Hello,
If node filtering is not activated, all point clouds are always shown. So if there was an obstacle the first time an area is visited, it won't be cleared if coming back when the obstacle is removed.
If node filtering is activated (you should see parameters like "filter_radius" and "filter_angle" for
map_assembler and
grid_map_assembler nodes, or "Node filtering radius" and "Node filtering angle" for
MapCloud rviz plugin), only the more recent cloud in the fixed radius/angle is shown or assembled in the map. So coming back at the same place where the obstacle is removed, the newest cloud (without the obstacle) would have priority on the old cloud (with obstacle), so the obstacle will be cleared indirectly by keeping only the newest cloud in the map.
This also work when a new object is introduced in a previously visited area, the object would be added to the map.
There are also other RTAB-Map's parameters like "RGBD/AngularUpdate" and "RGBD/LinearUpdate" which will make rtabmap node to not add new point clouds if the robot has not moved/rotated more than fixed values.
For planning in dynamically environment, I recommend to update a local map directly with the inputs from the sensors while using rtabmap to generate a global map. I would refer to
move_base documentation for example about local costmap generation from sensors. The
rtabmap stereo navigation tutorial shows an example with move_base.
Regards,
Mathieu