feat(policies): Add X-VLA (#2405)

* first commit

* more fixes

* add franka action

* update testing script

* add changes

* update files

* logits matching

* add imagenet as a norm type

* logits matching atol1e-2

* more eval fixes

* more changes

* xvla works on libero

* remove seed

* more refactoring

* more fixes

* more changes

* more changes

* more fixes

* migrate policy revert

* major pre-commit cleanup

* renaming

* revert to self.transformer

* refactor

* new changes

* clean

* update libero

* more changes

* make it work

* more changes:

* remove imagenet dependency

* style

* more

* more refactor

* remove proprio

* add loss

* more

* more

* add freeze/unfreeze options

* add testing

* upgrade transformers version

* update testing

* add installation

* remove .sh file

* fix testing

* silent linter in xvlatest

* fix failing test

* upgrade test, fix failing

* fix testing

* more fixes to testing

* require cuda in tests

* temp check

* add xvla docs

* fix styling

* update libero doc

* remove timm dep

* add different dtype support

* remove timm skip

* remove white lines

* Enhance X-VLA finetuning documentation with optimizer details (#2537)

Added detailed instructions for implementing a custom optimizer and modifying parameter retrieval for X-VLA finetuning.

Signed-off-by: Jinliang Zheng <54488861+2toinf@users.noreply.github.com>

* fix style

* iterate on review

* iterate on cpilot

* revert xvla dep

* free up ci

* test(xvla): remove main test (#2565)

* Add xvla custom optim and dtype (#2567)

* add custom optim

* add custom optim

* add auto mode

* more changes

* add identity to all

* add auto

* release

* add docs

* make image smaller docs

* smaller image in doc

* evan smaller image doc

* finalize doc

---------

Signed-off-by: Jinliang Zheng <54488861+2toinf@users.noreply.github.com>
Signed-off-by: Steven Palma <imstevenpmwork@ieee.org>
Co-authored-by: Jinliang Zheng <54488861+2toinf@users.noreply.github.com>
Co-authored-by: Michel Aractingi <michel.aractingi@huggingface.co>
Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>

docs/source/libero.mdx

@@ -62,6 +62,11 @@ lerobot-eval \
 
 - Pass a comma-separated list to `--env.task` for multi-suite evaluation.
 
+### Control Mode
+
+LIBERO now supports two control modes: relative and absolute. This matters because different VLA checkpoints are trained with different mode of action to output hence control parameterizations.
+You can switch them with: `env.control_mode = "relative"` and `env.control_mode = "absolute"`
+
 ### Policy inputs and outputs
 
 When using LIBERO through LeRobot, policies interact with the environment via **observations** and **actions**: