This is a Deep Q-Network (DQN) reinforcement learning agent which navigates a fixed wing aircraft in a simulator to a target waypoint while avoiding stationary and moving obstacles.

This is our submission to our final project of McGill University's ECSE 526 - Artificial Intelligence course.

To run this, one needs to set up MIT's director visualization tool as it will be used to display the simulator. Instructions on how to build this can be found here.

Following that, you must set up an alias for the directorPython executable that was built. This can be simply done by running:

alias director=/path/to/director/build/install/bin/directorPython

You can add this to your shell profile to avoid running this every time.

Finally, you need to install Tensorflow for Python 2.7. This can be done by following the steps here.

Once everything is setup, you can simply run:

director simulator.py

or

director simulator.py --help

for advanced options.

If everything works out, a window should appear with a plane model that flies around slowly learning how to reach the green circle and avoiding the white circle as in the screenshot above. The white circles denote obstacles, whereas the green circle denotes a target waypoint. The rays protruding from the plane represent the distances measured by the plane's sensor.

The learning process will take several hundreds of episodes, but rerunning the simulation will proceed from where it left off by reusing the model.ckpt file.

This project was inspired by the work found here.