feat(normalization): add validation for empty features in NormalizerProcessorStep and UnnormalizerProcessorStep (#2087)

* feat(normalization): add validation for empty features in NormalizerProcessorStep and UnnormalizerProcessorStep

* refactor(normalization): streamline feature reconstruction logic in _NormalizationMixin

* refactor(tests): remove unused preprocessor initialization in test_act_backbone_lr

---------

Co-authored-by: Pepijn <138571049+pkooij@users.noreply.github.com>

src/lerobot/processor/normalize_processor.py

@@ -108,8 +108,10 @@ class _NormalizationMixin:
         """
         # Track if stats were explicitly provided (not None and not empty)
         self._stats_explicitly_provided = self.stats is not None and bool(self.stats)
+        # Check if self.features is not empty
+        if not self.features:
+            raise ValueError("Normalization features cannot be empty")
         # Robust JSON deserialization handling (guard empty maps).
-        if self.features:
         first_val = next(iter(self.features.values()))
         if isinstance(first_val, dict):
             reconstructed = {}

tests/policies/test_policies.py

@@ -234,7 +234,6 @@ def test_act_backbone_lr():
     assert cfg.policy.optimizer_lr_backbone == 0.001
 
     dataset = make_dataset(cfg)
-    preprocessor, _ = make_pre_post_processors(cfg.policy, None)
     policy = make_policy(cfg.policy, ds_meta=dataset.meta)
     optimizer, _ = make_optimizer_and_scheduler(cfg, policy)
     assert len(optimizer.param_groups) == 2

tests/processor/test_normalize_processor.py

@@ -534,6 +534,18 @@ def test_empty_observation():
     assert result == transition
 
 
+def test_empty_features_raises_error():
+    """Test that empty features dict raises ValueError."""
+    norm_map = {FeatureType.VISUAL: NormalizationMode.MEAN_STD}
+    stats = {OBS_IMAGE: {"mean": [0.5], "std": [0.2]}}
+
+    with pytest.raises(ValueError, match="Normalization features cannot be empty"):
+        NormalizerProcessorStep(features={}, norm_map=norm_map, stats=stats)
+
+    with pytest.raises(ValueError, match="Normalization features cannot be empty"):
+        UnnormalizerProcessorStep(features={}, norm_map=norm_map, stats=stats)
+
+
 def test_empty_stats():
     features = {OBS_IMAGE: PolicyFeature(FeatureType.VISUAL, (3, 96, 96))}
     norm_map = {FeatureType.VISUAL: NormalizationMode.MEAN_STD}