diff --git a/src/lerobot/policies/rlearn/modeling_rlearn.py b/src/lerobot/policies/rlearn/modeling_rlearn.py
index e8f1db5b1..c6399226a 100644
--- a/src/lerobot/policies/rlearn/modeling_rlearn.py
+++ b/src/lerobot/policies/rlearn/modeling_rlearn.py
@@ -377,9 +377,7 @@ class RLearNPolicy(PreTrainedPolicy):
         inputs = {k: v.to(device) for k, v in inputs.items()}
         
         # Process in batch through DINOv3 model
-        # Use inference mode for stable, fast frozen encoder forward
-        with torch.inference_mode():
-            vision_outputs = self.vision_model(**inputs)
+        vision_outputs = self.vision_model(**inputs)
         
         # Prefer mean-pooled patch tokens over pooler/CLS to ensure input-dependent variation
         if hasattr(vision_outputs, 'last_hidden_state') and vision_outputs.last_hidden_state is not None: