In the Components of an AV system section, we outlined several modules that were necessary for a self-driving system. In this section, we'll look at how to implement one of them—the driving policy—with the help of DL. One way to do this is with RL, where the car is the agent and the environment is, well, the environment. Another popular approach is imitation learning, where the model (network) learns to imitate the actions of an expert (human). Let's look at the properties of imitation learning in the AV scenario:
- We'll use a type of imitation learning, known as behavioral cloning. This simply means that we'll train our network in a supervised way. Alternatively, we could use imitation learning in a reinforcement learning (RL) scenario, which is known as inverse RL.
- The output of the network is the driving policy, represented...
