How to use the repository

Models

• The models can be found under ddpg/CARLA or stable_baselines/carla_env_MIF or A3C_continuous

• For the stable baselines models use the parameters MODE and SMODE to define the model and the running mode

• For the DDPG use the run.py for configuration and start up

• Models can also be reloaded in the stable baselines file, we pushed some of the models - the most recent ones perform the best

• We have also pushed a custom PPO implementation

• to switch input states, modify flags

    self._data_gen  
  
    self._use_front_ae  
  
    self._use_birdseye  
  

  in the carla_env script (warning: switch training model to CNN/MLP policy in carla_env_MIF to fit the input)

Birds-eye-view

• The implementation of the encoder-decoder architecture can be found under semantic_birdseyeview
• There are also some models pushed
• separate requirements.txt in semantic_birdseyeview
• use the script train_model_rgb to train the birds-eye-view based on rgb images (should located in a "camera_storage" folder as npy)
• use the script train_model_semantic to train the birds-eye-view based on segmented images (should located in a "camera_storage" folder as npy)

Wrapper

• We used the carla_rllib wrapper from Sven Müller
• We have done some adjustments for the terminal condition and the sensor types under carlarllib/wrapper/carla_wrapper.py
• The observation space and reward function are modified in carlarllib/environment/carla_environment/carla_environment.py

Prerequisites

virtualenv -p python3 carla_env  

source carla_env/bin/activate  

pip install -r requirements.txt  

Bashrc:

CARLA_ROOT="<dir>/carla_0.9.6"  
export CARLA_ROOT  

Installation CARLA 0.9.6

0. get CARLA 0.9.6
   download precompiled version

   edit CARLA_ROOT in bashrc(add following lines if necessary):
   (pythonpath necessary to point python to carla_rllib (maybe not necessary in future versions))

   export CARLA_ROOT=<dir>/carla_0.9.6/ 
   export PYTHONPATH="${PYTHONPATH}:<dir>/carla_rllib/"  
   export PATH=$PATH:$CARLA_ROOT 

1. Clone wrapper repo

   git clone https://ids-git.fzi.de/svmuelle/carla_rllib.git

2. set up virtual environment, Python should be 3.5.0

3. checkout branch develop

   git checkout develop

4. Install the required python packages into your virtual environment:

   cd carla_rllib  
   pip install .  

5. run test files:

   python carla_env_test.py

6. If you have errors, check Versions, python, carla client and carla server have to be the same version if carla pip installs are existent, check versions

   pip list --local  

7. remove pip version if not 0.9.6 and use egg files to reinstall carla

   pip uninstall carla
   cd <dir>/carla_0.9.6/PythonAPI/carla/dist/  
   easy_install carla-0.9.6-py3.5-linux-x86_64.egg  

8. run CARLA_UE4.sh (or SDL_VIDEODRIVER=offscreen SDL_HINT_CUDA_DEVICE=0 sh CarlaUE4.sh )

9. run wrapper test files (in carla_rllib/examples)

   python carla_env_test.py