[Doc] Better doc for inverse transform semantic

ghstack-source-id: 444ad87d1ab0a829e8ce1848b5838859d5ee7494
Pull Request resolved: #2459

docs/source/reference/envs.rst

@@ -722,6 +722,9 @@ Since each transform uses a ``"in_keys"``/``"out_keys"`` set of keyword argument
 also easy to root the transform graph to each component of the observation data (e.g.
 pixels or states etc).
 
+Forward and inverse transforms
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
 Transforms also have an ``inv`` method that is called before
 the action is applied in reverse order over the composed transform chain:
 this allows to apply transforms to data in the environment before the action is taken
@@ -733,6 +736,20 @@ in the environment. The keys to be included in this inverse transform are passed
 
         >>> env.append_transform(DoubleToFloat(in_keys_inv=["action"]))  # will map the action from float32 to float64 before calling the base_env.step
 
+The way ``in_keys`` relates to ``in_keys_inv`` can be understood by considering the base environment as the "inner" part
+of the transform. In constrast, the user inputs and outputs to and from the transform are to be considered as the
+outside world. The following figure shows what this means in practice for the :class:`~torchrl.envs.RenameTransform`
+class: the input ``TensorDict`` of the ``step`` function must have the ``out_keys_inv`` listed in its entries as they
+are part of the outside world. The transform changes these names to make them match the names of the inner, base
+environment using the ``in_keys_inv``. The inverse process is executed with the output tensordict, where the ``in_keys``
+are mapped to the corresponding ``out_keys``.
+
+.. figure:: /_static/img/rename_transform.png
+
+   Rename transform logic
+
+
+
 Cloning transforms
 ~~~~~~~~~~~~~~~~~~