This project develops a hybrid collision avoidance system for autonomous vehicles, combining Deep Reinforcement Learning (DRL) with a Model Predictive Control (MPC) approach. Utilizing the CARLA simulator, the system is designed to navigate scenarios where an ego car approaches a stationary vehicle. Essentialy, the RL agent works as a path planner module and the MPC controls the vehicle.
The project employs Recurrent Proximal Policy Optimization (Recurrent PPO) from Stable Baselines3 for training the reinforcement learning model. The state space for the RL agent includes:
- x_dist: Lateral distance from the parked car
- y_dist: Longitudinal distance from the parked car
- current_speed
- current_acceleration
- current_yaw
The action space consists of five continuous values (x1, y1, x2, y2, desired speed). These values are used to generate a cubic spline curve that defines the path for the MPC controller, which manages vehicle navigation based on the desired speed and path.
The model demonstrates effective navigation and collision avoidance:
-
Generated Path Visualization: GIF showing the model's path planning in real-time at 90km/h with red markings to denote the RL-generated path at each timestep.
-
Performance Metrics: Graph depicting the reward received per timestep during training, indicating model performance and learning progress.
- CARLA Simulator
- Stable Baselines3
- Additional dependencies as listed in
requirements.txt
To start the simulation and train the model:
- Adjust the parameters in
train.pyand other configuration files as needed. - Run
train.pyto commence training and simulation. - Monitor outputs and adjust parameters to optimize the training and path generation.
Contributions to enhance the system or adapt it for different scenarios are welcome. Please fork the repository, propose changes via pull requests, or raise issues for discussion.
Special thanks to Carla-MPC by yangggzhang for the initial Model Predictive Control (MPC) implementation. This project uses a modified version of the Carla-MPC to integrate with our reinforcement learning model for enhanced path planning and control.