Skip to content

tekinengin/RL-frenet-trajectory-planning-in-CARLA

 
 

Repository files navigation

RL-frenet-trajectory-planning-in-CARLA

This repository is a framework that creates an OpenAI Gym environment for self-driving car simulator CARLA in order to utilize cutting edge deep reinforcement algorithms and frenet trajectory planning.

Installation

  • Simulation works as server-client. CARLA launches as server and uses 2000:2002 ports as default. Client can connect to server from port 2000, default, and interract with environment.

Client Installation

  1. git clone https://github.com/MajidMoghadam2006/RL-frenet-trajectory-planning-in-CARLA.git
  2. cd RL-frenet-trajectory-planning-in-CARLA/
  3. pip3 install -r requirements.txt (requires Python 3.7 or newer)
  4. cd agents/reinforcement_learning
  5. pip install -e .# installs RL algorithms as python packages

Simulation Server Installation

Use pre-compiled carla versions - (CARLA 9.9.2 Recommended)

  1. Download the pre-compiled CARLA simulator from CARLA releases page
  2. Now you can run this version using ./CarlaUE4.sh command
  3. Create a virtual Python environemnt, e.g. using conda create -n carla99, and activate the environment, i.e. conda activate carla99
  4. If easy_install is not installed already, run this: sudo apt-get install python-setuptools
  5. Navigate to PythonAPI/carla/dist
  6. Install carla as a python package into your virtual environment (get help): easy_install --user --no-deps carla-X.X.X-py3.7-linux-x86_64.egg

Now you may import carla in your python script.

Some Features

  • Reinforcement Learning/Gym Environment parameters are configured at /tools/cfgs/config.yaml
  • DDPG/TRPO/A2C/PPO2 are configured to save models during training with intervals and also best models with max_moving_average (window_size=100 default)

Example Training:

  • We need to start two different terminals.

Terminal-1

  • cd CARLA_0.9.9/
  • ./CarlaUE4.sh -carla-server -fps=20 -world-port=2000 -windowed -ResX=1280 -ResY=720 -carla-no-hud -quality-level=Low [CARLA documentation](https://carla.readthedocs.io/en/latest/)

Terminal-2

  • cd RL-frenet-trajectory-planning-in-CARLA/
  • python3 run.py --cfg_file=tools/cfgs/config.yaml --agent_id=1 --env=CarlaGymEnv-v1

Example Test:

Initilize the best recorded agent and associated config file given the agent_id. Test runs as --play_mode=1 (2D) as default.

Terminal-1

  • cd CARLA_0.9.9/
  • ./CarlaUE4.sh -carla-server -fps=20 -world-port=2000 -windowed -ResX=1280 -ResY=720 -carla-no-hud -quality-level=Low

Terminal-2

  • cd RL-frenet-trajectory-planning-in-CARLA/

  • python3 run.py --agent_id=1 --env=CarlaGymEnv-v1 --test

  • Pre-trained agents DDPG(ID:1), TRPO(ID:2), A2C(ID:3), PPO2(ID:4)

  • Besides the config.yaml file you can also use following parameters:

  • --num_timesteps; number of the time steps to train agent, default=1e7

  • --play_mode: Display mode: 0:off, 1:2D, 2:3D, default=0

  • --verbosity: 0:Off, 1:Action,Reward, 2: Actors + 1, 3: Observation Tensor + 2, default=0

  • --test: default=False

  • --test_model: if want to run a specific model type:str without file extension example (best_120238)

  • --test_last: if True will run the latest recorded model not the best

  • Carla requires a powerful GPU to produce high fps. In order to increase performance you can run following as an alternative:

  • DISPLAY= ./CarlaUE4.sh -carla-server -fps=20 -world-port=2000 -windowed -ResX=1280 -ResY=720 -carla-no-hud -quality-level=Low

Printing the Results

  • python3 monitor_plot.py --agent_ids 1 2 --window_size 100 --colors red blue --lr DDPG TRPO --alpha 0.1 --n_steps 1e5

Important Directories

  • RL Policy Networks : agents/reinforcement_learning/stable_baselines/common/policies.py
  • Env and RL Config File: tools/cfgs/config.yaml
  • Gym Environment: carla_gym/envs/ # Gym environment interface for CARLA, To manipulate observation, action, reward etc.
  • Modules: tools/modules.py # Pretty much wraps everything

To cite this repository in publications:

  title={An End-to-end Deep Reinforcement Learning Approach for the Long-term Short-term Planning on the Frenet Space},
  author={Moghadam, Majid and Alizadeh, Ali and Tekin, Engin and Elkaim, Gabriel Hugh},
  journal={arXiv preprint arXiv:2011.13098},
  year={2020}
}

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%