I found a solution that works for me:
In my launch-file, I have a -Tag that let's me include a script:
<param name="robot_description" command="$(find robot_description_pkg)/scripts/create_urdf.py"/>
(This tag is also evaluated before nodes are launched, so I don't have to care for race conditions).
In my create_urdf.py, I read my values from the db and call the xacro.py with my base-urdf.xacro and the new params.
My base.urdf.xacro then contains entries like:
<joint name="z_joint" type="prismatic">
<parent link="insertion_x"/>
<xacro:arg name="insertion_x_max" default="0"/>
<limit effort="37" velocity="1" lower="0" upper="$(arg insertion_z_max)" />
<joint_properties damping="0.1" friction="0.5" />
</joint>
If my db doesn't provide a value for the upper_limit, the default is used, otherwise the new value is inserted into the final urdf.
The return value of the xacro.py-call is then written to the parameter-server, so that a robot-state publisher can get it.
<node pkg="robot_state_publisher" type="robot_state_publisher" name="robot_state_publisher" respawn="true" respawn_delay="50"/>
Using the param-Tag to start the script is maybe a bit hackish, but it provides a way to call my script (and to only start the robot_state_publisher after I have updated the model).
With this approach, I only have to change the values in the DB and the robot will have the new values after the next launch.
