}
+{"episode_id": 13, "episode_return": 101.0, "episode_return_env": 101.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 55, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 101, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 101}
+{"episode_id": 14, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 56, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 26}
+{"episode_id": 15, "episode_return": 112.0, "episode_return_env": 112.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 57, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 112, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 112}
+{"episode_id": 16, "episode_return": 97.0, "episode_return_env": 97.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 58, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 97, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 97}
+{"episode_id": 17, "episode_return": 94.0, "episode_return_env": 94.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 59, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 94, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 94}
+{"episode_id": 18, "episode_return": 42.0, "episode_return_env": 42.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 60, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 42, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 42}
+{"episode_id": 19, "episode_return": 86.0, "episode_return_env": 86.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 61, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 86, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 86}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/resolved_config.yaml

@@ -0,0 +1,139 @@
+experiment:
+  name: starvla_gymnasium_train_eval
+  seed: 42
+backend:
+  run_mode: eval
+executor:
+  mode: simulated
+  simulated_inference_pool: false
+  simulated_worker_capacity: null
+env:
+  name: gymnasium
+  task_name: inverted_pendulum_rgb_state
+  env_id: LatencyBench/InvertedPendulumDiscrete-v0
+  registration_imports:
+  - latency_bench.envs.gymnasium_mujoco
+  env_fps: 25.0
+  obs_fps: 25.0
+  frame_stack: 1
+  base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose exactly
+    one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+    push_right_hard.'
+  action_labels:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_values:
+  - 0
+  - 1
+  - 2
+  - 3
+  - 4
+  noop_action_id: 2
+  make_kwargs:
+    base_env_id: InvertedPendulum-v4
+    render_mode: rgb_array
+    force_values:
+    - -3.0
+    - -1.5
+    - 0.0
+    - 1.5
+    - 3.0
+    base_make_kwargs: {}
+  state_space:
+    labels:
+    - cart_position
+    - pole_angle
+    - cart_velocity
+    - pole_angular_velocity
+  action_order:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_map:
+    push_left_hard: 0
+    push_left_soft: 1
+    zero_force: 2
+    push_right_soft: 3
+    push_right_hard: 4
+  noop_action: zero_force
+  obs_resize:
+  - 224
+  - 224
+latency:
+  method: fixed
+  add_latency_info: false
+  fixed_latency_ms: 80.0
+scheduler:
+  hold_policy: hold
+  ordering_policy: issue_order_fifo
+policy:
+  type: starvla
+  checkpoint_path: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2
+  device: cuda
+  task_manifest_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/manifest.json
+  gymnasium_training_task_contract:
+    task_name: inverted_pendulum_rgb_state
+    env_id: LatencyBench/InvertedPendulumDiscrete-v0
+    registration_imports:
+    - latency_bench.envs.gymnasium_mujoco
+    env_fps: 25.0
+    obs_fps: 25.0
+    frame_stack: 1
+    base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose
+      exactly one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+      push_right_hard.'
+    action_labels:
+    - push_left_hard
+    - push_left_soft
+    - zero_force
+    - push_right_soft
+    - push_right_hard
+    action_values:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    noop_action_id: 2
+    make_kwargs:
+      base_env_id: InvertedPendulum-v4
+      render_mode: rgb_array
+      force_values:
+      - -3.0
+      - -1.5
+      - 0.0
+      - 1.5
+      - 3.0
+      base_make_kwargs: {}
+    state_space:
+      labels:
+      - cart_position
+      - pole_angle
+      - cart_velocity
+      - pole_angular_velocity
+  latency_prompt_map_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_prompt_map.json
+  latency_prompt_key: 2
+  action_prefix:
+    mode: none
+evaluation:
+  eval_episodes: 20
+  eval_parallel_envs: 5
+  eval_max_steps: 3600
+  eval_max_decision_steps: 3600
+  eval_deterministic: true
+  eval_raw_reward: false
+  eval_latency_values: null
+logging:
+  output_dir: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2
+  video:
+    enabled: false
+    num_bins: 1
+  save_step_records: false
+  save_action_records: false
+  save_latency_records: false
+  simulated_pipeline_profile: false

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 37.0, "episode_return_env": 37.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 42, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 37, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 37}
+{"episode_id": 1, "episode_return": 22.0, "episode_return_env": 22.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 43, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 22}
+{"episode_id": 2, "episode_return": 29.0, "episode_return_env": 29.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 44, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 29, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 29}
+{"episode_id": 3, "episode_return": 31.0, "episode_return_env": 31.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 45, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 31, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 31}
+{"episode_id": 4, "episode_return": 22.0, "episode_return_env": 22.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 46, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 22}
+{"episode_id": 5, "episode_return": 28.0, "episode_return_env": 28.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 47, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 28, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 28}
+{"episode_id": 6, "episode_return": 25.0, "episode_return_env": 25.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 48, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 25, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 25}
+{"episode_id": 7, "episode_return": 30.0, "episode_return_env": 30.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 49, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 30, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 30}
+{"episode_id": 8, "episode_return": 25.0, "episode_return_env": 25.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 50, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 25, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 25}
+{"episode_id": 9, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 51, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 26}
+{"episode_id": 10, "episode_return": 30.0, "episode_return_env": 30.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 52, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 30, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 30}
+{"episode_id": 11, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 53, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 26}
+{"episode_id": 12, "episode_return": 33.0, "episode_return_env": 33.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 54, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 33, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 33}
+{"episode_id": 13, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 55, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 26}
+{"episode_id": 14, "episode_return": 22.0, "episode_return_env": 22.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 56, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 22}
+{"episode_id": 15, "episode_return": 40.0, "episode_return_env": 40.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 57, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 40, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 40}
+{"episode_id": 16, "episode_return": 27.0, "episode_return_env": 27.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 58, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 27, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 27}
+{"episode_id": 17, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 59, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 26}
+{"episode_id": 18, "episode_return": 22.0, "episode_return_env": 22.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 60, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 22, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 22}
+{"episode_id": 19, "episode_return": 26.0, "episode_return_env": 26.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 61, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 26, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 26}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/resolved_config.yaml

@@ -0,0 +1,139 @@
+experiment:
+  name: starvla_gymnasium_train_eval
+  seed: 42
+backend:
+  run_mode: eval
+executor:
+  mode: simulated
+  simulated_inference_pool: false
+  simulated_worker_capacity: null
+env:
+  name: gymnasium
+  task_name: inverted_pendulum_rgb_state
+  env_id: LatencyBench/InvertedPendulumDiscrete-v0
+  registration_imports:
+  - latency_bench.envs.gymnasium_mujoco
+  env_fps: 25.0
+  obs_fps: 25.0
+  frame_stack: 1
+  base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose exactly
+    one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+    push_right_hard.'
+  action_labels:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_values:
+  - 0
+  - 1
+  - 2
+  - 3
+  - 4
+  noop_action_id: 2
+  make_kwargs:
+    base_env_id: InvertedPendulum-v4
+    render_mode: rgb_array
+    force_values:
+    - -3.0
+    - -1.5
+    - 0.0
+    - 1.5
+    - 3.0
+    base_make_kwargs: {}
+  state_space:
+    labels:
+    - cart_position
+    - pole_angle
+    - cart_velocity
+    - pole_angular_velocity
+  action_order:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_map:
+    push_left_hard: 0
+    push_left_soft: 1
+    zero_force: 2
+    push_right_soft: 3
+    push_right_hard: 4
+  noop_action: zero_force
+  obs_resize:
+  - 224
+  - 224
+latency:
+  method: fixed
+  add_latency_info: false
+  fixed_latency_ms: 160.0
+scheduler:
+  hold_policy: hold
+  ordering_policy: issue_order_fifo
+policy:
+  type: starvla
+  checkpoint_path: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4
+  device: cuda
+  task_manifest_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/manifest.json
+  gymnasium_training_task_contract:
+    task_name: inverted_pendulum_rgb_state
+    env_id: LatencyBench/InvertedPendulumDiscrete-v0
+    registration_imports:
+    - latency_bench.envs.gymnasium_mujoco
+    env_fps: 25.0
+    obs_fps: 25.0
+    frame_stack: 1
+    base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose
+      exactly one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+      push_right_hard.'
+    action_labels:
+    - push_left_hard
+    - push_left_soft
+    - zero_force
+    - push_right_soft
+    - push_right_hard
+    action_values:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    noop_action_id: 2
+    make_kwargs:
+      base_env_id: InvertedPendulum-v4
+      render_mode: rgb_array
+      force_values:
+      - -3.0
+      - -1.5
+      - 0.0
+      - 1.5
+      - 3.0
+      base_make_kwargs: {}
+    state_space:
+      labels:
+      - cart_position
+      - pole_angle
+      - cart_velocity
+      - pole_angular_velocity
+  latency_prompt_map_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_prompt_map.json
+  latency_prompt_key: 4
+  action_prefix:
+    mode: none
+evaluation:
+  eval_episodes: 20
+  eval_parallel_envs: 5
+  eval_max_steps: 3600
+  eval_max_decision_steps: 3600
+  eval_deterministic: true
+  eval_raw_reward: false
+  eval_latency_values: null
+logging:
+  output_dir: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4
+  video:
+    enabled: false
+    num_bins: 1
+  save_step_records: false
+  save_action_records: false
+  save_latency_records: false
+  simulated_pipeline_profile: false

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/eval/post_train/step_5000.json

@@ -0,0 +1,318 @@
+{
+  "per_latency": {
+    "gymnasium/latency_0": {
+      "latency": 0,
+      "num_episodes": 20,
+      "mean_reward": 56.2,
+      "mean_length": 56.2,
+      "std_reward": 8.86904729945669,
+      "std_length": 8.86904729945669,
+      "episode_rewards": [
+        63.0,
+        46.0,
+        63.0,
+        58.0,
+        54.0,
+        54.0,
+        46.0,
+        58.0,
+        46.0,
+        60.0,
+        53.0,
+        48.0,
+        79.0,
+        46.0,
+        72.0,
+        58.0,
+        48.0,
+        65.0,
+        50.0,
+        57.0
+      ],
+      "episode_lengths": [
+        63,
+        46,
+        63,
+        58,
+        54,
+        54,
+        46,
+        58,
+        46,
+        60,
+        53,
+        48,
+        79,
+        46,
+        72,
+        58,
+        48,
+        65,
+        50,
+        57
+      ],
+      "decoded_action_hist": {},
+      "fixed_episode_seeds": true,
+      "eval_seed": 42,
+      "episode_seeds": [
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null
+      ],
+      "episode_indices": [
+        0,
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19
+      ]
+    },
+    "gymnasium/latency_2": {
+      "latency": 2,
+      "num_episodes": 20,
+      "mean_reward": 71.55,
+      "mean_length": 71.55,
+      "std_reward": 28.073964807272947,
+      "std_length": 28.073964807272947,
+      "episode_rewards": [
+        114.0,
+        45.0,
+        64.0,
+        66.0,
+        42.0,
+        115.0,
+        82.0,
+        94.0,
+        98.0,
+        46.0,
+        114.0,
+        34.0,
+        52.0,
+        100.0,
+        45.0,
+        112.0,
+        54.0,
+        60.0,
+        42.0,
+        52.0
+      ],
+      "episode_lengths": [
+        114,
+        45,
+        64,
+        66,
+        42,
+        115,
+        82,
+        94,
+        98,
+        46,
+        114,
+        34,
+        52,
+        100,
+        45,
+        112,
+        54,
+        60,
+        42,
+        52
+      ],
+      "decoded_action_hist": {},
+      "fixed_episode_seeds": true,
+      "eval_seed": 42,
+      "episode_seeds": [
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null
+      ],
+      "episode_indices": [
+        0,
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19
+      ]
+    },
+    "gymnasium/latency_4": {
+      "latency": 4,
+      "num_episodes": 20,
+      "mean_reward": 27.45,
+      "mean_length": 27.45,
+      "std_reward": 4.271709259769443,
+      "std_length": 4.271709259769443,
+      "episode_rewards": [
+        37.0,
+        22.0,
+        29.0,
+        31.0,
+        22.0,
+        28.0,
+        25.0,
+        30.0,
+        25.0,
+        26.0,
+        30.0,
+        26.0,
+        33.0,
+        26.0,
+        22.0,
+        36.0,
+        27.0,
+        26.0,
+        22.0,
+        26.0
+      ],
+      "episode_lengths": [
+        37,
+        22,
+        29,
+        31,
+        22,
+        28,
+        25,
+        30,
+        25,
+        26,
+        30,
+        26,
+        33,
+        26,
+        22,
+        36,
+        27,
+        26,
+        22,
+        26
+      ],
+      "decoded_action_hist": {},
+      "fixed_episode_seeds": true,
+      "eval_seed": 42,
+      "episode_seeds": [
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null,
+        null
+      ],
+      "episode_indices": [
+        0,
+        1,
+        2,
+        3,
+        4,
+        5,
+        6,
+        7,
+        8,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        17,
+        18,
+        19
+      ]
+    }
+  },
+  "aggregate": {
+    "stage": "post_train",
+    "step": 5000,
+    "task": "gymnasium",
+    "model_alias": "openvla",
+    "fixed_episode_seeds": true,
+    "eval_seed": 42,
+    "total_episodes": 60,
+    "mean_reward": 51.733333333333334,
+    "mean_length": 51.733333333333334,
+    "std_reward": 25.08244184462288,
+    "std_length": 25.08244184462288,
+    "task_count": 1,
+    "macro_mean_reward": 51.73333333333333,
+    "macro_mean_length": 51.73333333333333,
+    "distributed_eval": false
+  }
+}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 63.0, "episode_return_env": 63.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 42, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 63, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 63}
+{"episode_id": 1, "episode_return": 46.0, "episode_return_env": 46.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 43, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 46, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 46}
+{"episode_id": 2, "episode_return": 63.0, "episode_return_env": 63.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 44, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 63, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 63}
+{"episode_id": 3, "episode_return": 58.0, "episode_return_env": 58.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 45, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 58, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 58}
+{"episode_id": 4, "episode_return": 54.0, "episode_return_env": 54.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 46, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 54, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 54}
+{"episode_id": 5, "episode_return": 54.0, "episode_return_env": 54.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 47, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 54, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 54}
+{"episode_id": 6, "episode_return": 46.0, "episode_return_env": 46.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 48, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 46, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 46}
+{"episode_id": 7, "episode_return": 58.0, "episode_return_env": 58.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 49, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 58, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 58}
+{"episode_id": 8, "episode_return": 46.0, "episode_return_env": 46.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 50, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 46, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 46}
+{"episode_id": 9, "episode_return": 60.0, "episode_return_env": 60.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 51, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 60, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 60}
+{"episode_id": 10, "episode_return": 53.0, "episode_return_env": 53.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 52, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 53, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 53}
+{"episode_id": 11, "episode_return": 48.0, "episode_return_env": 48.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 53, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 48, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 48}
+{"episode_id": 12, "episode_return": 79.0, "episode_return_env": 79.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 54, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 79, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 79}
+{"episode_id": 13, "episode_return": 46.0, "episode_return_env": 46.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 55, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 46, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 46}
+{"episode_id": 14, "episode_return": 72.0, "episode_return_env": 72.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 56, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 72, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 72}
+{"episode_id": 15, "episode_return": 58.0, "episode_return_env": 58.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 57, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 58, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 58}
+{"episode_id": 16, "episode_return": 48.0, "episode_return_env": 48.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 58, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 48, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 48}
+{"episode_id": 17, "episode_return": 65.0, "episode_return_env": 65.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 59, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 65, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 65}
+{"episode_id": 18, "episode_return": 50.0, "episode_return_env": 50.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 60, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 50, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 50}
+{"episode_id": 19, "episode_return": 57.0, "episode_return_env": 57.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 61, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 57, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 57}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/resolved_config.yaml

@@ -0,0 +1,139 @@
+experiment:
+  name: starvla_gymnasium_train_eval
+  seed: 42
+backend:
+  run_mode: eval
+executor:
+  mode: simulated
+  simulated_inference_pool: false
+  simulated_worker_capacity: null
+env:
+  name: gymnasium
+  task_name: inverted_pendulum_rgb_state
+  env_id: LatencyBench/InvertedPendulumDiscrete-v0
+  registration_imports:
+  - latency_bench.envs.gymnasium_mujoco
+  env_fps: 25.0
+  obs_fps: 25.0
+  frame_stack: 1
+  base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose exactly
+    one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+    push_right_hard.'
+  action_labels:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_values:
+  - 0
+  - 1
+  - 2
+  - 3
+  - 4
+  noop_action_id: 2
+  make_kwargs:
+    base_env_id: InvertedPendulum-v4
+    render_mode: rgb_array
+    force_values:
+    - -3.0
+    - -1.5
+    - 0.0
+    - 1.5
+    - 3.0
+    base_make_kwargs: {}
+  state_space:
+    labels:
+    - cart_position
+    - pole_angle
+    - cart_velocity
+    - pole_angular_velocity
+  action_order:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_map:
+    push_left_hard: 0
+    push_left_soft: 1
+    zero_force: 2
+    push_right_soft: 3
+    push_right_hard: 4
+  noop_action: zero_force
+  obs_resize:
+  - 224
+  - 224
+latency:
+  method: zero
+  add_latency_info: false
+scheduler:
+  hold_policy: hold
+  ordering_policy: issue_order_fifo
+policy:
+  type: starvla
+  checkpoint_path: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0
+  device: cuda
+  task_manifest_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/manifest.json
+  gymnasium_training_task_contract:
+    task_name: inverted_pendulum_rgb_state
+    env_id: LatencyBench/InvertedPendulumDiscrete-v0
+    registration_imports:
+    - latency_bench.envs.gymnasium_mujoco
+    env_fps: 25.0
+    obs_fps: 25.0
+    frame_stack: 1
+    base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose
+      exactly one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+      push_right_hard.'
+    action_labels:
+    - push_left_hard
+    - push_left_soft
+    - zero_force
+    - push_right_soft
+    - push_right_hard
+    action_values:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    noop_action_id: 2
+    make_kwargs:
+      base_env_id: InvertedPendulum-v4
+      render_mode: rgb_array
+      force_values:
+      - -3.0
+      - -1.5
+      - 0.0
+      - 1.5
+      - 3.0
+      base_make_kwargs: {}
+    state_space:
+      labels:
+      - cart_position
+      - pole_angle
+      - cart_velocity
+      - pole_angular_velocity
+  latency_prompt_map_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_prompt_map.json
+  latency_prompt_key: 0
+  prompt_mode: latency_neutral
+  action_prefix:
+    mode: none
+evaluation:
+  eval_episodes: 20
+  eval_parallel_envs: 5
+  eval_max_steps: 3600
+  eval_max_decision_steps: 3600
+  eval_deterministic: true
+  eval_raw_reward: false
+  eval_latency_values: null
+logging:
+  output_dir: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0
+  video:
+    enabled: false
+    num_bins: 1
+  save_step_records: false
+  save_action_records: false
+  save_latency_records: false
+  simulated_pipeline_profile: false

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/episode_metrics.jsonl

@@ -0,0 +1,20 @@
+{"episode_id": 0, "episode_return": 114.0, "episode_return_env": 114.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 42, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 114, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 114}
+{"episode_id": 1, "episode_return": 45.0, "episode_return_env": 45.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 43, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 45, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 45}
+{"episode_id": 2, "episode_return": 64.0, "episode_return_env": 64.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 44, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 64, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 64}
+{"episode_id": 3, "episode_return": 66.0, "episode_return_env": 66.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 45, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 66, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 66}
+{"episode_id": 4, "episode_return": 42.0, "episode_return_env": 42.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 46, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 42, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 42}
+{"episode_id": 5, "episode_return": 115.0, "episode_return_env": 115.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 47, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 115, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 115}
+{"episode_id": 6, "episode_return": 82.0, "episode_return_env": 82.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 48, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 82, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 82}
+{"episode_id": 7, "episode_return": 94.0, "episode_return_env": 94.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 49, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 94, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 94}
+{"episode_id": 8, "episode_return": 98.0, "episode_return_env": 98.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 50, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 98, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 98}
+{"episode_id": 9, "episode_return": 46.0, "episode_return_env": 46.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 51, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 46, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 46}
+{"episode_id": 10, "episode_return": 114.0, "episode_return_env": 114.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 52, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 114, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 114}
+{"episode_id": 11, "episode_return": 34.0, "episode_return_env": 34.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 53, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 34, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 34}
+{"episode_id": 12, "episode_return": 52.0, "episode_return_env": 52.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 54, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 52, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 52}
+{"episode_id": 13, "episode_return": 100.0, "episode_return_env": 100.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 55, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 100, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 100}
+{"episode_id": 14, "episode_return": 45.0, "episode_return_env": 45.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 56, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 45, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 45}
+{"episode_id": 15, "episode_return": 112.0, "episode_return_env": 112.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 57, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 112, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 112}
+{"episode_id": 16, "episode_return": 54.0, "episode_return_env": 54.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 58, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 54, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 54}
+{"episode_id": 17, "episode_return": 60.0, "episode_return_env": 60.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 59, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 60, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 60}
+{"episode_id": 18, "episode_return": 42.0, "episode_return_env": 42.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 60, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 42, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 42}
+{"episode_id": 19, "episode_return": 52.0, "episode_return_env": 52.0, "game_score": null, "mean_latency_ms": 80.0, "metadata": {"action_prefix": {"mode": "none"}, "checkpoint_path": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 25.0, "env_id": "LatencyBench/InvertedPendulumDiscrete-v0", "episode_seed": 61, "frame_ms": 40.0, "gpu_class": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 25.0, "output_dir": "/home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 52, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 80.0, "p99_latency_ms": 80.0, "return_raw": null, "survival_steps": 52}

evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/resolved_config.yaml

@@ -0,0 +1,140 @@
+experiment:
+  name: starvla_gymnasium_train_eval
+  seed: 42
+backend:
+  run_mode: eval
+executor:
+  mode: simulated
+  simulated_inference_pool: false
+  simulated_worker_capacity: null
+env:
+  name: gymnasium
+  task_name: inverted_pendulum_rgb_state
+  env_id: LatencyBench/InvertedPendulumDiscrete-v0
+  registration_imports:
+  - latency_bench.envs.gymnasium_mujoco
+  env_fps: 25.0
+  obs_fps: 25.0
+  frame_stack: 1
+  base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose exactly
+    one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+    push_right_hard.'
+  action_labels:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_values:
+  - 0
+  - 1
+  - 2
+  - 3
+  - 4
+  noop_action_id: 2
+  make_kwargs:
+    base_env_id: InvertedPendulum-v4
+    render_mode: rgb_array
+    force_values:
+    - -3.0
+    - -1.5
+    - 0.0
+    - 1.5
+    - 3.0
+    base_make_kwargs: {}
+  state_space:
+    labels:
+    - cart_position
+    - pole_angle
+    - cart_velocity
+    - pole_angular_velocity
+  action_order:
+  - push_left_hard
+  - push_left_soft
+  - zero_force
+  - push_right_soft
+  - push_right_hard
+  action_map:
+    push_left_hard: 0
+    push_left_soft: 1
+    zero_force: 2
+    push_right_soft: 3
+    push_right_hard: 4
+  noop_action: zero_force
+  obs_resize:
+  - 224
+  - 224
+latency:
+  method: fixed
+  add_latency_info: false
+  fixed_latency_ms: 80.0
+scheduler:
+  hold_policy: hold
+  ordering_policy: issue_order_fifo
+policy:
+  type: starvla
+  checkpoint_path: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2
+  device: cuda
+  task_manifest_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/manifest.json
+  gymnasium_training_task_contract:
+    task_name: inverted_pendulum_rgb_state
+    env_id: LatencyBench/InvertedPendulumDiscrete-v0
+    registration_imports:
+    - latency_bench.envs.gymnasium_mujoco
+    env_fps: 25.0
+    obs_fps: 25.0
+    frame_stack: 1
+    base_prompt: 'Keep the pole upright by applying horizontal cart forces. Choose
+      exactly one action from: push_left_hard, push_left_soft, zero_force, push_right_soft,
+      push_right_hard.'
+    action_labels:
+    - push_left_hard
+    - push_left_soft
+    - zero_force
+    - push_right_soft
+    - push_right_hard
+    action_values:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    noop_action_id: 2
+    make_kwargs:
+      base_env_id: InvertedPendulum-v4
+      render_mode: rgb_array
+      force_values:
+      - -3.0
+      - -1.5
+      - 0.0
+      - 1.5
+      - 3.0
+      base_make_kwargs: {}
+    state_space:
+      labels:
+      - cart_position
+      - pole_angle
+      - cart_velocity
+      - pole_angular_velocity
+  latency_prompt_map_path: /home/ubuntu/talha/playground/Datasets/rl_games/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_prompt_map.json
+  latency_prompt_key: 2
+  prompt_mode: latency_neutral
+  action_prefix:
+    mode: none
+evaluation:
+  eval_episodes: 20
+  eval_parallel_envs: 5
+  eval_max_steps: 3600
+  eval_max_decision_steps: 3600
+  eval_deterministic: true
+  eval_raw_reward: false
+  eval_latency_values: null
+logging:
+  output_dir: /home/ubuntu/talha/results/Checkpoints/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2
+  video:
+    enabled: false
+    num_bins: 1
+  save_step_records: false
+  save_action_records: false
+  save_latency_records: false
+  simulated_pipeline_profile: false