Reinforcement Learning and Data Collection for Self-Driving in Carla
Reinforcement Learning and Data Collection for self-driving in Carla with Stable Baselines 3
Usage
Reinforcement Learning
Go to the Carla-RL folder, and run:
python train_sac.py --model-name sem_sac --width 96 --height 96 --repeat-action 4 --start-location highway --sensor semantic --episode-length 600
More training/evaluation options
python train_sac.py -h
Modify reward, goal condition, etc.
See Carla-RL/carla_env.py, and modify the relevant variable values in the step() function.
Data Collection and Training Vision-based Predictors
collect_data_roadline.py: Data collection scripts. Saves image observations of the agent as well as other state information (e.g. distance to the center of the lane). The collect data can be used to train predictors offline. Sample Usage: python collect_data_roadline.py. Run python collect_data_roadline.py -h for more data collection configuration options.
train_and_val_predictor_single.py: training and validation of image-to-distance predictors using the collected dataset.
You can define pytorch custom datasets in the datasets folder and additional pytorch models in the models folder.
test_data_collection.py provides a basic script which you can use to verify whether your data collection process works as expected. (e.g. whether the sensory observations are correctly incorporated into the dataset)
Visualization of sensors
Modify manual_control.py to include the new sensors you added to the car, and (after you launched Carla) run
python manual_control.py
to see them in action.
References
The Official Carla Simulator Repo: includes basic tutorials and demos for controlling the car and get sensory infomation from the environment.
Carla-RL Repo from Sentdex: includes a basic version of the gym-style reinforcement learning environment for Carla, and a basic DQN model.
Oatomobile Research Framework for Self-Driving: A library for self-driving research with high-level APIs, baseline agents, and graphics setup.