feat(self-obstruction): remove gripper self-obstruction as a blurry trigger

src/lerobot/annotations/camera_curation/curator.py

@@ -104,6 +104,10 @@ def _parse_verdict(camera_key: str, result: Any, cfg: CameraCurationConfig) -> C
     raw_label = result.get("view_label")
     label = str(raw_label).strip().lower().replace(" ", "_") if raw_label else ""
     view_label = label if is_valid_view_label(label, cfg.view_vocabulary, cfg.allow_combos) else None
+    if view_label is not None:
+        # Canonicalize combo order (``wrist_left`` -> ``left_wrist``) so the set
+        # of possible keys is deterministic regardless of the VLM's word order.
+        view_label = _order_combo(view_label.split("_"), cfg.view_vocabulary)
     if raw_label and view_label is None:
         logger.warning(
             "camera %s: VLM returned view_label=%r which is not in the vocabulary %s; leaving unlabeled",

src/lerobot/annotations/camera_curation/prompts/camera_curation.txt

@@ -5,9 +5,15 @@ together to judge the camera, not any single moment.
 Do two things and return them as one JSON object.
 
 1. QUALITY. Decide whether this camera view is usable for training a policy.
-   Mark it UNUSABLE if it is blurry / out of focus, badly over- or
+   Mark it UNUSABLE only for a genuine capture defect: blurry / out of focus,
-   under-exposed, mostly occluded, static/frozen, corrupted, or otherwise does
+   badly over- or under-exposed, static/frozen, corrupted, or otherwise not
-   not clearly show the scene. Otherwise it is usable.
+   showing the scene clearly. Otherwise it is usable.
+
+   Do NOT mark a view unusable because the robot's own gripper, arm, or
+   end-effector is visible or partially blocks the view — that is EXPECTED,
+   especially for a wrist-mounted camera where the gripper is normally in frame.
+   Only true image-quality problems count; the robot occluding its own
+   workspace does not.
 
 2. VIEW LABEL. Choose the single best label for where this camera is mounted /
    what it looks at, using ONLY this closed vocabulary:

tests/annotations/test_camera_curation.py

@@ -124,6 +124,15 @@ def test_curate_cameras_parses_and_validates(tmp_path):
     assert verdicts["observation.images.c"].view_label is None  # no frames
 
 
+def test_curate_cameras_canonicalizes_combo_order():
+    cfg = CameraCurationConfig(view_vocabulary=VOCAB)
+    frames = {"observation.images.a": [_tiny_image()]}
+    # VLM emits the combo in the "wrong" order; we normalize to <side>_wrist.
+    vlm = _queued_vlm([{"usable": True, "blur_reason": None, "view_label": "wrist_left"}])
+    (verdict,) = curate_cameras(frames, cfg, vlm)
+    assert verdict.view_label == "left_wrist"
+
+
 def test_build_name_mapping_and_collision():
     cfg = CameraCurationConfig(view_vocabulary=VOCAB)
     existing = {"observation.images.cam_0": {}, "observation.images.cam_1": {}, "observation.images.top": {}}