This repository contains the algorithms implementation for vehicles scheduling, dispatching and planning in complicated scenarios such as intersection, junction etc.
Goal: To enable CAVs(Connected and Autonomous Vehicle) to imitate the driving policy of HDVs(Human-Driven Vehicle), we apply inverse reinforcement learning to learn weight parameters for our designed cost function so that the learned policy can generate trajectories close to the ground truth trajectories.
Scenario: semi-structured driveways(the driveways are mostly curved lines, however no roadlanes are given and the vehicles are not obliged to follow general traffic rules)
Cost Function: the peculiarity of the scenario results in the special design of cost functions. A reference design for cost functions in structured roadways are given here. However, in this paper the author considers structured driveways with conventional driving behaviors(in-lane following and lane changing), which is easier since the dimension of the action space is degraded. Here in this semi-structured driveway scenarios, we need to adapt these cost function terms accordingly.
The cost function consists of a linear combination of driving features. Different features represent different objectives in trajectory planning. In total, 10 features are included that reflect three main objectives:
- Safety
- Efficiency
- Comfort
The combination of all features including their associated weights defines a driving policy. The proposed features should be generic to all possible driving maneuvers.
- Acceleration:
$$f_1=\frac{1}{N}\sum_i\frac{|a_i|}{a_{max}}$$ - Jerk:
$$f_2=\frac{1}{N-1}\sum_i|a_{i+1}-a_i|$$ - Yaw rate:
$$f_3=\frac{1}{N-1}\sum_i|\phi_{i+1}-\phi_i|$$