This repository contains a ROS package that implements a classic DWA local planner including a terrain traversability analysis. The code was developed as part of a 6-month research internship in the Computer Science and Systems Engineering Laboratory (U2IS) at ENSTA Paris. The aim is to add a traversal cost to each of the trajectories generated by the local planner. The cost in question is predicted by a convolutional neural network (CNN) trained in a self-supervised way on images of terrains that the robot has already traversed, and is constructed from the vibrational data recorded by the robot's inertial measurement unit (IMU). A guide to how the network has been trained and how it can be used for local navigation is currently being prepared.
Figure: Visualization of live prediction of the traversal cost for 5 candidate local trajectories (a color gradient from green to red is used to represent the trajectories from the most traversable to the least traversable).
Once the package has been added to a ROS workspace, and the code compiled with catkin, the local planner should be recognized as a plugin. You can check that it is registered using the following command:
$ rospack plugins --attrib=plugin nav_coreThe plugin should appear under the name dwa_ext_local_planner.
The local planner can then be used in the classic way with move_base. To do this, simply configure move_base to use the dwa_ext_local_planner/DWAExtPlannerROS local planner, and load the custom settings for it (the configuration is described in the config/dwa_ext_planner.yaml file, part of which can be reconfigured dynamically via config/DWAExtPlanner.cfg).
The structure of the Dynamic Window Approach (DWA) is inspired by the official implementation (see here). It has been simplified and does not currently take account of obstacles. It favours local trajectories that are close to the global trajectory, that do not cause the robot to oscillate and that cross terrain with a lower vibrational cost. The code consists of the following files:
- dwa_ext_local_planner[.cpp/.h]: local planner code (to respect the structure of the official DWA implementation).
- dwa_ext_local_planner_ros[.cpp/.h]: code for integrating the local planner into ROS, the planner inherits from the nav_core::BaseLocalPlanner class.(to respect the structure of the official DWA implementation).
- traversability_cost_function[.cpp/.h]: implementation of the traversal cost function (vibrational cost) which inherits from the base_local_planner::TrajectoryCostFunction class.
In addition to these files, the files depth[.cpp/.h] implement a class for managing depth images, and compute normal maps from them (it can be useful depending on the input of the CNN).
To see the Doxygen documentation, open the file doc/html/index.html in your favourite browser.