Download Carla version 0.9.14 from the official source. After downloading, unzip the contents. You do not have to put the contents under the root of this project. To start the simulator, run CarlaUE4.exe.
Install the necessary packages by running the following command: pip install -r requirements.txt
Run Script train_agent.py
-
Train New Agent: Use
train_new_model("agent_name", total_timesteps=200000, hyperparams=hyperparams2)to train a new agent. You can specify the name of the agent, the total number of timesteps for training, and the hyperparameters. -
Train Existing Agent: Use
train_exist_model(model_path, total_timesteps=100000, hyperparams=hyperparams2)to continue training an existing agent. Specify the model path, the total number of timesteps for continued training, and the hyperparameters.
Run Script eval_agent.py
- Evaluate Agent: Use
eval_model(model_path)to evaluate an agent. Specify the path to the model you wish to evaluate.
Note: A pre-trained model was provided. If you want to evaluate this pre-trained agent, please download the model from this link and place the model zip file under path training/Saved_Models. In this case, you can directly execute eval_agent.py as the default path in the script is the path of this model.
A demo video of the pre-trained agent with 400,000 training timesteps Demo Video
Implementation details of the gym environment are in carla_env/carla_env_multi_obs
[semantic segmentation camera sensor data], [relative position, relative velocity, relative speed]
| Color | Semantic Segmentation |
|---|---|
| Purple | Route |
| Light Green | Lane |
| Blue | Vehicle |
| Black | Obstacle |
| Action | Space | Range |
|---|---|---|
| Throttle | Continuous | [-1, 1] |
| Steer | Continuous | [-1, 1] |
| Brake | Continuous | [0, 1] |
| Plan | Description | Status |
|---|---|---|
| Learning to be Malicious | Train the agent to imitate human-driven malicious attacks for automated benchmark testing | ⏳ |