Reinforcement Learning Based Controller for a Race car simulation

Simulation setup provided as part of Automated and Connected Driving Challenges course

Proximal Policy Optmization was used.

Reward function:

lap complete : +1000

crash : -100

everytimestep: +linear_velocity/100

Result:

after ~1M iterations, the model is able to navigate the race track without major collisions for a lap. lap time was recorded to be 11-16 seconds. This can be improved with further training.

RLplay.mp4

To train the controller :

roslaunch racing train_controller.launch

TRAINING.mp4

To run the trained controller :

edit the code to include the proper location of trained model. (./model/PPO_racing_cart3)

roslaunch racing RaceCar.launch