fix(datasets): skip zero-width features in compute_episode_stats

`LeRobotDataset.save_episode()` raised
`ValueError: cannot reshape array of size 0 into shape (0)` whenever a
declared non-string feature had a zero-width dimension (e.g. `shape=(0,)`).
The root cause was `compute_episode_stats` running stats on every
non-string/language feature, then `RunningQuantileStats.update` calling
`batch.reshape(-1, batch.shape[-1])` on the empty array.

Skip features whose declared `shape` contains a zero dim, mirroring the
existing skip for `string` / `language` dtype features.

Fixes #3654

src/lerobot/datasets/compute_stats.py

@@ -519,6 +519,9 @@ def compute_episode_stats(
         if features[key]["dtype"] in {"string", "language"}:
             continue
 
+        if any(d == 0 for d in features[key].get("shape", ())):
+            continue
+
         if features[key]["dtype"] in ["image", "video"]:
             ep_ft_array = sample_images(data)
             axes_to_reduce = (0, 2, 3)

tests/datasets/test_compute_stats.py

@@ -687,6 +687,29 @@ def test_compute_episode_stats_string_features_skipped():
     assert "q01" in stats["action"]
 
 
+def test_compute_episode_stats_zero_width_features_skipped():
+    """Test that features with a zero-width dim (e.g. shape=(0,)) are skipped."""
+    episode_data = {
+        "empty": np.zeros((100, 0), dtype=np.float32),  # Zero-width feature
+        "action": np.random.normal(0, 1, (100, 5)),
+    }
+    features = {
+        "empty": {"dtype": "float32", "shape": (0,)},
+        "action": {"dtype": "float32", "shape": (5,)},
+    }
+
+    stats = compute_episode_stats(
+        episode_data,
+        features,
+    )
+
+    # Zero-width features should be skipped
+    assert "empty" not in stats
+    assert "action" in stats
+    assert "q01" in stats["action"]
+    assert stats["action"]["mean"].shape == (5,)
+
+
 def test_aggregate_feature_stats_with_quantiles():
     """Test aggregating feature stats that include quantiles."""
     stats_ft_list = [