🚗 CARLA-SB3-RL-Training-Environment

This repo provides an out-of-the-box training and evaluation environment for conducting multiple experiments using DRL in the CARLA simulator using the library Stable Baselines 3 including the configuration of the reward function, state, and algorithm used.

Video with examples of the pretained models provided: here

This work has been developed as part of the Bachelor's Thesis "Application of Deep Reinforcement Learning in autonomous driving" by Alberto Maté at UC3M.

📥 Installation

Install CARLA 0.9.13 from here. (Note: CARLA 0.9.13 is the only version tested with this repo)
Create a venv and install the requirements:

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Export CARLA installation path to $CARLA_ROOT:

export CARLA_ROOT=<path to carla installation>

🛠️ Usage

Configuration file

The configuration is located in config.py. It contains the following parameters:

algorithm: The RL algorithm to use. All algorithms from Stable Baselines 3 are supported.
algoritm_params: The parameters of the algorithm. See the Stable Baselines 3 documentation for more information.
state: The state to use as a list of atributes. For example, steer, throttle, speed, angle_next_waypoint, maneuver, waypoints, rgb_camera, seg_camera, end_wp_vector, end_wp_fixed, distance_goal See the carla_env/state_commons.py file for more information.
vae_model: The VAE model to use. This repo contains two pretrained models: vae_64 and vae_64_augmentation. If None, no VAE is used.
action_smoothing: Whether to use action smoothing or not.
reward_fn: The reward function to use. See the carla_env/reward_functions.py file for more information.
reward_params: The parameters of the reward function.
obs_res: The resolution of the observation. It's recommended to use (160, 80)
seed: The random seed to use.
wrappers: A list of wrappers to use. Currently there are two implemented: HistoryWrapperObsDict and FrameSkip. See the carla_env/wrappers.py file for more information.

Training

To train a model, run:

python train.py --config <number of the config to use> --total_timesteps <number of timesteps to train>

For example:

python train.py --config 0 --total_timesteps 1000000

The training results will be saved in the tensorboard folder. You can open it with:

tensorboard --logdir tensorboard

Evaluation

To evaluate a model, run:

python evaluate.py --config <number of the config to use> --model <path to the model to evaluate>

The evaluation routes can be changed inside carla_env/envs/carla_env.py in the eval_routes variable. Choose two points in the map and add them to the list.

Train and evaluate multiple models

To train and evaluate multiple models run the run_experiments.py script. It will train and evaluate all the models specified in the run_experiments.py file.

python run_experiments.py

Train VAE models

In this repo you can also train and eval a VAE model. To train a VAE model, run:

python vae/train_vae.py --epochs <number of epochs to train>

There are also some script to recollect data (RGB and segmentation images) from CARLA. To collect data from CARLA manually, run:

python carla_env/envs/collect_data_manual_env.py

To collect data from CARLA automatically using a RL agent in early stages, run:

python carla_env/envs/collect_data_rl_env.py

alberto-mate/CARLA-SB3-RL-Training-Environment