drl_rcc_torcs

Options

• End-to-end neural network
  • Run the simulator with the "perfect" driver. Collect the images and the steering angle.
  • Create a network based on VGG
  • Error function.
  • Training
• Actor/Critic
• Segmentation

Links

• Simulated Car Racing Championship Competition Software Manual
  • https://arxiv.org/abs/1304.1672
  • https://arxiv.org/pdf/1304.1672.pdf
• Gym-TORCS
  • https://github.com/ugo-nama-kun/gym_torcs
• Deep Reinforcement Learning approach to Autonomous Navigation
  • https://github.com/bhanuvikasr/Deep-RL-TORCS
• End-to-End Deep Learning for Self-Driving Cars
  • https://devblogs.nvidia.com/parallelforall/deep-learning-self-driving-cars/
• Teaching a Machine to Steer a Car
  • https://medium.com/udacity/teaching-a-machine-to-steer-a-car-d73217f2492c
• Community Steering Models
  • https://github.com/udacity/self-driving-car/tree/master/steering-models/community-models
• Reinforcement learning with docker and torcs
  • https://github.com/bn2302/rl_torcs

Notes

• If you want to compile for yourself, use /home/gym_torcs/vtorcs-RL-color/src/libs/raceengineclient/raceengine.cpp
• to get from little screen to big screen: down, down, enter; down, enter; right, right, enter; enter
• run with command line --random option to enable random track selector

Training Data

• https://www.dropbox.com/s/r7ln2y2plyezjy0/racingdata.zip?dl=0

todo

• image augmentation, road segmentation
• test network and solver
• adjust hyperparameters