Hello Matt, I used Inverse RL to obtain the reward function for different expert behaviors, and then used the obtained reward functions to train an RL agent. I built over your RL code with some minor changes in the environment, here is my implementation. https://jangirrishabh.github.io/2016/07/09/virtual-car-IRL/
I added the color sensing capability back (saw it in your 1st post) and changed the environment slightly, for example the obstacles are static now (oh yes I know overfitting, but its okay for just trying out IRL and getting results faster) let me know if you like it :-D