Is it possible to replace the local planner in the default ROS Navigation stack to one that uses reinforcement learning?

edit

Definitely possible. You will need to implement a "base local planner plugin" of nav_core::BaseLocalPlanner . You can follow this tutorial to create a plugin ( http://wiki.ros.org/pluginlib/Tutoria... ).

The nav_core::BaseLocalPlanner class has a few pure virtual function so you will have to re-implement them, but basically you just need to implement your planner in the function computeVelocityCommands.

You can also access the costmap_2d object (which is given by the initialize function) so you can perform any logic/algorithm to detect/avoid obstacles.