*****************************
Phase 3: Moving a long cuboid
*****************************

.. image:: ../images/robot_phase_3.jpeg


Phase 3 works in the same way as phase 2 with only the following differences:


Different Object
================

Instead of the cube, a small cuboid is used now.

- Size: 2x8x2 cm
- Weight: 16 g

.. image:: ../images/cuboid_2x2x8.jpg


New Singularity Image
=====================

The Singularity image has been updated according to changes in phase 3.

The new image can be downloaded here: :ref:`singularity_download_image`.

.. note::

    Don't forget to update your custom images accordingly (i.e. rebuild them
    with the new image as base).


Object Tracking
===============

Due to the smaller size, the tracking of the cuboid is more difficult than with
the bigger cube.  Therefore there will likely be more noise on the estimated
object pose.

We also improved the confidence computation a bit which results in values that
are overall a bit lower than they were in phase 2.

Filtered Pose
-------------

The ``filtered_object_pose`` in
:class:`~trifinger_object_tracking.py_object_tracker.ObjectPose` is now computed
as follows:

- If confidence > 0.75, ``filtered_object_pose = object_pose``.
- Otherwise it is set to the last observed pose with confidence > 0.75.  Each
  time a pose is repeated like this, its confidence is reduced a bit.


Reward Functions and Evaluation
===============================

With the long cuboid, it is difficult to detect the rotation around the long
axis (y-axis in the object frame) reliably.  Therefore the reward function for
difficulty level 4 is changed a bit to compensate for that.  Instead of using
the full 3d-rotation, only the angle between the y-axes of the goal and actual
orientation is used.  This means that the rotation around the long side is not
considered for the reward.

Further we changed the evaluation a bit to use median instead of mean for
averaging the different runs of the same difficulty level.  This should lead to
similar results in most cases but single but large outliers will have less
influence on the resulting score.

Apart from this, no changes were made compared to phase 2.  For details see

.. toctree::
   :maxdepth: 1

   tasks_phase_3