diff --git a/src/lerobot/policies/smolvla/modeling_smolvla.py b/src/lerobot/policies/smolvla/modeling_smolvla.py
index f4f158951..ff656febe 100644
--- a/src/lerobot/policies/smolvla/modeling_smolvla.py
+++ b/src/lerobot/policies/smolvla/modeling_smolvla.py
@@ -383,8 +383,6 @@ class SmolVLAPolicy(PreTrainedPolicy):
         original_action_dim = self.config.action_feature.shape[0]
         actions = actions[:, :, :original_action_dim]
 
-        actions = self.unnormalize_outputs({ACTION: actions})[ACTION]
-
         if self.config.adapt_to_pi_aloha:
             actions = self._pi_aloha_encode_actions(actions)
 
diff --git a/src/lerobot/policies/tdmpc/modeling_tdmpc.py b/src/lerobot/policies/tdmpc/modeling_tdmpc.py
index 0ee2e3a43..ca2c23c94 100644
--- a/src/lerobot/policies/tdmpc/modeling_tdmpc.py
+++ b/src/lerobot/policies/tdmpc/modeling_tdmpc.py
@@ -138,8 +138,6 @@ class TDMPCPolicy(PreTrainedPolicy):
         if ACTION in batch:
             batch.pop(ACTION)
 
-        batch = self.normalize_inputs(batch)
-
         if self.config.image_features:
             batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
             batch[OBS_IMAGE] = batch[next(iter(self.config.image_features))]
diff --git a/src/lerobot/policies/vqbet/modeling_vqbet.py b/src/lerobot/policies/vqbet/modeling_vqbet.py
index 41e021d8f..c3cc02eaa 100644
--- a/src/lerobot/policies/vqbet/modeling_vqbet.py
+++ b/src/lerobot/policies/vqbet/modeling_vqbet.py
@@ -131,7 +131,6 @@ class VQBeTPolicy(PreTrainedPolicy):
         # NOTE: for offline evaluation, we have action in the batch, so we need to pop it out
         if ACTION in batch:
             batch.pop(ACTION)
-        batch = self.normalize_inputs(batch)
         batch = dict(batch)  # shallow copy so that adding a key doesn't modify the original
         # NOTE: It's important that this happens after stacking the images into a single key.
         batch["observation.images"] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)