Carla Project RL End-to-End Auto Driving with Reinforcement Learning + Imitation Learning

https://github.com/JinmingWang/CarlaProjRL

• Carla simulator
• Reinforcement Learning, A2C
• Imitation Learning combined with RL
• Very simple path planning with grid world
• Point following
• Obstacle avoidance
• Random Repeat Random Actions
• Lidar sensor
• Point cloud to image, then to grid world

1. RL combined with IL to make the algorithm converge faster while does not need too much human demonstration
2. Process point cloud to lidar map image, then to grid world, with the ego vehicle at the center, a model is trained using A2C to find the next action in the grid world {left, right, forward, backward, idle}, running this model many times can generate a path from the agent to the target point.
3. Another model is trained using A2C to find the speed and steer action from the generated path drawn on the grid world.
4. One single random action is applied N times for the next N steps to allow better exploration, N is a random number with specified mean and standard deviation.

$ git clone https://github.com/JinmingWang/CarlaProjRL && cd CarlaProjRL
$ git submodule init
$ git submodule update

$ conda create -n carla-rl-pytorch python=3.8 torchvision pytorch==1.13 cudatoolkit==11.6 cudnn -c pytorch -c nvidia
$ conda activate carla-rl-pytorch
$ pip install carla==0.9.14 pyyaml pynvml tensorboard
$ cd leaderboard && pip install -r requirements.txt && cd ..
$ cd scenario_runner && pip install -r requirements.txt && cd ..

# Start Training (Carla server must be running):
$ python Train.py

# Train with existing saved memory on disk:
$ python TrainWithMemory.py

# Evaluation with test cases (Carla server must be running):
$ python Evaluate.py

# Play with environment (Carla server must be running):
$ python EnvRoute.py
# or
$ python EnvPracticeField.py

• The model is not learning, A2CAgent may contain potential problems.