[Doc] Tutorial: recurrent training on sequence batches (#3860)

Author: theap06

docs/source/index.rst

@@ -116,6 +116,7 @@ Intermediate
    tutorials/torchrl_envs
    tutorials/pretrained_models
    tutorials/dqn_with_rnn
+   tutorials/recurrent_sequence_training
    tutorials/mujoco_cube_bowl_macros
    tutorials/collector_trajectory_assembly
    tutorials/evaluator

docs/source/reference/data_layout.rst

@@ -453,6 +453,11 @@ Pair with :class:`~torchrl.cuda_memory_profile` to verify the win:
 See also
 --------
 
+* :ref:`Recurrent training on sequence batches <recurrent_sequence_tuto>` —
+  a runnable tutorial that samples ragged trajectory slices with
+  :class:`~torchrl.data.replay_buffers.SliceSampler` and feeds them straight
+  into a recurrent policy, the contiguous-batch pattern described on this
+  page.
 * :doc:`collectors_replay` — concrete ``ndim`` patterns and the full
   ``trajs_per_batch`` API.
 * :ref:`Auto-wrapping recurrent transforms <Environment-policy-arg>` —

docs/source/reference/recurrent_state_lifecycle.rst

@@ -359,6 +359,14 @@ What to check, in order
 See also
 --------
 
+- :ref:`Recurrent training on sequence batches <recurrent_sequence_tuto>` —
+  a runnable tutorial that walks through this lifecycle end to end:
+  collection, :class:`~torchrl.data.replay_buffers.SliceSampler` replay, and
+  a full-sequence ``set_recurrent_mode(True)`` forward.
+- ``examples/replay-buffers/recurrent_slice_sampler_pipeline.py`` — a
+  minimal, runnable script version (GRU policy, the collector writing
+  directly into the buffer, ``SliceSampler`` auto-detecting the trajectory
+  key).
 - :class:`~torchrl.modules.LSTMModule` — the module that consumes
   ``is_init`` and gates hidden-state resets.
 - :class:`~torchrl.modules.GRUModule` — same lifecycle, single hidden