• If you have problems with the config file, take pr2.yaml as an example, this one is used for tests and will always be up to date.

• The collision avoidance is generally not smart enough to find complicated paths around objects, but it can move along surfaces.
• Giskard will stop planning if it finds a local minimum and returns success. For example, it might get stuck in a stable position behind an object that is blocking the path. There is no check if the goal is properly satisfied, because this is not necessarily wanted.
• The collision avoidance can get unstable if there are to many objects in unfortunate configurations close to the robot.
• The API to program constraints is not (yet) easy to use and therefore not documented here.
• pybullet doesn't like mimic joints very much, complicated mimic chains seem to be buggy, which results in inaccurate collisions

If you are debugging Giskard it might be useful to look into it behavior tree.

sudo apt install ros-<your-distro>-rqt-py-trees

And then run

rosrun rqt_py_trees rqt_py_trees

and select giskards log topic.