diff --git a/src/lerobot/policies/pretrained.py b/src/lerobot/policies/pretrained.py
index 724f920f3..61a856f0b 100644
--- a/src/lerobot/policies/pretrained.py
+++ b/src/lerobot/policies/pretrained.py
@@ -40,6 +40,7 @@ T = TypeVar("T", bound="PreTrainedPolicy")
 
 class ActionSelectKwargs(TypedDict, total=False):
     noise: Tensor | None
+    return_extra: bool
 
 
 class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
@@ -187,20 +188,32 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
         raise NotImplementedError
 
     @abc.abstractmethod
-    def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]) -> Tensor:
+    def predict_action_chunk(
+        self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]
+    ) -> Tensor | tuple[Tensor, dict[str, Tensor]]:
         """Returns the action chunk (for action chunking policies) for a given observation, potentially in batch mode.
 
         Child classes using action chunking should use this method within `select_action` to form the action chunk
         cached for selection.
+
+        By default returns just the action `Tensor`. If `return_extra=True`, returns `(action, extra)`
+        where `extra` is a (possibly empty) `dict[str, Tensor]` of auxiliary outputs a policy may
+        expose (e.g. world-model predictions). Policies that produce nothing extra may ignore the kwarg.
         """
         raise NotImplementedError
 
     @abc.abstractmethod
-    def select_action(self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]) -> Tensor:
+    def select_action(
+        self, batch: dict[str, Tensor], **kwargs: Unpack[ActionSelectKwargs]
+    ) -> Tensor | tuple[Tensor, dict[str, Tensor]]:
         """Return one action to run in the environment (potentially in batch mode).
 
         When the model uses a history of observations, or outputs a sequence of actions, this method deals
         with caching.
+
+        By default returns just the action `Tensor`. If `return_extra=True`, returns `(action, extra)`
+        where `extra` is a (possibly empty) `dict[str, Tensor]` of auxiliary outputs a policy may
+        expose (e.g. world-model predictions). Policies that produce nothing extra may ignore the kwarg.
         """
         raise NotImplementedError